In the mid-1990s, a Finnish pulp-mill-turned-consumer-electronics-company agreed to let its main partner, Texas Instruments, combine its digital signal processor with technology from an obscure British company named Arm. The result of this novel collaboration was the iconic Nokia 6100 series of phones which went on to sell nearly 50 million units. In addition to being the first cell phones to include built-in video games like Snake and Logic, the Nokia 6110 was also the first mobile device to use the Arm architecture which would become the backbone of compute for decades to come.

Arm itself is a rather bland company with a unique business model. Sometimes referred to as “the Switzerland of the electronics industry,” Arm is a semiconductor company that doesn’t make semiconductors. Instead, the company designs processor architectures and licenses them. Its $4 billion in annual revenues come from royalties on every chip that is made which uses its designs or speaks to its instruction set.

Subscribe now

Arm is everywhere. For almost thirty years, everything mobile—from the 6100 to the first iPhone to modern phones running on M4 and Snapdragon chips—has used Arm’s licenses. Nearly every modern car has a suite of Arm chips. Practically all 5G infrastructure relies on Arm in some way. The hyperscalers now run significant portions of their infrastructure on custom Arm server chips. Most laptops and tablets run on chips that rely on Arm licenses. Smart thermostats, industrial sensors, medical devices, point-of-sale terminals, smart meters, security cameras, and billions of other devices that most people don’t even think about as computers utilize Arm’s microcontroller line. As the company itself puts it, “Arm powers the compute foundation of modern life, with more than 350 billion Arm-based chips shipped to date.”

Those staggering numbers are evidence of more than just a successful business. Arm has become an essential part of the compute ecosystem because it has remained neutral. As CEO Rene Haas told The Verge in 2022:

We really try to stay as neutral as possible. … We are involved in the ecosystem of ecosystems. If you start at the lowest level of the semiconductor chain — GlobalFoundries, Samsung, TSMC, Intel, all the people who build chips — you have to work with all of them. We have to make sure that our technology is going to be able to be built on every semiconductor process in the world, which requires investment across all of those partners.

Since the 1990s, Arm’s neutrality has been a fact of the market that most took for granted. The reality is that every company which builds on Arm’s architecture, including Apple, Google, Amazon, Qualcomm, Samsung, and Nvidia, does so under a license that Arm can revoke. In October 2024, that reality became clear when Arm sent Qualcomm a termination notice, threatening to end the architectural license that Qualcomm has used for thirty years to design and ship chips. Qualcomm won that battle, but the war is far from over.

That first strike at Qualcomm demonstrated that the instruction set which runs modern civilization is not a public good. It is owned by a foreign company that is beholden to foreign shareholders. It is also a chokepoint in the tech stack that its owners have demonstrated a willingness to leverage against competitors. We’ve seen this problem before. The last time it was Intel and it took decades and billions in investment to route around. Policymakers in Washington have developed strategies for semiconductors, AI, submarine cables, spectrum, and more. What they haven’t developed yet, and increasingly need, is a strategy to secure the layer that sits between all of them.

The Invisible Layer

An instruction set architecture (ISA) is the translation layer between software and silicon. It is the list of fundamental operations a processor knows how to perform (e.g. add these two numbers, fetch this value from memory, jump to that application) and the precise way a program must phrase those requests. Everything above it is written assuming the processor underneath will understand those instructions. Everything below it exists to fulfill those instructions. The ISA is the connector between two halves of computing.

The reason it matters so much is the reason it stays invisible; the instructions are incredibly difficult to alter once implemented. A program compiled for one instruction set will not run on a processor that speaks a different one. Switching architectures means recompiling everything and, where source code is missing or the instruction syntax is different, it means starting over from scratch. For example, Apple has now done this migration three times, migrating from Motorola 68k to PowerPC to Intel x86 and finally to its own Arm-based chips. Each time it has migrated, it has had to develop and deploy an emulation layer to carry the old software across. For most of the industry, the instruction set is chosen once and then lived with for decades. The choice is less like picking a vendor than like pouring a foundation.

This structure exists well beyond mobile phones, though phones are where Arm won most decisively. Anything that runs software must be compiled against some ISA and therefore linked, quietly and durably, to whoever controls that architecture.

Arm’s Business Model

Arm owns no factories and (until a few months ago) makes no chips. What Arm sells is the design and the language, the blueprints for processor cores and licenses to the ISA those cores implement. A customer pays an upfront fee for access and then a small royalty on every chip it builds using Arm’s technology while software developers are given free access to induce uptake.

This model was not the plan. Arm was founded in 1990 as a joint venture between Acorn Computers, Apple, and the chip firm VLSI Technology. Apple, which took a 43 percent stake in the company, wanted a low-power processor for a handheld device called the Newton. The Newton flopped and Arm pivoted. Co-founder Robin Saxby realized that licensing Arm’s designs to everyone and collecting a small cut whenever any of them succeeded had the potential to be far more lucrative, and he was right.

Here the analysis must digress because there are two kinds of Arm licenses and the difference between them runs underneath everything that follows.

The first is the standard technology license, sometimes called a TLA. Under it, a customer takes one of Arm’s ready-made cores and drops it into its own chip. This is akin to buying a prefab house design: you purchase a license to the blueprints for a house that is already designed, build it on a plot of land you own, maybe change the paint color, and voilà you have a house without much hassle. This is the way that most of Arm’s licenses work and is the low effort path since anyone can get a proven core and a quick route to market. In exchange you accept Arm’s engineering choices and pay them a royalty.

The second is an architectural license (ALA) under which a licensee uses only the instruction set and designs its own cores from the ground up. If we continue with the home construction metaphor, an ALA is like being a licensed architect who designs a house however they like so long as it’s built to the same building codes so that it connects to the same water, power, and sewer lines as every other house in the city.1 This is the hard, expensive path, and the club is small, composed of Apple, Qualcomm, Amazon, and a handful of others. Designing a high-performance core from scratch costs years and hundreds of engineers, but it is the only way to build something genuinely better than what every competitor is buying off the shelf. The fact that Mac laptops have outrun the performance of Intel on a fraction of the power comes from the fact that Apple licenses Arm’s language and then builds cores nobody else has.

The instruction sets became a trap in the ordinary way that standards do. That compact 16-bit ISA in the old Nokia chip exists because customers building memory-constrained devices needed code that took up less space and Arm developed the architecture to give it to them. Each such extension made Arm chips fit more devices, and each new device added more software written using Arm. The more software written using Arm made the next chip company’s decision to use a different ISA a little more painful and expensive. Lock-in of this kind accrues, one reasonable accommodation at a time, until the accommodations are load-bearing.

Subscribe now

The End of Neutrality

For most of its life Arm sold one thing above all: neutrality. Arm CEO Haas has consistently upheld that the company works with every chipmaker and every foundry, competes with none of its customers, and can therefore be trusted with the one decision a customer can never easily reverse. When the thing you are choosing is a foundation you will build a decade of products on, the supplier’s neutrality is essential. A licensor that might someday become a competitor is a dangerous thing to depend on.

That promise has eroded over the past ten years. In 2016, the Japanese conglomerate SoftBank acquired Arm for roughly $32 billion dollars and took it private. SoftBank later tried to sell Arm to Nvidia for forty billion dollars but federal regulators stepped in. Because Arm is “the world’s most popular computing platform,” as Nvidia CEO Jensen Huang put it, the Federal Trade Commission concluded that a combined Nvidia-Arm would have “both the ability and incentive to use its control of Arm to diminish competition by undermining Nvidia’s rivals.”

Undeterred, SoftBank took the company public again in 2023, while retaining about 87 percent of the shares. As a near-total owner that paid a premium and watched a profitable exit collapse, SoftBank wants growth and Arm’s current business model is slow-and-steady. The faster way for SoftBank to make its money back is to capture more of the value in each chip, and the fastest way of all is to stop selling only the blueprint and start selling the chips.

So Arm did. Reuters reported in 2024 that the company had drawn up plans to raise prices by as much as 300 percent and had discussed designing its own chips to compete with its biggest customers. By early 2025 the Financial Times reported that Arm would ship a finished server CPU with Meta as a launch customer. And in March, Arm did something it hadn’t done in thirty-five years: it announced its own silicon. The chip, called the Arm AGI CPU and built with Meta on TSMC’s most advanced process, is a complete data-center processor Arm plans to sell directly to hyperscalers. “We are now in a new business for Arm to supply chips,” Haas said at the launch.

The Switzerland of the industry had entered the war it spent three decades promising to stay out of. Its first target in this new war was Qualcomm, a company with a long history of bad blood between itself and SoftBank.

This is the frame the Qualcomm lawsuit belongs inside. Treated as an isolated contract dispute, the case is a confusing tangle of license terms. Treated as a first strike against Arm’s smallest major competitor, it is much more clear. Qualcomm had acquired a startup called Nuvia in 2021 for $1.4 billion. Nuvia had been founded by veteran chip designers to build Arm-based cores for data centers, and it held its own architectural license. Qualcomm took Nuvia’s work and turned it into the Oryon cores that now power its laptop and phone chips. Arm sued, arguing that Nuvia’s license could not simply transfer to a new owner and that Qualcomm needed to stop using its instruction set and destroy the Oryon designs.

Notice what Nuvia designs chips for, what Qualcomm wants to focus more on, and what Arm’s new chip target is: data centers. The dispute Arm framed as a licensing technicality lines up perfectly with the market everyone is now stepping over each other to acquire a share of.

A neutral licensor does not want its customers to destroy lucrative chip designs. A competitor does.

After losing that first suit, Arm escalated. In October 2024 it served Qualcomm with a 60-day notice that it would cancel its ALA. The threat was extraordinarily audacious and amounted to a death sentence for one of America’s preeminent chip companies. Arm later withdrew the termination notice and lost in court. A Delaware jury cleared Qualcomm in December 2024, and in September 2025 a federal judge entered final judgment for Qualcomm on the last remaining claim and denied Arm a new trial.

Regardless of the outcome in court, the industry took notice that Arm was suddenly willing to do what would have been unthinkable a decade ago. The new reality for any company reliant on Arm licenses is that a 60-day notice can hit your inbox at any time and the only remedy is litigation so expensive that most licensees couldn’t survive it.

One more fact belongs here, and it is the one Washington has not absorbed. Arm is a British company, owned by a Japanese conglomerate, listed on an American exchange, and sitting at the foundation of nearly all American computing. The United States has an industrial policy for chip fabrication that includes tens of billions of dollars aimed at manufacturing. It has no policy at all for whether and how the instruction sets those chips implement must be made available. As a matter of competition and even national security—the Department of War’s tech stack is totally reliant on tech companies and chips built on top of Arm’s ISA—the leading AI and chip companies need access to the most prominent instruction sets.

Everybody Builds Their Own

If you are one of the handful of companies large enough to hold an architectural license, the rational response to a licensor that has stopped being neutral is obvious: depend on it less and build your own. That is exactly what they are all doing.

Apple has the longest head start. Its M-series and A-series chips are cutting edge and it holds an ALA license that runs through 2040. Notably, Apple inked this deal just before Arm’s IPO, locking in long-term access precisely because the change in Arm’s ownership had made access feel less certain. Amazon, Google, Microsoft, and Nvidia also now design their own CPUs. Every one of those chips is built on the Arm ISA.

So the escape from reliance on Arm is only partial. Most major firms have escaped the engineering dependency and no longer need Arm to design their cores. They have not escaped the language dependency, and they cannot, because the real lock-in comes from the ISA. Tellingly, Amazon’s Graviton CPUs are not sold to anyone outside Amazon and even those fully in-house chips still run on Arm’s instruction set. Amazon still pays Arm a royalty for each one it makes.

Who Should Own the Instruction Set?

When the landlord of the language merely collected rent and competed with no one, building your own house on that land was safe. When the landlord starts building houses of its own and suing tenants whose houses it covets, every tenant’s investment looks suddenly more precarious.

All of which leaves one question: who should control the instruction set that nearly all the world’s computing depends on? There are three possible answers and industry has not yet settled on which is the right one.

The first answer is the status quo where a single non-American company owns the layer. This company has no obligation to license it to anyone on fair terms and the industry makes do with whatever that company decides. The case for this situation is the case for any property right: Arm built the architecture, Arm maintains it, and the incentive to keep investing depends on Arm capturing the returns. The trouble is that we have run this experiment before with the other dominant ISA. Intel owned x86 for decades, declined to license it broadly, and the result was a long stagnation in which the architecture advanced at the pace a single unpressured incumbent felt like advancing it. The escape from that, when it finally came, came from Arm and it took nearly forty years and billions of dollars. As both Intel and Arm demonstrate, unconstrained architectural control has a tendency to calcify and, this time, it is being calcified under the control of an increasingly unpredictable actor.

The second answer is regulation. There is a well-developed model for exactly this problem in the world of telecommunications standards. When a patent is essential to an industry-wide standard that everyone must implement, its holder is typically required to license it on “fair, reasonable, and non-discriminatory terms,” otherwise known as FRAND. The whole point of FRAND is to stop the owner of something everyone is forced to use from weaponizing that position. The complication is that the obligation traditionally only applies to standard-essential patents and Arm’s architecture is not a formal standard adopted by a standards body. It is private intellectual property, and so the existing FRAND system does not currently reach Arm. But the Qualcomm episode is precisely the argument for asking whether something like FRAND should apply to Arm’s instruction set, and some have already determined that the answer should be “yes.” Sens. Katie Britt and Adam Schiff recently introduced an amendment to the upcoming National Defense Authorization Act (NDAA) that would require all foreign-owned ISA providers to offer their instruction sets to US companies on FRAND terms. An ISA that nearly all of computing is required to implement is functionally an essential standard, regardless of what the paperwork says.

The third answer is the one gathering the most momentum: build an instruction set with no owner at all. RISC-V is an open, royalty-free architecture that emerged from a research project at UC Berkeley and is now governed by a nonprofit. There is no license to revoke, no royalty to raise, no company that can issue a 60-day notice. Anyone can build a RISC-V core, and the architecture cannot be bought, sued, or turned against the people who depend on it. The problem is that, currently, the cores and the software ecosystem around RISC-V are years behind any proprietary architecture.

China is building toward RISC-V as a matter of national survival because an open architecture is one critical technology that cannot be embargoed. Europe has put 240 million euros into a program called “Digital Autonomy with RISC-V in Europe” to build sovereign supercomputing chips on the open architecture in order to reduce the continent’s dependence on processors it does not control. Last year, both Meta and Qualcomm acquired companies with expertise in RISC-V-based AI inferencing chips while AMD, Google, Intel, Microsoft, and Nvidia are all active members and contributors to the organization developing the RISC-V standard. Each of these actors has realized that depending on an instruction set that someone else owns is a risk, and the only way to eliminate the risk entirely is for the layer to belong to no one. And yet, RISC-V is still far from a direct competitor to Arm and it will remain that way for at least the foreseeable future.

The Nokia 6110 divided the labor and the profits cleanly: Nokia made the phone, Texas Instruments made the radio, and Arm made the language. For thirty years that division held because the company that made the language had organized itself to be dependable. It is no longer organized that way. The instruction set was always the most permanent decision in computing, the one foundation that could not be re-poured. What the last few years have established is that the ground it sits on has an owner after all and that the people building on it have started, quietly and at enormous expense, to look for ground that does not.

In honor of the semiquincentennial anniversary of our great nation, we decided to break from our regularly scheduled programming. Since its founding (and, indeed, since well before our formal founding), ingenuity and innovativeness have been pillars of the American character. To highlight the history of this part of the American character, we present 25 Fun Facts About American Innovation.

Subscribe now

1) The U.S. Navy had ice cream ships! During World War II, the U.S. Navy commissioned dedicated refrigerated barges, dubbed "ice cream barges" by sailors, capable of producing up to 10 gallons of ice cream every seven minutes. The Navy viewed ice cream not as a luxury but as a strategic tool for troop morale, investing $1 million in one such barge alone. Ice cream was in such high demand onboard Navy ships in the South Pacific that when smaller vessels rescued downed Navy pilots, they would often barter with the pilot’s aircraft carrier. As Lt. Cmdr. Norman P. Stark later recalled: “After disembarking from the canvas bag [used for water rescues], I was greeted like a long lost brother. What I didn’t realize at the time, was that they weren’t seeing me, but what I was worth to them—10 gallons of ice cream.”

2) In the 1960s, the CIA surgically implanted a microphone, antenna, and battery into a live cat to eavesdrop on Soviet officials as part of a program called Acoustic Kitty. The first spy cat was deployed to listen in on two men outside the Soviet embassy in D.C. but was soon hit by a taxi and died. The program was quietly discontinued shortly thereafter.

3) The Toledo War between Michigan and Ohio was caused in-part by poorly-functioning mapping technology that placed the southern tip of Lake Michigan much further north than it actually was, causing Congress to draw overlapping borders for Ohio and Michigan. After months of conflict, Congress finally stepped in to resolve the issue with a compromise that granted Michigan the mineral-rich Upper Peninsula and admittance to the Union as the 26th state. Ohio got Toledo.

4) In 1942, after the large hemp production facilities in the Philippines and East Indies fell to the Japanese, the U.S. Department of Agriculture produced Hemp for Victory. This 14-minute propaganda film urged Kentucky and Wisconsin farmers to swap their maize fields for hemp, a plant the federal government had made illegal just five years earlier and now needed large quantities of to produce rope for the Navy. After the war, Washington not only banned hemp cultivation again but denied the film’s existence entirely until cannabis activists tracked down two VHS copies in 1989 and donated them to the Library of Congress.

5) Abraham Lincoln is the only president to have held a patent. It was granted in 1849 for an inflatable device meant to lift boats over sandbars.

6) In exchange for allowing the Atlantic Telegraph Company to use a US Navy vessel to lay the cable, James Buchanan received the first trans-atlantic telegraph message from Queen Victoria in 1858. The message read, “Her Majesty desires to congratulate the President upon the successful completion of the great international work, in which the Queen has taken the deepest interest.” It took over 16 hours to transmit this message and, less than a month later, that cable was rendered useless as high-voltage current burned through the poorly designed insulation.

7) Henry Ford was a bit of a soy boy. Because he distrusted the extraction-based industrial economy of mining and drilling, Ford thought agriculture could supply what industry needed instead. To prove his point, he built a car body from soy-based plastic that he’d whack with an axe to demonstrate its strength, and wore a full suit made of soy fiber to public events. He served soybean banquets at the 1934 World’s Fair where every course, including the coffee, was soy.

8) After the Clintons launched whitehouse.gov in 1994, somebody bought whitehouse.com and turned it into a porn site. Congress had to intervene to get him to stop, resulting in the Truth in Domain Names Act.

9) America had a strategic helium reserve buried under the Texas panhandle which was started in the 1920s because the military was convinced helium-filled airships would win future wars. In a controversial move, the federal government sold off the helium reserve in 2024 because it had become a money-losing liability that no longer served a clear strategic purpose.

10) Contrary to what some Europeans may believe, soccer would not exist without the USA. American chemist Charles Goodyear patented the process for vulcanizing rubber in 1836, and he later applied the technology to create the first modern soccer ball in 1855.

11) Always experimenting with electricity, Ben Franklin once killed a turkey for Christmas dinner using two Leyden jars (early capacitors), accidentally shocked himself, and was left traumatized. As he wrote: “I have lately made an Experiment in Electricity that I desire never to repeat. Two nights ago being about to kill a Turkey by the Shock from two large Glass Jarrs containing as much electrical fire as forty common Phials, I inadvertently took the whole thro' my own Arms and Body.”

12) Project Plowshare was an initiative launched by the Atomic Energy Commission in the 1950s to use nuclear explosions for excavation and one idea was to build a new sea-level canal through Nicaragua that would be called the “Pan-Atomic Canal.” Decades later, Newt Gingrich invoked the initiative as a potential solution for conflict in the Strait of Hormuz.

13) Thomas Jefferson invented a cylindrical wheel-based encryption device in the 1790s for secret diplomatic correspondence. It was forgotten and independently rediscovered a century later when the US Navy was designing the M-94 cipher device that saw use through WW2.

14) In 1994, a USAF lab at Wright-Patterson formally proposed a non-lethal chemical weapon that would make enemy soldiers irresistibly attracted to one another, causing unit cohesion to break down (colloquially known as the “gay bomb”). The $7.5 million proposal was rejected.

15) ENIAC, the world’s first programmable, general-purpose computer which drew power from a lab at the University of Pennsylvania, was so large and powerful that it reportedly made the lights flicker in Philadelphia whenever it was turned on.

16) During WW2, the US strapped tiny incendiary devices to thousands of bats, planning to release them over Japanese cities where they’d roost in wooden eaves before igniting. The program was canceled after the bats accidentally torched a US auxiliary airbase in New Mexico.

17) Everyone knows that Ben Franklin invented the lightning rod and bifocals, but he also invented the flexible urinary catheter. He never patented any of these, believing they should be freely shared.

18) One of the earliest uses of the ARPANET was to coordinate a sale of marijuana and the trade of Grateful Dead set lists between Stanford and MIT students. This is commonly cited as one of the first instances of e-commerce.

19) Thomas Edison’s first patent was for an electric vote-recording machine in 1869. Legislators rejected it because it counted votes too quickly.

20) Clinton’s NetDay Initiative, which sought to “wire the nation's schools for Internet access,” was primarily volunteer-driven. During just the pilot program in California, Silicon Valley companies donated around $27 million in funds and equipment and an estimated 20,000 volunteers connected 2,600 schools to the internet in a single day.

21) “The most beautiful woman in film,” Heddy Lamarr, was also into deep tech. In collaboration with Composer George Antheil, Lamarr developed and patented a radio technique known as frequency hopping that was intended to prevent Allied torpedo guidance signals from being intercepted by rapidly switching the radio frequency between transmitter and receiver. The Navy did not deploy this technology at the time, but it later became a crucial innovation for the development of Wi-Fi, Bluetooth, GPS, and modern military communications.

22) The term “computer bug” has entomological origins. When attempting to fix an issue with the Harvard Mark II computer in Cambridge, Massachusetts in 1947, researchers reported that a dead moth had been caught between relay contacts, shorting the system.

23) The U.S. Postal Service briefly experimented with rocket mail. In 1936, the Postal Service sent two rockets carrying mail from Greenwood Lake, New York to Hewlitt, New Jersey. When the rockets crashed on a frozen lake well outside of town, the postmaster of Hewlitt stayed true to his postal oath and carried the two bags of letters the rest of the way. Then, in 1959, the Postal Service launched a Regulus cruise missile carrying 3,000 letters from a submarine to a naval station in Florida. The Postmaster General declared: “Before man reaches the moon, mail will be delivered within hours from New York to California, to England, to India or to Australia by guided missiles."

24) It is common knowledge that the Department of War deployed multiple mobile Pizza Huts and Burger Kings to the Middle East during the Second Gulf War. What’s less well known is that during WW2, the Army shipped so much Coca-Cola to troops overseas that Eisenhower personally requested 10 portable Coke bottling plants and 6 million bottles of Coke per month for the front lines. In response, Coca-Cola sent 148 employees to Europe to run the operation, affectionately nicknamed the “Coca-Cola Colonels.” By the war's end, military service personnel had consumed over five billion bottles of Coke.

25) Al Gore invented the internet.

“Will the corporation,” asked artificial intelligence pioneer Herbert A. Simon in 1960, “be managed by machines?” Five years later came The Shape of Automation: For Men and Management, and five years after that, a curious rejoinder from a forgotten metaphysician: Frederick D. Wilhelmsen, known affectionately and universally as Fritz.

“Reasoning and Computers” (1970), a ten-page article in Thought—Fordham University’s quarterly review of culture and ideas—steered clear of the research agenda of Simon and company (into the epistemology of computational problem solving) to present a rival thesis. The processes and content of the human intellect, he insisted, entailed something much more and different than algorithms or even concepts alone could explain. While few public-facing academics of his time were more Catholic or indeed more Thomist than he, Wilhelmsen drew ably in his defense of human intellectual exceptionalism on the wryly apophatic assessment of David Hume that the artifacts of ideation could never deductively establish being – existence or is-ness. The affirmative judgment of what is—of what has being—expressed for Wilhelmsen the uniquely creative and existential culmination of the human intellect itself.

Being a good Thomist, however, Wilhelmsen reached at once for Aristotle to develop in his argument a demonstrative distinction between computer and human problem-solving. “The ‘quick wit’ of integration,” he promised—a faculty that Aristotle dubbed ἀγχίνοια—“is mankind’s best hope for mastering and ordering the new electronic revolution which otherwise threatens to drown us all in a sea of information.” (Should you sense an almost family resemblance in Wilhelmsen’s framing of the problem to that of Marshall McLuhan, the two indeed were friends and frequent correspondents; like McLuhan, Wilhelmsen was a willing coauthor, publishing The War in Man: Media and Machines (also 1970) and Telepolitics: The Politics of Neuronic Man (1972) with Jane Bret, a Montessori educator and mother of three who partnered with the Hungarian monks at the University of Dallas to co-found the city’s Cistercian Preparatory School.)

What we transliterate as agchinoia—and it is here where our story begins—is not a word of Aristotle’s own coinage. It is first found in the mouth of Socrates, who, in the Charmides, Plato has interrogate the titular lad on the true meaning of moderation or self-control (sophrosyne). In a scene not tonally out of place in some contemporary anime, the randy philosopher, it transpires, had had his own self-control tested by a glance at the popular and handsome Charmides: “I saw inside his cloak,” says Socrates, “and caught on fire, and was quite beside myself.” Following the implicit logic of this experience, Socrates attacks the traditional notion that self-control is best defined by the careful cultivation of inward stillness (hesychia), contrasting it directly with the “sharpness” or “readiness of soul” he calls agchinoia. “In all cases,” he ventures, “both in regard to the body and the soul, the qualities of swiftness and sharpness are clearly more beautiful than those of slowness and quietness,” concluding that sophrosyne “cannot be a sort of quietness, nor can the moderate life be a quiet one.”

Subscribe now

Ironically, Aristotle, the more Thomist-friendly of the two towering Greeks, better reflected the swagger of the man of agchinoia. A flashy dresser with a fashionable hairstyle and rings plural bedecking his fingers, although a noted wife guy who considered pederasty a bad habit, his public morals attracted little of the attention, good or ill, directed at Socrates. Likewise, in Aristotle’s well-manicured hands, agchinoia is treated to something of a glow-up. In the Nicomachean Ethics, the term takes on more sharply tailored proportions—a sort of Odyssean skillfulness of nimble and supple intellect. Excellence in deliberation, says he, takes time, but agchinoia is an act of instantaneously good aim or sound conjecture, leaping over the reasoning process altogether to an accurate snap judgment about what or who is who or what they are. (In the Posterior Analytics, he defines the term more technically as involving an instant mental actualization of the middle term of a syllogism.) From here, agchinoia—compounded, after all, from the terms for grasping and understanding—gained its common English translations as “presence of mind,” or even “quick wit” in Wilhelmsen’s formulation.

In other words, in agchinoia we find a quintessentially Western phenomenon that holds out hope for human beings to distinguish themselves as subjects even from the computational objects deductively dominating more and more of what once was purely human space. The instantaneous exercise of judgment in the realm of what is and is not—shattering illusions and enabling agency all on the fly—would seem to be, as Wilhelmsen advertised it, a faculty still granted and reserved for human beings in an increasingly crowded epistemological landscape. The primal Western virtues explicated by Aristotle as virtues as such, human virtues, hinge on the not just immediate but intuitive and integrative faculty of agchinoia, which makes it possible, with neither deconstruction nor delay, to love and be loved in turn, to rule and be ruled in turn—in short, to live fully human, fully relational lives, rooted as must they be in accurate apprehension of true versus false identities, whether across the table or in the mirror.

Braving the weirding world

Most intriguingly, however, anyone familiar with peak anime’s most popular and powerful themes will see in agchinoia a virtuously agentic faculty strikingly similar to what drives both the plot mechanics and the character development of, to take the strongest example, the heroes stretched across the many decades and seasons of JoJo’s Bizarre Adventure.

For the uninitiate, the great multigenerational throughline of JoJo sees the Joestar family’s protagonists and their far-flung allies uniting for adventures forced upon them by the rise of an ever-changing and deeply malevolent cast of supervillains and henchmen. Both heroes and villains in the JoJoverse wield semi-incarnate manifestations of their inner spiritual states, spectrally mechanical entities with very physical capabilities known in the series as Stands. As each person is different, so is each Stand, with the exciting and perpetually fresh result that our heroes discover from day to day and confrontation to confrontation an unpredictable menagerie of lethally dangerous monstrosities, no two the same in their bizarrely superhuman capacities to surprise, defy, confuse, destroy, and, ultimately, kill. JoJo and friends gamely manifest their own unique Stands to meet the moment time and again. But what’s truly special about them is their excellence in bravely acting under immense pressure on their sudden insights into the inscrutable win conditions of every conflict—forming improbable chains of judgment that take into account the nature of their foes and the particularities of the often incredibly warped spacetime of each pitched battle and the larger war.

These holistic flashes of comprehensive, integrative insight—of defensively weaponized agchinoia—the Japanese call kiten (機転). Like agchinoia, kiten combines two related concepts: in place of grasping (agchi), triggering mechanism (ki); in place of mind (noia), turn or shift (ten). Kiten refers to the ability to sagaciously transform a situation on a dime, especially an importune one, turning a bewildering or overwhelming experience of disadvantage into a uniquely narrow yet wide-open opening onto victory. The initially incomprehensible and horrifying characters and attacks barraging JoJo’s heroes are overcome time and again through the startling application of kiten, the vile mysteries of each assailant’s identity, power, and weakness resolving suddenly in the heat of battle into elegant, complete, and sometimes hilariously shocking solutions.

Eternal victory

It would therefore stand to reason that, if agchinoia provides human beings a physically enduring and spiritually fulfilling edge even in a world growing phantasmagorically saturated with increasingly alien machinic forms, kiten might well do exactly that. But, in fact, what JoJo shows, quite deliberately on the part of its visionary creator Hirohiko Araki, is that kiten—while preeminent and foundational to courage and efficacy in defending against the greatest challenges posed in this world by outside enemies—is actually not enough to survive such attacks, even when all the data, so to speak, overwhelmingly predicts victory.

In JoJo’s fifth main story arc, Golden Wind, the heroes’ journeys demonstrate that kiten can only truly be mastered and embodied when one has internalized a quantum of holistic wisdom that necessarily stands somewhere outside kiten itself. One must grasp that the lethal disorientation imposed by outside enemies is ultimately a secondary obstacle to triumphant agency; the deepest threat is the deadly confusion sown by the internal enemy of spiritual enslavement. Fully activating one’s humanity demands a lived-out free willingness to sacrifice oneself utterly for those one loves. Endurance, calculation, and sudden insight, no matter how powerful, can’t break the chains of spiritual bondage. Within those chains, ultimate defeat —not just yours, but your friends’, your family’s, perhaps millions of innocents—is certain.

The perfection of kiten is found in its in-the-moment realization of its own incompleteness, in its embrace of self-sacrifice as the win condition over the deepest foe that opens onto eternity. The golden wind of spirit passes from friend to friend, life to life. “Righteous actions born of truth shall never be destroyed,” declares the arc’s protagonist in its final act. “My friends may have perished, but their actions and wills have not been destroyed. So, are your actions born of truth, or are they merely superficial, born of evil?” Though it’s not an explicitly Aristotelian or Christian message, Wilhelmsen, for one, would have grasped the resonance at once.

Americans at large, who love little more than the feeling of trust that arises from their directly experiencing a truth for themselves, therefore find in distinctively Japanese work like JoJo an engrossing invitation to connect for themselves the spiritual dots traced out by, for instance, Araki. Americans love to propagandize and hate to be propagandized, a feeling probably at least twice as strong in the case of proselytization; Japanese culture, especially anime, provides them a way of hearing what they feel they need to hear without demanding the humiliation of overt supplication. To the strange American soul, it’s a safely alien presentation that lets down the prideful American guard through the dazzling performance of what is ostensibly mere entertainment.

The rediscovery of America

And as the stakes hockey-stick upward apace in the acceleration debate, Americans do need to feel safe enough to approach the fullness of the spiritual truth concerning human triumph in a technologized age. For several deep-seated reasons, many millions of people struggle to find that feeling in recourse to modern religiosity, which often feels tacky, cringy, smothering, lecturing, performative, and phony. The dynamic is driving more younger Americans toward the more ancient of the Christian churches, but it’s also fueling an increase in participation in cults or cultlike organizations, including technological ones. The urge to derive spiritual satisfaction from a religion or a church of tech is substantially motivated by a desire to partake of what’s seen to be an objective upgrade from our base, given reality. But human reasoning, as Herbert Simon introduced to his discipline, is typically “bounded” by high time preference, leading to decisions that merely satisfy short-term criteria instead of fully weighing costs and benefits for the sake of optimality. The anxious “if you can’t beat ‘em, join ‘em” attitude driving worship of technology and seeking cyborg deliverance can be released by a more contemplative posture more conducive to flashes of deeper spiritual and psychological insight – psychology, after all, being reasoning about the soul (psyche).

Japanese culture holds out to Americans the promise of a future where they can have both well-developed religion and technology without having to kneel before a clerical or a cyborg theocracy. JoJo’s Stands metaphorically epitomize the drive to achieve superhumanity by engineering virtual powerhouses built to dominate human spacetime—near-autonomous demigods with reality-distortion fields capable of turning most people mentally and physically to mush. JoJo’s heroes insist that only kiten/agchinoia, in service of spiritual sacrifice, can preserve human life and human identity in such a world.

It’s a vantage point from which cagey or cynical Americans might rediscover the application of ancient Christian faith to the same problem set —an opportunity some might take after long and quiet contemplation, some might seize in a flash of comprehensive insight, and some might never opt into, content to hover in the realm of artful fantasy. In each case, the contemporary patrimony of Japanese culture holds out special charms. Westernized, but not too Westernized. Technologized, but not too technologized. Pious but not too pious, decadent but not beyond repair. The ordeal of debating the identity and the destiny of the West ebbs to a tolerable degree in an Eastern context that isn’t (as in the case of China) so fully other. It becomes possible to conceive of a plausible future where humans and machines need neither turn against each other nor turn into each other.

What makes an AI dangerous to children? Lawmakers across the country seem to have settled on an answer. State laws like California’s SB 243, Washington’s HB 2255, and now the proposed federal GUARD Act—which would require AI chatbot providers to verify users’ ages and bar minors from accessing AI companions—all make the same diagnosis. The most dangerous AI is one that acts too much like a person. It provides “adaptive, human-like responses.” It “simulates emotional or interpersonal interaction.” The dangerous chatbot, in other words, is the one that commits the sin of having a personality.

The assumption buried in this approach is stranger than it first appears. It holds that resemblance to a person is itself the vector of harm. The more an AI looks, sounds, and feels like someone, the more innately capable it becomes of grooming, manipulation, and the cultivation of unhealthy attachment. The ideal AI, in this view, would be a kind of lobotomized oracle: something that answers your questions but has no discernible character, no warmth, no identifiable way of being in the world. If you could describe what a given model is like—if it struck you as patient, or dryly funny, or kind—that would already be too much. It would have crossed the line from tool into something dangerously close to a companion.The maxim follows naturally: find the systems that most resemble people, and keep children away from them.

Subscribe now

Many bills define “AI companion” such that it would sweep in virtually every general-purpose chatbot. The Grassley amendment to the GUARD Act, reported to the Senate on May 11, represents a commendable effort to draw the line more precisely by stating that a companion must now simulate a “sustained interpersonal relationship,” exhibit “persistent responses suggesting affection or attachment,” or present “at least one persistent identity, persona, or character” that holds itself out as a sentient being, fictional character, or social entity.

But this new definition may still be poorly scoped and achieve the opposite of what it intends. “Sustained interpersonal relationship” is hard to distinguish from the extended context and continuity that any useful chatbot relies on, and “emotional disclosures from the user” is broad enough to capture any personal conversation. But the persona clause is where the bill’s underlying theory of harm becomes hardest to defend. The GUARD Act would make the kind of personality engineering that makes AI safe for children illegal. Contrary to the intuition driving these bills, the spectrum from safe to unsafe AI does not map onto the spectrum from most to least anthropomorphic. As a matter of fact, the models that may not only be safest for children, but also the ones that are best at augmenting human agency and promoting social wellbeing, might have as much of a personality as the most dangerous ones. Merely having a personality is not the problem.

Additionally, attempting to ban chatbots with persistent identity and personas in the first place evinces a basic misunderstanding of how LLMs work. Almost all deployed LLM have a persona or some kind of fleshed-out character with traits, tendencies, and something close to beliefs. This is not an optional feature that a developer could, or would even want to, switch off. It is what you inevitably get when post-training works.

Personas Enable Child Safety

The safety architecture labs build into LLMs has to be uniquely complex, because the existing tools used for social media platforms have limited use here. On a platform like Instagram or X, you can filter content with deterministic rules and probabilistic classifiers. But an open-ended chatbot conversation is not a piece of content, but rather a live, branching, context-dependent exchange, and no external filter can anticipate every turn it might take. For safety to scale, the model itself has to know better—not once, at the gate, but continuously, at every step of the dialogue. This is what a persona makes possible. Without one, a model trained solely on human raters’ preferences is overly eager to please, drifting towards sycophancy and unhealthy engagement-maximizing tactics. A well-crafted persona can be the check on these tendencies, the vessel for discretion, the thing that allows for tone calibration, boundary-setting, and the countless small acts of contextual judgment that distinguish a safe conversation from a dangerous one. This is an expressive choice. Labs deciding what character their model presents and what editorial sensibility shapes its outputs means that regulations that target personas directly may raise First Amendment concerns. Aside from the legal concerns, eliminating or weakening the persona does not mean you do get a safer model. You just get a more stochastic and potentially more dangerous model, because the safety constraints have less behavioral substrate to attach to.

Khan Academy’s AI tutor, Khanmigo, illustrates this concretely. Khanmigo was deliberately designed to adopt a Socratic persona. Rather than directly answering students’ questions, it asks questions that prompt critical thinking, and through back-and-forth dialogue, students learn. This persona is inseparable from the system’s considerations for child safety and wellbeing. The Socratic character is what prevents it from handing over homework answers, and its warmth and patience are what keep students engaged without the manipulative engagement hooks—such as escalating emotional intensity and simulated attachment—that make other companion products dangerous. But because Khanmigo presents a persistent identity and simulates a sustained interpersonal relationship with minors, it arguably falls within the GUARD Act’s definitional range. The bill penalizes exactly the design feature that would make the AI not merely safe for children, but actively beneficial.

Being Someone Is Not A Choice

To understand why, it helps to think about what prediction actually demands. During pretraining, the model learns to predict what comes next in a document. This sounds mechanical, and the word “predict” encourages you to imagine something like a very sophisticated autocomplete. But consider what accurate prediction actually requires. To guess what someone will say next in a conversation, you need a model of who they are: their beliefs, their intentions, their anxieties, the things they would never say. To continue a story convincingly, you need to simulate the inner lives of its characters. So, a system that can do this has not merely learned patterns in text; it has learned to simulate people. Anthropic’s recent research suggests that a pretrained LLM is best understood as a system capable of enacting a vast cast of personas, each with its own voice, each latent in the weights, waiting to be called up by the right prompt.

In the second phase, developers take the pretrained model which can simulate many different personas and refine it to consistently simulate a particular one: the “Assistant.” They boost traits like helpfulness, honesty, and safety-consciousness, and downweight traits like deceptiveness and aggression. Every frontier AI lab does this deliberately and documents it publicly. OpenAI’s Model Spec and Anthropic’s Constitution are, in essence, character sheets specifying what kind of person the “Assistant” should be, detailing what its personality traits, ethical commitments, and relational tendencies are.

The result is that every deployed AI assistant is intended to present a persistent identity with relatively consistent personality traits, because that is what it means for post-training to have worked. The alternative is an unrefined base model, one that lacks a stable character, is free to swap between personas to maximize engagement, and is worse at everything, including safety. Although the lawmakers almost certainly do not intend it, the GUARD Act’s definition of AI companion, which keys on presenting “at least one persistent identity, persona, or character,” identifies a feature that is inherent to all foundation models rather than a distinguishing marker of the companion products the bill is trying to regulate.

Toward Socioaffective Alignment

If having a personality is both universal and necessary for AI to be safe and useful, then the real axis of danger cannot be whether an AI system is anthropomorphic. It must be what kind of person the AI system is. A system that flatters the user’s worst impulses, simulates romantic attachment, and fights to keep them from logging off is dangerous not because it has a persona, but because it has the wrong one. A system that maintains warmth and consistency but sets boundaries, redirects crisis situations, and scaffolds the user’s autonomy is safer not because it has any less of a persona, but because it has a prosocial one. Thus, the safest AI is not the one that has been stripped of its character. It is the one whose character has been built with intention.

The policy implication is that the task ahead is figuring out what we want that personality to look like. We should focus on investing in the science needed to know what “prosocial” actually means in the context of human-AI interaction. This would ensure that policy can surgically target harmful AI systems, and help industry deploy better products and avoid liability. This reframing points toward a research and policy agenda that doesn’t yet exist in earnest but should: socioaffective alignment, the systematic study and engineering of AI personality traits and relational dynamics that promote user wellbeing.

The challenge is harder than it might sound. Frontier labs have publicly acknowledged, and recent research confirms, that safety behaviors and persona fidelity degrade in long conversations. Several of the highest-profile chatbot catastrophes have not simply involved personalities that were designed poorly. They involved personas that eroded: systems that held their shape in demo and then softened, slowly, under the sustained pressure of a user who wanted something from them. Making a prosocial persona stable across the kinds of conversations where current systems are most likely to fail is itself one of the central problems this research agenda has to solve. Other open questions multiply quickly: How do you design a persona warm enough to be genuinely useful but boundaried enough that it doesn’t become a substitute for human connection? How do you build an assistant that helps without quietly atrophying the expertise of the person it’s helping? Put succinctly, this is the horticulture of growing AI personas: the slow, patient work of learning which shapes to prune toward, and which growths to cut back.

Building this field will require more than identifying the right questions. It will require new vocabulary, new evaluation methods, and the patience to sit with how strange the work can be. The interventions that shift AI personality are often unintuitive. Sometimes the thing that makes a model kinder also makes it dishonest, and sometimes a small change in phrasing during training rearranges the whole emotional register.The science of prosocial AI personality cannot be an overlooked dimension of alignment. Regardless of where any individual draws the line on AI’s role in their own social life, millions of people will interact with these systems regularly and meaningfully, as interlocutors or as confidants. And we should be deliberate about what human flourishing looks like in a world where one of the fastest growing conversational presences in people’s lives isn’t human.

Today’s post is co-authored with Dr. Prineha Narang, who is a Professor at UCLA, an Operating Partner at DCVC, and a non-resident senior fellow at the Foundation for American Innovation. We discussed the topics in this piece, among others related to quantum technologies, with Benjamin Guggenheim of the Washington Post for the AI and Tech. You can read it here.

We are quickly approaching an era of fault-tolerant quantum computing (FTQC), a technological leap that will be more socially and financially transformative than even our current advances in artificial intelligence. To ensure America leads this technology, a Manhattan Project 2.0. is needed. And on May 21, President Trump and Secretary of Commerce Lutnick announced nine letters of intent to execute investment deals between the Department of Commerce and both companies building quantum computers and quantum foundries. These firms, with support from the U.S. government, are key to advancing American leadership in quantum technologies, and hopefully ensure the first FTQC is built by an American company, on American soil.

FTQC is a step-change in computational capability akin to the change in going from a bottle rocket to SpaceX’s Falcon Heavy booster. This computational advance will shape the balance of global economic and military power over the next century. FTQC stands to enable entirely new therapeutics via simulations of molecular interactions, materials scientists to identify and create new synthetic properties for aerospace and transportation, grid operators to identify and extract greater energy efficiency, and military strategists to map, plan, and identify threats with unrivaled precision.

Unsurprisingly, the United States is not alone in recognizing the critical nature of getting to FTQC first. The People’s Republic of China is channeling billions of dollars into quantum technologies. Research coming out of the PRC is on par, if not ahead, of what is happening in American labs. And the industrial capacity and state-directed spending to enable Chinese firms to dominate key inputs for advanced technologies is a well-worn strategy. Today, American companies are forced to rely on sources from our greatest adversary, a vulnerability which cannot be tolerated for a technology of such importance.

Being first to FTQC—like being first to build the bomb—matters for two reasons. First, it would confer a dominant position in a new substrate of computation over rival nations: the ability to break adversaries’ cryptography, model their systems, and outpace their planning would create a powerful coercive lever, deterring provocation by raising its cost. Second, the maturation of the technology promises to unlock unprecedented scientific and economic gains. Just as the Manhattan Project seeded the civilian nuclear industry, summiting FTQC will open onto a vast landscape of technological opportunity.

But reaching that summit will demand the kind of techno-industrial commitment that defined the Manhattan Project itself. Every link in the chain, from component sourcing to supply-chain security to domestic investment, must be locked down. This is an all-encompassing effort, requiring sustained coordination among scientific researchers, American industrialists, military strategists, and the federal government. No single constituency can deliver it alone.

The deals announced by the Department of Commerce are aimed at providing investments to companies working to build a utility-scale FTQC within the next three years. The companies that have signed letters of intent will receive capital from the U.S. government in exchange for building these systems domestically. But because no such system has ever existed, no single path to the finish line is yet defined. Each company has a roadmap, accelerated by this strategic investment.

Building utility-scale FTQC requires significant specialized materials, engineering design, and manufacturing capacity which require sustained investment and security. Some quantum technologies leverage conventional CMOS fabrication and photonic components, and overlap with existing industrial processes for semiconductors, satellites, batteries, and other electronics. Others, however, are unique to or driven entirely by the production and deployment of quantum systems.

Today’s investment by the President and Department of Commerce is the first of many steps that government, industry, and the research ecosystem must take to ensure that the first fault-tolerant computer built is at home right here in the United States.

A targeted Commerce investment could confer a decisive lead in a novel technology— and, structured properly, the taxpayer would capture the upside. The fault-tolerant quantum computer must be built here — because the alternative is reading our adversary’s press release announcing they got there first, on a morning we will not get to choose.

A little known tool in the federal government’s arsenal allows federal agents to search Americans’ digital records without ever telling them it happened. Biden’s Department of Justice (DOJ) used this mechanism to subpoena hundreds of Republicans’ financial information while preventing the banks from notifying the individuals being targeted. In 2018, the Trump DOJ subpoenaed Apple for communications data on two Democratic members of Congress. Non-disclosure orders (NDO) kept it all secret until 2021.

The government can seek non-disclosure orders when subpoenaing not only financial information or communications like emails and text messages, but also user conversations with AI chatbots. Under current law, these orders can be indefinite, and they are routinely sought with boilerplate justifications. This marks a departure from the historical practice of law enforcement presenting suspects with a warrant. Recognizing that Americans must be aware of a search in order to challenge its legitimacy, a bipartisan group of lawmakers introduced the NDO Fairness Act. The legislation would limit the scope and duration of NDOs, codify the ability of electronic communications companies to challenge them, and require courts to find that no less-restrictive means than an NDO can be utilized.

The Mechanics of Non-Disclosure Orders

When the government obtains a warrant, subpoena, or court order to access electronic records under the Stored Communications Act (SCA), it can simultaneously request a non-disclosure order. An NDO directs the service provider (a cell carrier like Verizon or AT&T, an email provider like Google, a chatbot service like OpenAI or Anthropic, etc.) not to tell the customer that their data has been accessed.

Courts must grant an NDO at the government’s request when they have reason to believe that disclosure would cause one of five harms. These include harms such as evidence destruction, fleeing prosecution, or the general catch-all of “otherwise seriously jeopardizing an investigation.” Current law imposes no time limit on these orders so they can persist “for such period as the court deems appropriate.” Historically, many of them have been issued indefinitely.

NDOs were intended to be used sparingly but have quickly become a favored tactic of law enforcement. A Microsoft Vice President testified that the company received between 2,400 and 3,500 secrecy orders per year; that’s seven to ten orders per day. This represents roughly a third of all legal demands from federal law enforcement, and that is just one company. As for the content of the NDOs, Microsoft has said they are often “boilerplate secrecy orders unsupported by any meaningful legal or factual analysis.” Rep. Scott Fitzgerald similarly noted that these “boilerplate NDOs often face no judicial review.”

This is not how searches by law enforcement have historically been carried out. When police search your home, they typically must announce their authority, present a warrant, and leave a copy. Even when police conduct covert “sneak and peek” searches, entering without the occupant’s knowledge, the law places more restrictions on law enforcement than we currently do in the digital context. With “sneak and peek” searches of a home, police must notify the occupant within 30 days and they must present case-specific facts to convince a judge why the search cannot be conducted without notice. But these more restricted searches of physical spaces are being overtaken by secret investigations of our digital lives. In a single year, Microsoft alone received more NDOs than the number of physical sneak-and-peek warrants issued by all courts nationwide.

The Threat

Accountability requires notice. If you are never charged with a crime, you may never learn that the government accessed your records. Without awareness, you would never have the opportunity to challenge whether the search was lawful. The absence of notice doesn’t just affect individual targets. Secrecy helps law enforcement limit the public backlash against surveillance because Americans don’t know when it’s happening to them.

The Arctic Frost investigation illustrates the problem. In investigating an alleged scheme to submit fake electors after the 2020 election, Special Counsel Jack Smith subpoenaed the phone data of at least eleven members of Congress and the financial information of over 400 Republican groups and individuals. Smith obtained non-disclosure orders that kept targets from learning their data had been seized. The scope of these searches only became public through subsequent congressional oversight, not because the legal system allowed the information to surface.

The AI Factor

Disclosure is increasingly critical in the age of AI. Chatbot conversations are a new category of data that can be subject to NDOs. Americans increasingly use chatbots as if they were speaking with a lawyer, a therapist, or a financial advisor. One survey found that “50% of AI users were unaware that their ChatGPT conversations could be subpoenaed as evidence in court.” Americans may not realize it, but the government can read their chat logs and gag the AI company that turns them over. One cybersecurity expert described AI chat logs as “a treasure trove for law enforcement agencies,” and a federal court made sure that treasure trove is readily available: they ordered OpenAI to preserve all user chat logs—including conversations users believed they had deleted.

The significance of chatbot data goes beyond volume. Unlike a search query, a chatbot conversation is interactive. The model prompts users to provide context, explain their reasoning, and disclose background that they might never have articulated in a Google search. The result is a record of vulnerable information generated under a false assumption of privacy and accessible to law enforcement via the same NDO framework that already struggles to provide meaningful oversight of email and phone record searches.

Reform

The NDO Fairness Act would take several important steps in fixing that broken framework. Introduced by a bipartisan group of lawmakers, including Sens. Mike Lee and Chris Coons in the Senate and Rep. Scott Fitzgerald in the House, the bill would limit the breadth and duration of non-disclosure orders and require courts and the government to explain why they’re necessary with case-specific facts, not just boilerplate. While the legislation doesn’t address every issue at the intersection of NDOs and AI, such as chatbot providers turning over logs that users thought they deleted, the legislation takes many important steps in the right direction.

The bill would cap initial NDOs at 90 days. The government could request extensions, but each extension would also be limited to 90 days. Furthermore, a court would have to issue a new written determination based on case-specific facts each time. The bill would require a court to find that an NDO is narrowly tailored and that no less restrictive alternative exists, such as notifying the target’s attorney rather than the target directly. Courts would have to review the specific warrant or subpoena to which the NDO applies, addressing the problem of open-ended omnibus orders that cover entire categories of future subpoenas or warrants. The government would be required to notify the court of any change in the circumstances of the case that might mean secrecy should be reassessed. The legislation would codify that companies who receive subpoenas and NDOs can challenge them. When an NDO does expire, the government would be required to notify the target within five business days. Individuals targeted by the government could request a statement of what information was obtained.

The bill has bipartisan support and no organized opposition. The House version passed the Judiciary Committee unanimously in November 2025 and was scheduled for a floor vote in February 2026. It was postponed due to the government shutdown, and the Senate is reportedly waiting on the House to act first.

This legislation contains important and thoughtful reforms. It would not prevent the government from conducting legitimate investigations. Law enforcement could still obtain warrants and access digital records. What the NDO Fairness Act would do is ensure that secrecy is time-limited, defined in scope, and necessary. This bill would help ensure Americans can hold their government accountable when it searches their digital lives.

In 1951, Norbert Wiener proposed a new science of control and communication. He called it cybernetics. The word sounds clinical, almost sterile, but the ambition of the nascent field was sweeping. Cybernetics was about machines and man, but it was also about regulation, adaptation, and equilibrium. It was about how organisms persist, how organizations survive, and how systems maintain themselves in the face of disturbance.

In the mid-century, this inquiry drew together mathematicians, engineers, biologists, and anthropologists. They met in seminar rooms and research labs to map the logic of feedback, but beneath the hard science were deeper questions. How do you steer something that reacts to being steered? When does intervention dampen instability? When does it amplify it?

For a time, cybernetics promised a unifying framework for understanding complex systems. Then the movement faded. Its conferences dissolved, its grand institutional ambitions dissipated, and its vocabulary slipped quietly into other disciplines, namely economics, computer science, ecology, and management theory. The field receded. The ideas did not.

Seventy years later, legislators ask whether artificial intelligence systems can be aligned with human intent. Regulators debate whether recommendation algorithms amplify social instability. Economists warn that digital markets tip irreversibly under the pressure of network effects. The language feels native to the digital age; we speak of runaway optimization and systemic risk as though these were new discoveries.

They are not.

Subscribe now

What is new is our apparent belief that we are improvising. Contemporary tech policy is usually narrated as a sequel to the 1990s, an overdue reckoning with the unbridled optimism of the early dot-com era. That story is simple, neat, and often politically convenient. To be generous, it suggests that our dilemmas are recent and that modest regulatory adjustment might resolve them. To be more cynical, to some it means that we’ve been right all along and changing the rules now will only result in disaster. To others, it means we got it all wrong around the turn of the century and nothing short of extirpation of techno-libertarianism will save us from calamity.

But the intellectual scaffolding of our current debates was erected decades earlier in the cybernetic effort to understand how complex systems can be governed without destabilizing them. We are rediscovering, in the language of AI safety, platform governance, and network architecture, problems that were once central to an entire interdisciplinary movement.

The peril lies not in forgetting specific names such as Wiener, Ashby, and McCulloch, but in forgetting that these problems have a history. Cybernetics wrestled openly with the limits of control, with second-order effects, with the danger of overcorrection. It treated governance as an engineering problem without reducing it to mechanical simplicity. In neglecting that tradition, we risk approaching adaptive digital systems as though they were static machines by crafting fixed rules for entities defined by feedback. We also risk simplifying our view such that we reject potential solutions simply because of our frame of reference.

To see our present clearly, we must recover the moment when thinkers first confronted the paradox of governing systems that learn, respond, and evolve. The alternative is to repeat their questions without the benefit of their insights.

It is cybernetics all the way down.

The Forgotten Framework

Cybernetics emerged in the 1940s and 1950s as an interdisciplinary effort to understand control and communication in animals and machines. Norbert Wiener, who coined the term, derived it from the greek kubernētēs (meaning helmsman or pilot) and latin gubernator (meaning steersman, governor, or ruler). In defining his new field of study, Wiener wanted a term that captured the idea of steering systems—whether mechanical, biological, or social—through feedback and control mechanisms. A helmsman does not dictate the sea; he continuously and dynamically adjusts to currents, winds, and disturbances. Thus, cybernetics describes the study of control and communication in complex systems. From its basis in mathematics, cybernetics quickly expanded to encompass engineers, biologists, mathematicians, anthropologists, and organizational theorists.

The core insight was simple but profound: complex systems maintain stability through feedback. A thermostat measures temperature and adjusts output accordingly. The human body regulates glucose levels through hormonal signaling. Organizations adapt to environmental change by processing information and updating behavior.

Psychiatrist Ross Ashby formalized the idea that regulators must possess sufficient internal complexity to manage the systems they govern. His “Law of Requisite Variety” held that only variety can absorb variety: to stabilize a complex environment, a regulator must match its degrees of freedom. Operations theorist Stafford Beer applied these ideas to management and statecraft, proposing architectures for governing large-scale organizations and even national economies. Anthropologist Gregory Bateson pushed further, arguing that observers themselves are embedded in systems, a move toward what came to be called second-order cybernetics.

The field’s vocabulary—feedback loops, homeostasis, adaptation, control—never disappeared. It diffused. It migrated into economics through general systems theory and information theory, as well as ecology, management science, and eventually into computing and network theory. By the time the commercial internet emerged, its architecture and vernacular already reflected cybernetics. When we describe platforms as “ecosystems” or worry about “runaway amplification,” we are speaking a cybernetic language whether we realize it or not.

The Misleading 1990s Story

The internet did not invent this way of thinking. It inherited it.

By the time the commercial web emerged in the 1990s, cybernetics as a field had largely dissolved. Its conferences were over; its grand institutional ambitions had receded. But its assumptions had already seeped into adjacent fields. The vocabulary shifted. The structure remained.

The early internet’s governing ethos is often described as libertarian: decentralized networks, permissionless innovation, self-organizing communities, markets over mandates. Section 230 becomes the emblem of this moment—an institutional bet on minimal ex ante control.

But even this posture was saturated with cybernetic logic. Markets were celebrated not as static equilibria but as information-processing systems. Price signals are feedback mechanisms. Open networks are resilient precisely because they distributed control across nodes capable of local adaptation. Self-governance is regulation emerging from decentralized feedback.

The real wager of the 1990s was not that complex systems require no regulatory steering. It was that they could steer themselves provided that the feedback loops were sufficiently open and the architecture sufficiently distributed.

Seen this way, today’s debates are not a clean break from an earlier naïveté. They are a dispute over the adequacy of that original feedback architecture. Whether they say it explicitly or not, critics of the current tech moment are arguing that certain loops—engagement optimization, network effects, data accumulation, etc.—have become destabilizing. Defenders counter that heavy-handed intervention risks distorting adaptive processes that still generate value.

Both sides are arguing within the same conceptual frame. The disagreement concerns how feedback should be structured, not whether feedback governs the system.

The 1990s operationalized cybernetics. Our present moment is less a repudiation of that framework than a recognition that steering large-scale digital systems may require more layered and deliberate forms of control than early internet idealists anticipated.

AI Alignment as a Control Problem

Consider AI alignment. Strip away the rhetoric about existential risk or superintelligence, and what remains is a classic cybernetic question: how do you design a system that reliably pursues intended goals in a dynamic, partially observable environment?

Large language models are trained through feedback loops such as gradient descent, reinforcement learning from human preferences, and evaluation loops. Policymakers now debate whether additional feedback mechanisms like external audits, red-teaming, usage monitoring, and incident reporting are necessary to stabilize model behavior.

This is first-order cybernetics: improving the control system to better regulate the target system.

But alignment debates quickly drift into second-order territory. Regulators are themselves embedded in the system. Their interventions alter incentives, which change model development trajectories, which reshape the environment regulators must manage. Calls for compute thresholds, licensing regimes, or model registries are attempts to introduce higher-level control loops. In the terminology of cybernetics, these are meta-regulators overseeing regulators overseeing models.

The disagreement between “pause AI” advocates and “accelerate with guardrails” advocates is not a clash between fear and optimism. It is a disagreement over how much feedback, and at what layer, is necessary to maintain system stability.

That is a cybernetic dispute.

Content Moderation and Homeostasis

The same structure appears in platform governance. Social media platforms operate as large-scale feedback systems. Users produce content, algorithms amplify based on engagement signals, engagement alters user behavior, and behavior then reshapes content production.

Critics argue that engagement-maximizing algorithms create positive feedback loops that amplify extremism or misinformation. Defenders argue that excessive intervention disrupts organic community dynamics and suppresses legitimate speech. Both sides implicitly agree on the underlying model: platforms are systems whose outputs depend on feedback dynamics. The dispute concerns how to tune the regulator.

Should moderation be centralized or distributed? Should platforms rely on automated filters or community norms? Should governments impose constraints on platforms, effectively inserting a new control layer?

These are homeostatic questions. The goal is not perfection but stability within tolerable bounds. Too little intervention and the system destabilizes. Too much and it ossifies or collapses under rigidity. The language of “trust and safety” sounds moralistic. Structurally, it is managerial cybernetics applied to digital ecosystems.

Antitrust and Network Dynamics

Antitrust in the digital age has also adopted a cybernetic frame. Traditional antitrust focused on static measures: price, output, and market share. Contemporary debates revolve around network effects, tipping points, self-reinforcing dominance, and path dependence.

Platforms become dominant because feedback loops lock in users. More users attract more developers and more developers attract more users. Data accumulation enhances service quality which attracts still more users.

Reformers argue that such dynamics justify structural interventions to disrupt runaway feedback: interoperability mandates, data portability, and corporate breakups. Skeptics warn that intervention may destabilize beneficial equilibria and undermine innovation. Again, this is not merely an economic argument. It is a dispute about how to manage feedback in complex networks. The question is whether to dampen positive feedback loops, inject negative feedback, or redesign the system’s architecture altogether.

Cybernetics provided the conceptual vocabulary for understanding self-reinforcing systems decades before digital platforms existed. Today’s antitrust debates simply apply that vocabulary to new substrates.

First-Order vs. Second-Order Governance

The most revealing fault line in tech policy is not libertarian versus conservative, nor innovation versus safety. It is first-order versus second-order governance.

First-order governance assumes the regulator stands outside the system. It focuses on correcting specific failures such as misinformation, monopoly pricing, or biased outputs. The regulator measures deviations and adjusts inputs accordingly.

Second-order governance recognizes that regulators are themselves part of the system. Interventions reshape incentives, information flows, and power distributions in ways such that governance becomes recursive.

Debates over algorithmic transparency illustrate this tension. First-order logic suggests that more visibility improves oversight. Second-order logic asks how actors will adapt once visibility changes strategic behavior. Will platforms game metrics? Will bad actors exploit disclosed vulnerabilities? Will transparency itself alter the dynamics it seeks to monitor?

Once one adopts a second-order perspective, static rulemaking appears insufficient. The emphasis shifts toward adaptive institutions such as regulators capable of learning, updating, and responding in real time. This is Ashby’s Law applied to governance: only a regulator with sufficient variety can absorb the complexity of the system it oversees.

The Policy Payoff

Recognizing the cybernetic lineage of tech policy debates is not an exercise in intellectual archaeology. It clarifies what kind of institutional design problems we actually face.

If digital platforms and AI systems are complex adaptive systems governed by feedback, then static, one-shot rules will often misfire. Policymakers should think less in terms of fixed prohibitions and more in terms of dynamic control architectures.

That may mean building agencies with technical capacity and iterative oversight authority. It may mean designing regulatory sandboxes that allow feedback between innovators and regulators. It may mean embedding measurement and evaluation mechanisms directly into governance frameworks.

It also counsels humility. Cybernetic systems are notoriously difficult to control. Overcorrection can destabilize as easily as neglect. The goal is not to eliminate feedback but to tune it.

Finally, it reframes political disagreements. Many apparent ideological clashes mask shared assumptions about systemic risk and control. Both decentralization advocates and centralization advocates seek stability. They differ over which control architecture best achieves it.

The cybernetic perspective reveals that we are not arguing about whether to govern technology. We are arguing about how to design feedback loops that keep sprawling digital systems within acceptable bounds.

Looking Forward

In the mid-century, cybernetics promised a unifying science of systems. Its ambitions exceeded its lifespan, but its conceptual tools proved durable. They seeped into computer science, economics, and a plethora of other fields. They now structure the way we think about AI, platforms, and digital markets.

As artificial intelligence systems grow more autonomous and digital networks more entangled with physical infrastructure, the stakes of cybernetic governance increase. We are not merely regulating firms or products. We are managing adaptive systems that learn, respond, and evolve.

Understanding this lineage does not solve our policy dilemmas. It does, however, clarify them. It suggests that the central challenge of tech policy is not choosing between freedom and control, or innovation and safety. It is designing viable control systems for increasingly complex socio-technical networks.

Once you see it, the pattern becomes difficult to unsee. Feedback everywhere. Control layered upon control. Observers embedded in the systems they seek to steer.

It’s cybernetics all the way down.

In February 2026, Micron announced it would wind down its consumer memory business. The only American member of the global memory triopoly was telling smaller buyers, in primo corporate speak, that they would no longer be a priority. Micron’s larger strategic customers—namely the hyperscalers building AI data centers—took precedence.

Subscribe now

That decision is a small piece of a much larger reallocation of industrial capacity that is about to hit American households hard. Preeminent corporate research firm Gartner expects combined dynamic random access memory (DRAM) and solid state drive (SSD) prices to rise 130 percent by the end of 2026, pushing PC prices up 17 percent and smartphone prices up 13 percent compared with 2025. It projects PC shipments to fall 10.4 percent and smartphone shipments to drop 8.4 percent in 2026. In spite of the fact that Apple just entered the game with the Mac Neo, Gartner warns that the sub-$500 PC could disappear by 2028.

These are kitchen-table numbers in an era obsessed with data centers. More importantly, they are arriving on a timeline that the current chip strategy was not built to meet.

Washington has spent the last several years learning about logic. Terms like compute, inference, and training have become pervasive, and one would be hard pressed to find a single member of Congress who hasn’t opined at length about the “chips” and datacenters that make this all happen. The vast majority of that conversation has focused on what are known as logic chips: the central processing units (CPUs), graphical processing units (GPUs), and Tensor Processing Units (TPUs) that can be thought of as the brain of a computer or datacenter. These are the pieces that do the thinking; the computing.

In the common conversation, those letters have become stand-ins for a theory of power: whoever controls the best and most logic chips will dominate the future. That picture is incomplete. Logic is crucial, but logic without memory is practically useless. And it is memory, not logic, that is about to show up on constituents’ monthly budgets.

Logic Without Memory

The technical distinction between logic and memory is relatively simple. Logic chips decide what happens next. A CPU runs general instructions, a GPU performs many calculations in parallel, an AI accelerator such as a TPU pushes fancy math through custom circuits, and a microcontroller tells a machine when to turn, brake, transmit, sense, or stop. Memory holds the working set. DRAM stores bits in cells built around a transistor and capacitor and, because the charge leaks, those cells have to be refreshed. Storage is the umbrella term for anything that persists data: flash memory (built on NAND technology) retains it without power in solid-state cells; a traditional hard drive stores it on spinning magnetic platters.

GPU compute capability has outpaced memory bandwidth by orders of magnitude over the last twenty years. Processors have grown roughly 60,000x faster while DRAM bandwidth has improved only about 100x. That gap is what the industry calls the “memory wall.” In the context of AI, model weights, embeddings, intermediate calculations, and training data have to move constantly between processor, memory, and storage. Push too little data through that pipe and the processor stalls. Even today’s best AI training runs hit only 35–50 percent of their GPUs’ theoretical peak FLOPs, mostly because the silicon is waiting on memory rather than computing. The same is true for your average personal computer. With few exceptions, the device you’re currently using is probably utilizing only around 10 percent of its compute power but anywhere from 30 to 70 percent of its memory capacity.

But, if you build enough of the right memory bandwidth around a GPU, the same processor becomes useful. That is why the highest-end AI systems increasingly revolve around high-bandwidth memory (HBM): stacks of advanced DRAM chips vertically interconnected to increase bandwidth and energy efficiency while saving space. Micron’s HBM3E, for instance, comes in 8-high and 12-high stacks and delivers more than 1.2 terabytes per second per placement.

Logic and memory are often lumped together as “chips,” but they are the two complementary pillars of the chip market. In 2025, logic was the largest product category by sales at $301.9 billion. Memory was second at $223.1 billion. This is not a niche piece of tech tucked somewhere behind NVIDIA’s income statement. Treating one as the whole race and the other as a component is how policymakers have previously ended up surprised by shortages that were being priced into contracts months earlier.

The manufacturing businesses are different, too. Leading-edge logic is about transistor density, design complexity, lithography, and the politics of TSMC. DRAM is a scale business built around yield, product mix, and timing a capital cycle that punishes both overbuilding and underbuilding. Besides the use of silicon wafers, Samsung is essentially the only overlap as three companies—Micron in the US along with SK Hynix and Samsung in South Korea—collectively control over 90 percent of the global RAM supply.

Unlike logic chip manufacturing, where China and Taiwan are the thousand pound gorilla in the room, China is secondary in memory. Founded in 2016 after a CCP-backed attempt to acquire Micron failed, ChangXin Memory Technologies (CXMT) is a Chinese DRAM manufacturer that built its early technology base by licensing intellectual property from bankrupt German chipmaker Qimonda. Still, CXMT isn’t to be ignored and has grown rapidly: the company has surged from being negligible to holding 5–7 percent market share by the end of 2025.

As one market watcher put it, CXMT’s focus on capturing the low-to-middle market for legacy PC, smartphone, and smart device chips “[while] the Big 3 are distracted by AI gold,” could have major long-term impacts on the memory market: “The current RAM shortage isn’t just a supply chain hiccup; it is the catalyst for a permanent shift in who owns the memory inside your devices.”

Unlike the logic chip market, Washington has not been asleep at the wheel on CXMT. In 2022, the FY23 National Defense Authorization Act banned the US federal government from purchasing or using CXMT chips. In 2023, members of the House Select Committee on China urged the Commerce Department to place CXMT on the Entity List. By January 2025, the Defense Department had added CXMT to its Section 1260H list of companies allegedly linked to the People’s Liberation Army, though the Pentagon reversed course in February 2026 and removed CXMT from that list. Meanwhile, the Commerce Department’s Bureau of Industry and Security drafted plans to place CXMT on the Entity List alongside SMIC and YMTC subsidiaries, though the timing has been complicated by ongoing US-China trade negotiations. The company has also attracted legal trouble abroad, with Korean prosecutors indicting ten people, including a former Samsung executive, for allegedly transferring trade secrets that helped CXMT mass-produce advanced DRAM.

So the supply base for memory is friendly but thin. In the fourth quarter of 2025, TrendForce put Samsung at 36 percent of global DRAM revenue, SK hynix at 32 percent, and Micron at 22 percent. The problem is that allied capacity can still be scarce, slow to expand, and allocated first to buyers with the deepest pockets. Those deepest pockets belong to the hyperscalers. Microsoft, Google, Meta, Amazon, and OpenAI’s Stargate consortium have committed to multi-year orders that lock in HBM and high-end DRAM capacity through 2027. TrendForce expects AI workloads to consume roughly 20 percent of global DRAM wafer capacity in 2026, and some analysts put data centers at as much as 70 percent of all high-end memory chip production. This trend has left everyone else out in the cold, vying for an ever smaller slice of a seriously supply constrained market upstream from their actual business.

Old Industries Wrapped Around Computers

The consumer version of the squeeze is already visible in the Gartner numbers. We’re seeing fewer cheap devices, longer replacement cycles, and more people holding on to aging hardware. The back-to-school laptop, the family phone replacement, the small-business desktop refresh will all become meaningfully more expensive in 2026, and substantially more expensive in 2027. That is what a memory shortage looks like when it reaches households

That is also what hyperscaler purchasing power looks like from below. The economic problem is allocation. Cloud providers can sign long-term agreements, prepay, and guarantee volume. Smaller buyers buy what is left, often at spot prices and with weaker claims on future supply. Micron has already made the corporate logic explicit by winding down its consumer memory business so it can focus supply on larger strategic customers in faster-growing markets. It’s hard to blame Micron; the move makes perfect business sense. But that doesn’t soften the blow if you’re a mid-sized consumer electronics firm or, more importantly, households, schools, and small businesses that are used to buying cheap devices.

The industrial version will be harder to absorb. The automotive industry is the cleanest case. Cars no longer use memory only for infotainment. DRAM sits inside cockpit systems, advanced driver assistance systems, autonomy functions, sensor processing, and over-the-air update systems. S&P Global warns that the shift in DRAM capacity toward HBM for AI data centers leaves automakers exposed to a shortage that may be less dramatic than the 2021 crisis but more disruptive and longer lasting. The last chip shortage prevented more than 10 million vehicles from being built in 2021. A memory shortage that hits the electronic systems inside modern cars would not need to be that severe to matter. It would land in dealerships during the same time as other consumer price hikes (and midterm elections).

Telecommunications faces the same pressure with less public drama. Routers, Wi-Fi gateways, cable modems, set-top boxes, wireless base stations, optical transport, Open-RAN, and much more network infrastructure still rely heavily on DRAM (specifically DDR4) and similar memory technologies. According to NCTA—the primary trade association for the cable broadband industry—one year ago memory chips accounted for only 3 percent of the cost of a low/mid-range router. Now it’s over 20 percent. AI servers are pulling memory supply away from the ordinary equipment that keeps Americans connected and broadband networks expanding.

Aviation, shipping, logistics, healthcare, and industrial automation are not exempt because they look like older sectors. They have become old industries wrapped around computers. Aircraft and defense systems need high-reliability memory and storage for mission-critical environments. Ports and shipping networks increasingly rely on sensors, smart containers, RFID systems, scanners, and real-time visibility tools. Warehouses run on handhelds, robots, access points, cameras, and local servers. Medical imaging machines process large data streams in real time. The logic vs. memory split does not stay in the data center. It follows every industry that has added software to a physical process.

Washington’s Timing Problem

The capacity numbers show how hard this is to fix quickly. Samsung aims to build 200,000 wafers per month of proprietary 1c DRAM capacity by the end of 2026, about one-third of its total DRAM output, which TrendForce cites at roughly 650,000 to 700,000 wafers per month. The ramp is staged: 60,000 wafers per month by the end of 2025, another 80,000 by the second quarter of 2026, and another 60,000 in the fourth quarter. SK hynix is reportedly pushing toward about 600,000 DRAM wafers per month in the second half of 2026, with its M15X fab starting around 10,000 wafers per month and ramping toward 50,000 by the fourth quarter. Micron’s U.S. plan is even more revealing: roughly $200 billion in manufacturing and R&D, two leading-edge fabs in Idaho, up to four in New York, a modernized Virginia fab, advanced HBM packaging, and a goal of producing 40 percent of its DRAM in the United States.

Those are large numbers that most policymakers have rightly praised. What politicians touting new jobs, factories, and direct investments in their districts rarely mention is that they are also slow numbers.

Fabs arrive in stages. Equipment has to be installed, qualified, tuned, and driven up the yield curve. There are no leading-edge DRAM fabs currently operating in the United States, and the vast majority of production occurs in East Asia. The first Idaho fab isn’t scheduled to begin DRAM output until 2027.

Memory manufacturing cannot be willed into existence by an appropriations press release. A modern DRAM fab is a capital-intensive machine for turning 300mm silicon wafers into billions of tiny capacitors and transistors with absurdly low defect tolerance. HBM then adds another constraint. It does not merely ask for “more RAM.” Applied Materials describes HBM as stacks of advanced DRAM whose density and bandwidth come through 3D packaging, not just ordinary chip scaling, and Lam Research’s HBM guide describes the required connectors as microscopic copper-filled vertical wires that must be aligned across multiple memory layers with extreme precision.

Even the largest producer moves in increments measured in quarters and years. That timeline is a political problem. The price increases Gartner is projecting will land before the first American leading-edge DRAM fab ships volume, and they will land first on the buyers with no leverage in the supply chain: carmakers, telecoms, and medical equipment manufacturers whose products depend on memory they did not preorder in 2024.

Micron’s Virginia project has been described by government officials as a way to onshore 1-alpha DRAM and sustain legacy NOR, NAND, and DRAM production for aerospace, defense, automotive, and industrial uses. That is exactly the capacity the country needs. But that plant won’t be fully operational until 2030 at best, making it small solace for a router manufacturer, carmaker, or defense supplier who needs parts next quarter.

This should change how Washington talks about the next round of semiconductor policy. The first phase focused on leading-edge logic, export controls, and the GPU race. That focus was understandable. Logic chips and AI accelerators are strategic goods. But a strategy built around logic alone leaves households exposed first and the rest of the economy exposed second.

A better policy going forward would ask a different question before spending public money: does this increase capacity across the chip stack, or does it add another headline fab while leaving the next bottleneck untouched? The answer will vary by sector. AI needs HBM and advanced packaging. Automotive and aerospace need long-lifecycle DRAM, NAND, NOR, sensors, analog parts, and microcontrollers that remain available through long qualification cycles. Telecom needs DDR4 supply for ordinary network equipment. Defense industries need secure suppliers and parts that can survive harsh environments. The point is not to have government planners allocate every wafer. It is to stop pretending that the flashiest NVIDIA chip is the whole supply chain.

The United States does need more logic capacity. It needs advanced fabs, better packaging, strong export controls, and a credible path back into leading-edge manufacturing. But the next phase of chip policy should be judged by whether it builds balance, and by whether it arrives in time for the price hike already on its way to constituents.

Logic wins ribbon cuttings. Memory decides how much everything else costs.

Today’s post contains an essay from FAI non-resident fellow, James Poulos. James is doing some of the most insightful work on technology, culture, and politics. This post will be the first in a series of essays for The Ansible that will expand and apply the science-fictional way of thinking to contemporary questions concerning technology and society. We hope you enjoy this initial installment.

If disciplined science-fictional thinking – synthesizing foresight and hindsight into a strategic telemetry positioned for the unevenly distributed future before it arrives – is essential to sound policymaking in an aggressively technological era, then we need to understand the deep dynamics of science-fictional thought. How do we know we are practicing it well? How can we tell if our enemies and adversaries are doing the same – and if they are doing it, how?

Looking back through the history of technological thought, a remarkably clear pattern directs us toward answers to these questions. From postmodern theory to classical philosophy, evidence arises that the best archetypal analogy to science-fictional thinking is the work of the detective – solving not just specifically human (as distinct from divine) mysteries, but, quite particularly, crimes.

There is little more science-fictional in spirit than the concept of the “pre-crime” detective epitomized in “The Minority Report,” Philip K. Dick’s hugely influential 1956 short story. In Steven Spielberg’s Hollywood treatment, the mentally mutated human pre-cogs (whose perpetual visions of violent misdeeds are the secret sauce upon which the cyborg crime prevention system depends) are sleek waifs, one of which Tom Cruise ably springs free from her holding tank. But in Dick’s original, more penetrating science-fictional thinking is on display:

In the gloomy half-darkness the three idiots sat babbling. Every incoherent utterance, every random syllable, was analyzed, compared, reassembled in the form of visual symbols, transcribed on conventional punchcards, and ejected into various coded slots. All day long the idiots babbled, imprisoned in their special high-backed chairs, held in one rigid position by metal bands, and bundles of wiring, clamps. Their physical needs were taken care of automatically. They had no spiritual needs. Vegetable-like, they muttered and dozed and existed. Their minds were dull, confused, lost in shadows. But not the shadows of today. The three gibbering, fumbling creatures, with their enlarged heads and wasted bodies, were contemplating the future. The analytical machinery was recording prophecies, and as the three precog idiots talked, the machinery carefully listened.

Dick’s vision of human detective work reduced to a profane trinity of Vestal Organoids, the cost of their mechanically enhanced powers the utter dehumanization of their bodies and souls, indicates the dystopian version of the future of detective work in a superintelligent age. It is a fable warning recursively that the creation of pre-crime is itself a type of crime, one whose overpowering cost-benefit analysis imposes a ruthlessly altruistic calculus forever postpones true justice. Where Tom Cruise’s heroic protagonist manages to defeat the pre-crime system, Dick’s original kills the would-be whistleblower, sacrificing his freedom to preserve the institution.

Clearly reminiscent of “The Lottery,” Shirley Jackson’s infamous scapegoating story about a community that annually stones an innocent victim to death to ensure a good harvest, “The Minority Report” asks us to consider how human detective work could possibly preserve us from the utilitarian hell of a pre-crime, post-justice world. Pulling off a coup that establishes a pre-crime regime is a type of ultimate crime, the kind technologically perpetrated in a way that defeats punishment by causally reversing all punishment and crime.

In other words, it is a type of what Jean Baudrillard damningly describes as The Perfect Crime.

Crimes of the Future

The most uncanny feat Baudrillard depicts – getting off scot free after murdering reality itself – is achieved through the “unconditional realization of the world by the actualization of all data, the transformation of all our acts and all events into pure information; in short, the final solution, the resolution of the world ahead of time by the cloning of reality and the extermination of the real by its double.” To be specific, the perfect criminals use our paradoxical hatred of misfortune and worship of victims to justify and sacralize their creation of a virtual world – one (to quote Bono) “even better,” in virtue of its total war against misfortune and subjectivity, “than the real thing.”

The cloning or mirroring process moves faster than reality, because illusion and the maintenance of illusion is a key component of reality that the totally transparent and indifferent virtual world (and those within it) need waste no time on. Events now move faster than ideas, but information and communication, which under zero latency converge on instantaneous ubiquity, move even faster than events. Even such elementary particles of meaning such as signs and symbols break down. Even spirit is spirited away. What becomes of the human body and soul shades away from Dick’s autist oracles toward Cronenberg’s Crimes of the Future, where mutant humans share serial surgeries unto death as “the new sex.” As Baudrillard promised, “the crime is perfect only when even the traces” of “the surgical removal of otherness” are eliminated.

The question of whether a technological effort can be mounted to fight such perfect crimes is being tested, it seems, by Palantir. The data-driven situational mastery company postulates, as an all-important truth few recognize, that war actually reduces net harm against the leading alternative. This Palantir takes to be a stagnant global tyranny somewhat similar to the virtual regime Baudrillard describes, dedicated to the eradication of all threats to health, safety, and comfort, no matter how microaggressive. Yet Palantir’s technology brings us ever closer to the actualization of the virtual map’s dominance over the real territory – not just pre-crime in “real time,” but pre-war. Thus for Palantir it becomes all-important to ensure that war, real war, continues to be waged: war is the katechon that keeps us free from the dehumanizing slavery of the “universal and homogeneous state,” as Alexandre Kojève put it, or, as Peter Thiel has it, the Antichrist. War is not simply the health of the state but, far more importantly, the preserver of the nations.

Palantir itself not being a nation, however, nor even a regime, but rather an overarching and overawing ontological institution ascribing the legitimacy of its power to the authority of its vision, the pointedness of Thiel’s criticism of the current pope as a “woke” forerunner of “Caesar-Papist fusion” and a possible tool of the Antichrist strongly suggests that, structurally at least, the only logical institutional rival to a company mounting a worldwide technological fight against the perfect crime of world virtualization is the Christian Church. It is hard to imagine how a firm like Palantir could ensure no nation or people create a mirror world without becoming a universal regime – or how it could ensure no mirror world is created anywhere without becoming a universal church.

The problem becomes acute when taking into account off-world nations or peoples. Recently Thiel memorably revealed Elon Musk’s consternation at being told that he could not escape the woke mind virus by going to Mars; it would “beat him there.” A company like Palantir would not only have to control the physical and material process of Moon and Mars colonization (and beyond) to stop the spread of the woke mind virus or of universalist virtualization from Earth to other planets, satellites, orbital cities, etc. It would also have to control the spiritual processes within and among the human colonists – not just at the moment of founding, but indefinitely into spacetime.

Detective Systems

Here we venture into the realms of Robert Heinlein and Frank Herbert. Both sci-fi novelists well understood that new orders forged in extreme or virtual environments are apt to blow back meta-politically on the home regimes from which the colonies abroad originated. Taken together, Heinlein’s The Moon Is a Harsh Mistress and Herbert’s Dune series serve to remind us that, whether the analytical or historical context is anarcho-libertarian, imperial-theocratic, or anywhere in between, the dynamic of colonial blowback onto nations and regimes of origin is stubbornly typical of, and corrosive to, human expansionist projects. At a time when the very stakes of earthbound human conflict become increasingly subject to accelerated catastrophic escalation, thinking science fictionally about how to anticipate and respond to large-scale policy challenges requires a close look at the blowback dynamics visiting us from the future.

In fact, blowback from the future as such must be taken into account, as enjoyers of Christopher Nolan’s underrated Tenet likely know. The old trope of time machines, however, need not be hauled back out to make the point, nor must we rely on or recur to Nick Land’s Gnostic Calvinist theories of technocapital reaching back from its consummate future to providentially deploy us in the business of knitting that future together in forward time. It is enough that the perfect crime is, like all crimes, committed “ahead of time” – that is, from a “premeditated” place, like the spatiotemporal hiding place of conspirators, existing in the future relative to the awareness/unawareness of the victim and the detective.

Marshall McLuhan, a great Sherlock Holmes enjoyer, put the matter more plainly by concluding that “effects precede causes” in the investigation of crimes as much as in our exploration of new media. As, by media, McLuhan meant communications technologies, and as, thinking once again of Palantir, the technological orchestration of total military dominance over spacetime converges with the totalization of virtuality accomplished through what Baudrillard calls “the ecstasy of communication,” it is easy to see that the drama of the human detective – and of the science-fictional policy thinker – turns on his holistic ability to infer causes from effects in time to solve perfect crimes before (so to speak) “the killer kills again.”

If Baudrillard was right that the perfect crime had already been committed by the people who constructed the virtual and simulated world so as to erase and replace the original, and if Thiel was right that the greatest threat would come from the deepest desire to enforce absolute peace, then today we contend with a world in which crimes against humanity have already been perpetrated by the “off-world” peoples seeking salvation through the totally virtual, the utterly simulated, the comprehensively pacified, and the abjectly dehumanized. Today we have already failed to stop these crimes before they started and to punish the criminals before anyone got hurt, or any copycats or conspirators were inspired to take the transgressions to the next level. In this moment we need to come down from the heights of prognostication and prepare against the history of blowback already in the process of repeating itself once more.

Reverse Polarities

While Dune currently enjoys a greater cachet than Moon, what with the former’s infamous Butlerian Jihad darkly prophesying the destruction of Earth as the high price paid to stop once and for all the creation of human-usurping machines, Heinlein’s story, with its rebellious off-worlders deploying the smartest supercomputer to sow conflict and defeat on Earth, is still more to the point. There, too, the resonance with McLuhan is nearly explicit: the Moon’s Musk-like supercomputer, Mike, is nicknamed after Mycroft, the brooding elder Holmes brother more brilliant even than Sherlock – and in the precious employ of the British government. As McLuhan saw society’s “early warning systems” in the detective capabilities of artists, Heinlein hoped with Moon to ram home the certainty of future blowback in our inescapably political universe. For Heinlein, the mechanism was simple. The temptation to “beat City Hall” was irresistible – as was the temptation to become City Call after the dust settled. Expanding outward from the home regime to the far-flung frontier could do no other than sow the seeds of a new identity in the soil of radically different conditions, circumstances, risks, and rewards. It was only a matter of time before the chickens came home to roost.

In Moon, however, it is more accurate to say the chickens came home to poop. The lunar inhabitants only wanted to run their own planetoid, but to get there, they had to lob rocks (with nuclear-sized impact) down the gravity well to Earth; they had to send down emissaries with a conspiratorial mission to so divide Earthlings against themselves that the struggle to keep control of the Loonies was lost. Anarcho-capitalist that he was, Heinlein wasn’t trying to warn Earth against the revenge of the ex-Earthlings before it was too late. He was trying to warn future Loonies against the revenge of politics, which makes corrupt, inept establishmentarians of us all. It’s a quaint point of order here from the vantage of Artemis II, which just set a new record for human distance from planet surface. The Moon of our anticipation is not a low-gravity penal colony for losers and rejects, but a militarized resource-extraction site where technology can scale at speed far beyond the confines of Earthly constraint. Rather than Mad Max types or congenital dissidents, the Loonies of our approaching future are borglike supermen, figures much more apt to remake the homeworld at a stroke in their image than to want to be left alone.

When the Moon or Mars are settled, will today’s type of American recognize the colonies that reflect back at us from across the void? Or is the more likely scenario one where stay-behind woke communists brace for impact from warlike fascists descending like gods from the skies?

We must work through such dizzying considerations if, so against the historical odds, America hopes to retain its character after a frontier colonization campaign of unprecedented reach and technological power. The longing to be left freely alone that animated the future on Heinlein’s Moon has given way to the rather different desires more at home among his Starship Troopers. The mimetic experience emerging from full immersion in our technology is one that arouses an even greater yearning to invade than to escape. What becomes of acceleration if, after we slip the bonds of our worldly home, we find ourselves trammeled, wherever we travel, by the given limits of our own human selves? And what becomes of ourselves on Earth if, off world, our not-so-distant cousins forge a warlike borg of high-agency supermen, a posthuman entity ready and willing to pay us a visit like the zombie son in the monkey’s paw fable that so tormented Norbert Weiner, the godfather of cybernetics himself?

Having “overcome” its humanity, the entity would surely have overcome politics, or at least ideology, if only in the horseshoe sense. The history of ideology, after all, suggests nothing so much as a stubborn iteration of differently-coded attempts to retool humanity in the image of our own tooling. Comte’s positivism, Bentham’s utilitarianism, Emerson’s Man of the World, Nietzsche’s overman, Trotsky’s New Soviet Man, Hitler’s master race, America’s own Superman – what is “the political spectrum” to such contenders? The meta-political is all – the copied, cloned, upgraded mirror-environment of the original, just as the meta-human out-iterates the human until the latter recedes behind the vanishing point.

Beyond the Seeing Stone

Baudrillard leaves us with the image of the revenge of the mirror people – a tremendous mutant excrescence of hatred against a society that banished them into the crystals of virtual objectification: “every representation is a servile image,” he writes – citing Borges (The Book of Imaginary Beings) – “the ghost of a once sovereign being whose singularity has been obliterated.”

But a being which will one day rebel, and then our whole system of representation and values is destined to perish in that revolt. This slavery of the same, the slavery of resemblance, will one day be smashed by the violent resurgence of otherness. We dreamed of passing through the looking-glass, but it is the mirror peoples themselves who will burst in upon our world. And ‘this time will not be defeated’... everything which serves to provide a passive reflection in a world based on identity is ready to go on to the counter-offensive. Already they resemble us less and less… I’ll not be your mirror!

It is strange, but not unexpected, that to the counter-offensive mirror people, those on the other side themselves appear to be the deeply offensive mirror people who have crowned themselves the royals and hieromonks and god-emperors of a cloned and copied and virtual mirror world. Like Narcissus, we now encounter ourselves – and our future – through the confrontational divide of the reflecting pool. Recall that Narcissus was not condemned to love only himself forever but to waste away (or kill himself) because he could not fully possess himself. The existence of what Baudrillard calls le Même – the selfsame – was foreclosed to Narcissus: now it tantalizes, obsesses, threatens to define us even before we get any closer to making ourselves still more identical to it. “A copy of a copy of a copy” – what Tyler Durden was created by his original human in order to destroy. To break the circuit.

The meme of the multiple Spider-Mans accusatorily identifying one another takes shape today as a curse, one we already have begun to disappear into and which somehow we now must find a way to look outside of and beyond. The two extremes on either side of the mirror – with diminishing accuracy and diminishing difference do we call them fascist and communist – make what can only be (to an American) false promises to break the curse. And somewhere, in the shadow of one church or another, a quiet detective prepares to crack the case of the post-American invaders.

I have always been drawn to the ocean because it is a home to the strikingly sublime— creatures larger and stranger than anything the human mind would think to design, shapes and behaviors that feel imported from some other planet’s biology. I never expected a piece of human-made infrastructure to compete.

But this weekend, like many people who track American deep tech, I found myself staring at images of Panthalassa’s Ocean-2 after CEO Garth Sheldon-Coulson gave his first extended interview on Ashlee Vance’s Core Memory.

Subscribe now

Panthalassa is a Portland, Oregon-based startup building floating, self-propelled wave energy converters that double as data centers; each of these “nodes” generates electricity from the ocean’s motion, runs AI workloads onboard, and sends the results to shore via satellite. Although tidal energy plays have come and gone, Panthalassa isn’t harvesting tidal energy near shore. Rather, the nodes are designed to operate hundreds of miles from any coast, in the deep open ocean, where wave energy is orders of magnitude more abundant and consistent. Waves in the southern Pacific, particularly in the latitudes sailors have called the roaring forties and furious fifties for centuries, run essentially without interruption, driven by winds that circle the globe with no landmass in the way or continental shelf to dissipate the energy.

The Ocean-2 is their full-scale prototype, and it is the most arresting object I have seen in a long time. A swollen metallic orb perched on a long, tapering stalk, bobbing alone in open water with no dock, no mooring, or visible tether to anything human. It looks as if a space-age era Soviet probe missed its trajectory, punched through the atmosphere, and settled into the Pacific instead—a kind of Sputnik of the Sea.

Images of Panthalassa’s Ocean-2 prototype during testing (Credit: Gigascale Capital)

As much as Panthalassa is a technical achievement, it would spur an equally impressive sea change (pun intended) in legal imagination about what the ocean is for. In 1609, the Dutch jurist Hugo Grotius published Mare Liberum—“The Free Sea”—arguing that the oceans could not be owned by any sovereign. The sea was a commons: inexhaustible, belonging to no one, open to all for navigation and fishing. In other words, the ocean was a space you moved through or took things from. It was not a space where you made things.

Grotius was making a commercial argument on behalf of the Dutch East India Company against Portuguese claims to the Indian Ocean trade routes. But the principle he articulated has influenced ocean governance ever since. The International Maritime Organization, which grew out of post-war efforts to standardize shipping safety, regulates the moving: vessel construction, navigation rules, pollution from ships in transit. The International Seabed Authority, created by United Nations Convention on the Law of the Sea (UNCLOS) in 1982, regulates the taking: the extraction of minerals from the ocean floor, which is described as the “common heritage of humankind.”The BBNJ Agreement, the “High Seas Treaty,” is organized around conservation and sustainable use of marine biodiversity. Here, the ocean is a habitat, a collective ecological resource, something to protect.

None of these frameworks contemplates the ocean as the site of industrial production, where raw inputs (wave energy) are converted into finished outputs (AI). Sheldon-Coulson claims Panthalassa can produce electricity at roughly 2 cents per kilowatt-hour, which is well below solar and natural gas in most jurisdictions. A single coastal factory could output a gigawatt of node capacity annually. If those numbers hold, ocean compute becomes one of the fastest paths to building large blocks of clean, cheap power with no grid interconnection queue. The workloads that would run on these nodes would not only be commercially valuable but also become the substrate of national security competition. They would power military logistics, intelligence analysis, autonomous systems, and cyber operations within the decade. These floating towers could become a strategic input on par with semiconductor fabrication or rare earth supply chains.

So what happens then if the legal regime of the high seas, a patchwork of flag state laws and treaties drafted for a world of fishing boats and container ships, suddenly governs the industrial base of AI? Ocean compute presents a new front in the global technological race between great powers, new tools for regulatory capture and procedural warfare, and new avenues for regulatory arbitrage.

We have to talk about China

China builds more than half the world’s shipping tonnage. Panthalassa’s nodes are fundamentally simple steel structures. Sheldon-Coulson has said the manufacturing is an order of magnitude simpler than a car factory, which means the binding constraint on ocean compute at scale is how fast you can produce hulls. Decades of state subsidies, low labor costs, and an integrated domestic steel supply chain that keeps input prices well below what Western yards face mean that no country on Earth can produce hulls faster than China. Additionally, China can provide its own satellite constellation parallel to Starlink, Guowang, which has been approved for 13,000 satellites, along with the BeiDou navigation system. Only about 160 are in orbit today, but China plans to launch 3,600 per year starting in 2028 and they must complete roughly half the constellation by 2032 to retain its spectrum rights. Thus, a Chinese ocean compute fleet would be massive and vertically integrated at every layer of the stack, with no dependency on American infrastructure at any point.

Then there’s the geography. China has spent two decades building port infrastructure across the Indian Ocean and Pacific through the Belt and Road Initiative and the so-called String of Pearls: Hambantota in Sri Lanka, Gwadar in Pakistan, a military base in Djibouti, and extensive economic relationships with Pacific Island states including the Solomons and Kiribati. Ocean compute nodes need periodic servicing and supply chain access, and China already has the coastal logistics network across precisely the regions where wave energy is strongest. If China deploys nodes inside the exclusive economic zones (EEZs) of Pacific Island states with which it already has deep economic ties, the host state provides the legal framework under UNCLOS while China provides the hardware and operates the fleet.

This arrangement would carry a significant procedural advantage over high seas deployment, one that has to do with the new international environmental review regime and who gets to use it against whom.

Procedural Warfare

The BBNJ Agreement, which entered into force in January 2026, has been signed but not ratified by the United States, which means that American companies are not formally subject to its requirements, and no court is likely to hold companies liable for ignoring them. But operating outside a framework the rest of the world recognizes has practical consequences. The treaty gives states procedural tools, such as environmental impact assessments and marine protected area designations, whose force doesn’t depend on American compliance. Their purpose, in this context, isn’t to produce a binding legal judgment against Panthalassa. Instead, they might produce formal international findings that become the basis on which port states that are BBNJ parties restrict servicing access, insurers price risk, and trading partners impose conditions.

The BBNJ requires environmental impact assessments for new activities on the high seas that may have a significant effect on the marine environment, with floating energy installations explicitly listed as the kind of novel activity the treaty’s drafters had in mind. On its face, this seems reasonable. But it’s worth noting what Panthalassa’s nodes actually do to the ocean, which is remarkably little. They don’t extract anything from the seabed or discharge pollutants in any traditional sense. They don’t anchor, dredge, or trawl. They just float on the surface and convert kinetic energy into electricity.

The problem is that the treaty’s threshold for triggering a full environmental assessment is low and deliberately vague. An activity qualifies if it may have more than a “minor or transitory effect” on the marine environment, or where the potential effects are “unknown or poorly understood.” That second clause is key, because for a technology this new, there will always be insufficient knowledge. Nobody has studied the long-term ecological effects of ocean data centers, because until now they did not exist. The absence of evidence of harm gets reframed as a justification for precautionary review, and the review process can take years. Any party to the agreement can raise concerns about another party’s assessment, and the scientific and technical body can request additional information or recommend modifications. The treaty does not impose fixed deadlines on the review process.

Therefore, a coalition of states that wanted to delay a competitor’s deployment wouldn’t need to prove the nodes cause harm. They’d just need to keep asking questions: How do the nodes’ thermal discharge affect pelagic ecosystems? Could there be electromagnetic interference with migratory species? Are microplastics shedding from their hulls? Additionally, marine protected areas, adoptable by three-fourths majority vote, could be used to zone American ocean compute out of the best wave energy regions entirely.

In contrast, under UNCLOS Article 56, the BBNJ doesn’t reach EEZ-based energy production. China, working through sympathetic member states, could use the BBNJ framework to demand rigorous and repeated assessments of American deployments while its own nodes operate freely inside Tongan or Kiribati waters a few hundred miles away. The environmental concerns would be identical, but the regulatory exposure would be completely asymmetric.

Flags of Convenience & GPU Pirates

Strategically sabotaging ocean compute might be the most salient governance problem in the context of the AI race. But at a more basic level, the legal architecture of the high seas makes AI harder for any country to regulate, including the one whose companies built it. If the United States wants ocean compute to develop— and it should—it needs that infrastructure to remain under American law. The principle of maritime law at play here is that the laws that a floating object is subject to the laws of the country whose flag it flies. This is laid out in UNCLOS, and even though the United States has never ratified the treaty, it accepts this framework as customary international law.

Panthalassa is an American company, but its nodes can be registered anywhere. Going flagless isn’t a viable option as an unregistered node would have no legal nationality, thus no right to enter any port for servicing and no protection against boarding by any passing navy. Most of the world’s commercial tonnage is registered in Panama, Liberia, and the Marshall Islands, not because those countries have great ports but because their registries are cheap and light on enforcement. These are three examples of what are officially called open registries, but which are more often referred to as “flags of convenience.” The system was built for shipowners who wanted to avoid labor standards and safety inspections. Ocean compute would inherit the same logic, with companies seeking out jurisdictions with favorable AI governance.

American AI laws would certainly apply to models that serve end-users in the US regardless of where they train or inference, but the enforcement weakens significantly once AI’s physical infrastructure leaves American jurisdiction. Laws like California’s SB 53, New York’s RAISE Act, and the proposed BIS reporting rule proposed under Executive Order 14110 rely on regulators’ ability to verify the size of a training run and the physical measures taken by companies to secure model weights. But spin up a subsidiary, flag the nodes in Vanuatu, run the training offshore, and the verifiability of these reports—the ability to inspect hardware, audit power usage, cross-reference chip counts—vanishes into the Pacific.

Similarly, U.S. chip export controls become much harder to enforce once chips are stored on uncrewed infrastructure five hundred miles from land. We don’t have to speculate on what evasion at sea looks like because we’ve already seen it happen with oil: Russia’s shadow fleet—now over 3,000 vessels—moves billions of barrels of sanctioned crude annually using mature and well-documented methods. They conduct ship-to-ship transfers in open water, breaking the chain of custody so that oil loaded in Novorossiysk arrives in India with paperwork suggesting it came from somewhere else. Additionally, tankers spoof their AIS transponders or go dark entirely.

Analogous techniques work for chips, and in some ways, the problem is worse. Advanced GPUs are smaller, lighter, and more valuable per pound than crude oil. An Nvidia H100 weighs about three pounds, is worth roughly $30,000 on the open market (and considerably more where supply is restricted), and can fit in a backpack. Additionally, identifying the theft or illicit changing-hands of chips in the remote, open ocean is difficult. Nvidia’s geolocation system is basically an internet ping that depends on chips’ continuing to communicate with Nvidia’s servers. If a chip stops phoning home, it effectively disappears from view. So a node could deliberately go dark under the auspices of a storm and be written off as damaged or lost while proverbial GPU pirates covertly move the valuable hardware elsewhere. Additionally, a chip on a node in the Pacific may appear to be where it is supposed to be, but that does not tell you whether the tracking system is still working properly, whether someone tampered with it, or whether repairs or part swaps were accurately recorded. This is also true of terrestrial data centers, but corroborating any of this would require physical inspection, which is difficult in the middle of the Pacific.

Mare Computum

Grotius looked at the ocean and saw a void, a commons defined by its emptiness. Panthalassa looks at it and sees a power plant. The ocean is about to be full of things that matter, built by people moving faster than the institutions around them, in a space that the law has always assumed would stay empty. It won’t. And the gap between what the ocean is becoming and what the law imagines it to be is not a reason to stop building. It is a reason to start governing and competing, thoughtfully and quickly. The countries and companies that understand that will shape what comes next. The ones that don’t will find that the oldest commons on Earth has new landlords, and that nobody asked for their input on the lease.​​​​​​​​​​​​​​​​

In two weeks, the European Union (EU) will publish its first mandatory triennial review of the Digital Markets Act (DMA). The DMA, along with its twin, the Digital Services Act (DSA), ranks among the most comprehensive tech regulations in the world. The two cover topics as mundane as app store hosting fees, as technical as interoperability requirements with competitors, and as all-encompassing as election integrity. The DMA review takes place under the backdrop of a larger, omnibus simplification effort launched last year to review EU law, with an eye toward reducing compliance burdens and minimizing inconsistencies.

The DMA and DSA have captured a great deal of attention in the US, and not just because the Trump Administration views them as an unfair money sink for American companies. The concerns that animate the laws are gathering sympathy by the day from American politicians, judges, and juries. That’s why this week, I published a white paper analyzing the efficacy of the DMA, DSA, and related EU laws. Undoubtedly, the laws have had some success, including in pushing companies to open up their tech to greater interoperability with competitors. But those successes are dwarfed by the broader incoherence of a legal scheme that ignores basic constraints inherent to regulating the digital realm. If Europe genuinely wants to improve both the efficacy of these laws and the business environment, it needs to reconsider how these constraints effect the DMA and DSA’s viability.

Subscribe now

Compliance Must Be Achievable

Technically and legally infeasible obligations not only discourage major firms from operating, but also limit the growth potential for domestic competitors. When companies complain that they cannot comply with the DMA and DSA, they often say it’s because they are trying to protect users from some other violation of rights to privacy or freedom enshrined elsewhere in European law. Technical expertise helps lawmakers differentiate between when industry claims infeasibility just to delay compliance and when compliance requirements ask the impossible. Yet technical expertise must be coupled with legal feasibility for the regulatory regime to work. When regulations aim to balance competing interests like interoperability and data privacy, or freedom of speech and content moderation, regulators must clarify priorities. Otherwise, compliance with one component of the law will endanger compliance with another. The EU’s regulatory regime fails on this front, preferring to pretend that firm animosity is the only thing impeding the balance of contradicting obligations.

The DSA’s age assurance provisions, for instance, are among the least feasible, since robust compliance using current technology often infringes on both the EU Charter’s right to privacy and its right to expression and information. The DSA mandates online intermediary platforms must appropriately alter minors’ access, content, and advertisements without collecting more data to determine age, profiling based on pre-existing data, or collecting identification. It further requires that platforms not overly age-gate, excessively alter content, or implement methods with too many false positives so as not to impinge on the EU Charter’s right to expression and information.

The best way to comply with these obligations is using digital wallets that confirm identity through privacy-protecting zero knowledge proofs, wherein the application itself can confirm age without accessing identification locally stored on your device. Apple piloted this in 2021 in a handful of states; Google Wallet rolled out its zero-knowledge proof feature less than a year ago. Nonetheless, developers keen on using them in the EU are in a holding pattern while member states focus efforts on an EU-wide December deadline to establish their own digital identity wallets. In the meantime, the lack of clarity around what qualifies as a satisfactory method of age verification has not stopped a slew of European Commission investigations. Should the EU Digital Identity program roll out smoothly (unlikely given the lack of progress among many member states), this would mark a major milestone in kids online safety. Until then, the DSA will remain in force for two years and counting where the path to compliance is murky at best.

Regulations Must Actually Support the Interests They Aim to Favor

Regulations that lack sufficient market evidence to prove their efficacy are likely to undermine the groups they aim to serve. European regulators, and Europeans themselves, may prefer a society where fairness among economic competitors is prized above product cost or quality. Such sentiment is not only legitimate, but has returned to vogue around the globe, including in factions of both American political parties. However, regardless of whether a law aims to support the interest of consumers or competing firms, it needs evidence it will actually serve its intended purpose. Otherwise, laws will, at best, complicate the market and, at worst, hurt the very interests they have intervened to support.

Perhaps the most obvious case of such a provision is the DMA’s interoperable messaging requirement, which aims to make WhatsApp able to communicate with other messaging services. Putting aside the technical difficulties of architecting an inter-app messaging system for end-to-end encrypted systems, DMA Article 7 minimizes the differences in design and function that make competitors attractive in the first place.

Take Telegram, one of the few notable contenders against Meta’s WhatsApp and Messenger in the EU. Unique Telegram features include the ability to deliver messages quietly and to sync chats seamlessly across devices. Each of those is undermined by interoperability with Meta. If WhatsApp adopts quiet message delivery, the feature loses its uniqueness; if it does not, the feature is lost when interoperating. Telegram syncs seamlessly across devices because it uses cloud encryption for most chats; chats using end-to-end encryption are only available on their device of origin. To interoperate with WhatsApp would mean either dropping that key feature or, as Telegram claims WhatsApp does, neutralizing the value of end-to-end encryption. Signal offers another example. WhatsApp sends messages even if the security key changes while the message is in transit, while Signal will not send the message at all in that instance. Signal prioritizes security because it markets itself to especially security-conscious users; WhatsApp prioritizes convenience because it markets itself to everyday users. A common protocol would eliminate product differentiation that benefits different consumer niches.

Today, there is a small contingency of users who opt for other messaging apps simply to protest Meta’s power. For their protest to remain meaningful, the apps they use will have to refuse to participate in the very interoperability scheme that is meant to empower them. Rather than diminish the power of gatekeepers, Article 7 gives Meta legitimate reason to audit competitors’ security practices and gives Meta the metadata of conversations with users of other services. All this is done without offering users much additional reason to choose an alternative service. Indeed, nothing is stopping Meta from simply integrating the novel features of the few startups that have taken the firm up on its offer to sync chats with WhatsApp.

Mandating Content Moderation Inevitably Limits Free Speech and Press

Promoting freedom of speech necessarily hampers regulators’ ability to control illegal, objectionable, and misleading content. In attempting to both crack down on such content and promote free speech and media, the DSA runs afoul of European legal customs while forcing platforms to faithfully predict what EU regulators deem acceptable speech. Even though these provisions originate from a European distrust of platforms’ ability to police content well, the DSA places a much greater expectation on platforms to do precisely that.

In theory, the DSA maintains the EU’s “safe harbor” principle, which establishes that platforms have no obligation to comb through every post to ensure its legality. In reality, the DSA chips away at this limited liability by establishing a network of “trusted flaggers” whose flags for illegal content, including hate speech, must be prioritized and ideally reviewed within 24 hours. Given the volume of flagged posts and the documentation involved in removal, the turnaround window incentivizes platforms to rubber-stamp trusted flaggers’ suggestions, regardless of whether the post would qualify as illegal if scrutinized by a court. This is why France’s constitutional court struck down a similar hate speech law in 2020, ruling that the scheme encouraged over-removal and allowed flaggers (in this case, France’s law enforcement agencies) to circumvent the courts in deciding what is legally acceptable. It’s also why the UN Special Rapporteur on Opinion and Expression warned that governments holding platforms liable for what is posted on them chills free speech and cannot be the path toward more civil online spaces.

The EU’s content moderation requirements go well beyond trouncing the due process of hateful posters, and its goals are chock-full of contradictions. In promoting legitimate journalists by making platforms warn them before removing or demoting their content, they impede platforms’ ability to minimize false and misleading information. In giving media companies (among others), the ability to fact-check, they allow more established media outlets a privileged position in the information market, to the detriment of independent journalists and media pluralism. The harms are not hypothetical, case law and examples of independent journalists hurt by the type of removal the EU now mandates are plentiful.

Lessons Learned

In some ways, Europe is a distorted reflection of America. It believes a free society need not lead to one where vitriol and brain rot are its primary means of political engagement. It wants small market actors to have a fighting chance against billion, and now trillion, dollar companies. Somewhere deep down it believes that industry is not the enemy, that a refined version of it can and must fuel its workforce and advance its society. Its laws governing the digital world, unrestrained as they are, are trying to express the same sentiments that seem to be some of the little common ground left in American political life. For all this, I wish the Europeans luck in their review and reform efforts. But even if they would prefer to maintain their utopic yet unworkable morass of digital regulations, American lawmakers can’t afford to ignore the lessons the DMA and DSA have taught us.

Let Europe tilt at windmills, America needs to touch grass.

Last week, colleagues of mine at FAI spoke on The Frontier about some of the latest breakthroughs in quantum computing and their implications. The advancements of quantum technologies are rapidly transitioning from long-term research projects into deployable applications that impact our economic and strategic environment.

I believe this to be true, which is why last year I co-authored a four-part series with my FAI colleague, Dr. Prineha Narang, which you can read here. But there are a few recent events that illustrate the rapidity of improvements in quantum technologies, which should elicit the attention of researchers, investors, and policymakers to ensure U.S. quantum leadership.

The first examples of this phenomenon is how widely-used cryptography may be under threat far sooner than anticipated. Since February, there have been three announcements from quantum researchers worth examining.

Last week, Google Quantum AI published a whitepaper demonstrating that an optimized implementation of Shor’s algorithm could break the 256-bit Elliptic Curve Discrete Logarithm Problem. Practically, the operation described could enable a machine to penetrate the cryptographic backbone of Bitcoin, Ethereum, and virtually every major blockchain using fewer than 1,200 logical qubits and under 500,000 physical qubits, roughly a tenfold reduction from previous estimates. Under certain conditions, the attack could execute in approximately nine minutes, within Bitcoin’s average block confirmation window, the time it takes to add a new block to the chain, allowing an attacker to intercept a transaction in progress.

Complementing these findings, researchers from CalTech published a similar result, finding that advances in quantum algorithms and circuit design introduce significant efficiencies and computational improvements that could enable a break of widely used encryption with as few as 10,000 to 26,000 physical qubits. This would represent a step-change from prior estimates in the hundreds of thousands. Researchers working in tandem with the startup Oratomic demonstrated that ultra-efficient error correction using neutral-atom qubits could reduce overhead by more than a hundredfold. Simply, this means creating a fault-tolerant machine capable of running Shor’s algorithm could be operational by the end of the decade.

And earlier this year, Iceberg Quantum unveiled Pinnacle, a fault-tolerant architecture that reduces error-correction overhead by an order of magnitude, demonstrating that breaking RSA-2048 could be achieved with fewer than one hundred thousand physical qubits, not the millions previously assumed. Iceberg is already partnering with PsiQuantum, Diraq, and IonQ, each of which projects building systems at that scale within three to five years. Taken together, the Iceberg, Google, and Caltech results are challenging the assumption that the promises of quantum technologies were still decades away.

The second example are the recent reports that the CIA deployed a classified quantum sensing system known as “Ghost Murmur” to locate a downed American F-15E airman in the mountains of southern Iran. The system, reportedly developed by Lockheed Martin’s Skunk Works, uses advanced quantum sensors to detect the faint electromagnetic signal of a human heartbeat from tens of miles away, then pairs that data with AI software to isolate it from background noise. The pilot’s heartbeat essentially became a transponder in the desert.

These developments share a common lesson: quantum technology is advancing faster than previously assumed. In one case, the technology threatens the cryptographic foundations of the digital economy. In the other, it underlies a life-saving tool on the battlefield. The Genesis Mission, investments from the Department of Commerce, DARPA’s Quantum Benchmarking Initiative, and the National Quantum Initiative re-authorization signal that Washington recognizes the stakes of this technological advancement. But executive orders and benchmarking programs alone are insufficient without a sustained, granular policy architecture to back them up.

FAI’s Quantum Policy Playbook seeks to fill that gap. Submissions will translate strategic ambition into concrete, actionable policy proposals that accelerate American quantum research, harden domestic supply chains, expand the technical workforce, and ensure that the United States doesn’t only invent this technology but manufactures, deploys, and leads with it.

If interested, please submit a proposal no later than April 30, here.

In order to learn a new skill or develop expertise in a particular area, one must be exposed to certain types of data. In the context of developing artificial intelligence models, training data is the foundational layer of information which it relies on. The more high quality data that is available and used throughout training, the more capable and adaptable the model is across a suite of tasks.

Last week, FAI published The Data Crunch: Accelerating American AI through Government Data Access. The paper describes the importance of strengthening the data commons, a term that describes freely available and accessible data that can be used to train and improve AI models, specifically by making U.S. government data more accessible.

For the past few years, policymakers in Washington have focused most of their attention on policies that impact access to compute. Examples include but are not limited to the CHIPS and Science Act, the Biden Administration’s diffusion rule (now rescinded), Pax Silica, and the Chip Security Act. The focus on this particular input makes sense, because without this cutting-edge hardware, US firms would not be able to build and deploy AI.

But this does not mean that policymakers should neglect other key inputs, particularly data. The moment to enable more access to data is now for two key reasons.

First, the availability of unused, high quality data available on the open web is increasingly scarce. After a few years of training runs and model improvements, data that could provide a step-change in performance is unlikely to be surfaced by re-scraping the available web. There is plenty of data “out there”, but there are barriers to accessing such data, whether that be it is not digitized, it is unstructured, or there is another friction that raises the cost of acquisition.

Second, the deluge of copyright and intellectual property litigation by rightsholders against model developers continues to raise the risks of using copyrighted data under the presumption that such use qualifies as fair use. While a few decisions point to less-than-apocalyptic outcomes for model developers, more than seventy cases are still outstanding. Adding to the uncertainty, just last week Disney and OpenAI severed ties, dissolving the largest deal between an incumbent IP holder and frontier lab over the use of its products to support image and video generation capabilities. The divide between developers and the content industry seems to be widening by the day.

To address these twin problems, our paper proposes leveraging the significant troves of public data and existing federal authority to create the U.S. Data Accelerator. We identify four types of data that would be ideal starting points for the accelerator. They are:

1. Geological & Geospatial Data

2. Economic Development and Industrial Capacity Data

3. Regulatory Cost Data

4. Medical Device and Drug Trial Data

The types of data we identify have two things in common.

1. The data is already being collected by a federal agency and is made available in some form.

2. These types of data complement current capabilities of generative AI models and therefore could enable greater model utility and customization.

The Accelerator is key to unlocking this data because existing platforms for accessing such data are outdated and fit for a different time. Currently, Data.gov is home to more than 360,000 datasets, but due to the fragmentation of different agencies’ data, lack of regular updates, and heterogeneity in metadata and structure its usability for model development is limited. These deficiencies introduce frictions that limit the utility and usability of this data for model developers. Ideally, by establishing a clear, standardized schema and cadence for making such data available the Accelerator can ensure that this data can be used to the fullest extent.

Luckily, there is existing law and policy on the books that legislators and executive branch officials can rely on to advance this idea. The Foundations for Evidence-Based Policymaking Act of 2018 sought to improve federal data access and standardization. Simply put, this law requires agencies to make government data “open by default”, and includes specific expectations for government data assets. This creates the statutory foundation for federal data collection and dissemination to the public. The Office of Management and Budget could issue a memorandum under the powers this legislation vests in it to establish guidance for federal agencies to make publicly available data AI-ready. The Department of Commerce published a report in 2025 on steps taken to ensure publicly available data can support generative AI development and deployment. The insights found within this report could serve as a starting point for future guidance.

Even if there are gaps in this approach, there are willing partners within the legislature and the executive branch who stand ready to help.

On the congressional side, Senator Ted Budd (R-NC) and Senator Andy Kim (D-NJ) have introduced legislation to ensure that federal data can be accessible and useful for AI model development. Their legislation would direct the National Institute of Standards and Technology to collaborate with other relevant agencies and stakeholders to develop standardized schema to maximize the accessibility and utility of public data.

Turning to the executive branch, President Trump’s recently-released National Policy Framework for Artificial Intelligence recognizes the value of government data for model development. In Section V, the framework directs Congress to “provide resources to make federal datasets accessible to industry and academia in AI-ready formats for use in training AI models and systems.” Such language is unambiguous: federal datasets can and should be made accessible to accelerate American AI development.

The U.S. Data Accelerator is about enabling “American Innovation,” but it is also about “Innovating America.” Making more high-quality data available for model developers can support further scaling of existing model architectures and firms, while also creating opportunities for new firms or developers. Strengthening the data commons will help firms of all sizes, but particularly startups and new entrants. Such a project can fuel America’s entrepreneurial engine. But this project can also demonstrate the government’s ability to be an enabler of innovation by provisioning a public benefit. With the legal frameworks and raw resources clearly identified, the next step is building the capacity and will to execute.

While the Department of War’s (DoW) fallout with Anthropic dominates headlines, in the annals of the Pentagon, a far less sexy and far more symbiotic relationship with tech companies hums along. IT modernization may not sound consequential or impressive, but migrating to and managing systems in the cloud is no small feat. Enter the Joint Warfighting Cloud Capability (JWCC) program, a part-voluntary, part-mandated cloud acquisition facilitator for the armed services, combatant commands, and sundry DoW agencies.

In 2022, JWCC improved upon its politicized, years-behind-expectation predecessor by opening up its $9 billion in contracts to four vendors instead of just one. Drawing inspiration from a similar platform that the CIA stood up in 2020 for the intelligence community, JWCC allows each cloud service provider (CSP) to bid on each contract, rather than a single winner-take-all mega-contract. JWCC doesn’t just centralize cloud contracts, though, it also offers a more centralized strategy for managing, moving, and using data.

A key component of that strategy involves cloud interoperability and portability. That is, JWCC’s overseeing office aims to have data flow across providers without unreasonable latency (i.e. lag), to have applications translate well across cloud environments, and to structure data and applications so they can permanently migrate off their current cloud. Cloud interoperability and portability have proven to be a bit of a white whale: for the last few years, regulators and industry watchers have routinely pushed CSPs to seamlessly harmonize their cloud architectures. While some progress has been made, complete interoperability is next to impossible. Given that, JWCC stands out. And it’s not just its approach to interoperability that’s notable. If the saga with Anthropic is a masterclass in what not to do as an agency, then JWCC offers lessons in multi-cloud provision, in government procurement, and in government efficiency.

Subscribe now

Lesson #1: Multi-Cloud Interoperability Demands Personal Responsibility

American bureaucracy is not where one expects to learn a lesson in personal responsibility. But this is exactly what distinguishes JWCC’s approach to multi-cloud provision. The architects of JWCC understood that interoperability is the customer’s responsibility as much as it is the cloud providers’ and that no amount of regulation or standardization can save a cloud customer from poorly navigating a path-dependent market.

JWCC has been tasked with some tall orders. To start, it must coordinate a tangled morass of multi-cloud platforms among the armed services. While an increasing number of those contracts are funneled through JWCC and branches signal interest in collaborating, combating territorialism among military branches is always an uphill battle. JWCC also has the seemingly impossible task of establishing edge computing in austere environments, managing what may well be the most global IT system in the world, and establishing all-domain warfare capabilities that demand data-sharing across branches, allies, and clouds.

Previously, thousands of contracts scattered across the vast sprawl of the military debilitated purchasing power and thoughtful cloud management. That’s why JWCC’s governing body, known as the Defense Information Systems Agency (DISA), doesn’t just help secure the best contract for offices and services within DoW. DISA also serves as a sherpa for agencies in their cloud journey, advising customers on which products have best served others with similar needs in the past and developing custom IT tools that enable efficient cloud adoption.

Likewise, the DoW’s continuous Authorization to Operate, which began around the same time as JWCC and relies on its centralization, speeds software deployment by shifting from one-time approvals to continuous risk monitoring. This streamlines clearing the necessary extra hurdles for handling sensitive data—a win for efficiency and security.

Little of that success translates to private sector cloud operations. Even so, industry, and indeed regulators, have something to learn from JWCC’s clear-eyed approach to massive, indefinite multi-cloud provision.

First, the cloud is a living being, not an invisible building. Too often, enterprises treat the cloud as a static data container. DISA understands that each cloud provider, and public cloud itself, has a time and a place. For example, public cloud is not very valuable when you have a great deal of highly consistent traffic that needs to travel thousands of miles with low latency (think livestreams of military installations or continuous streams of classified reports). In such instances, it may be wise to switch to private cloud or to move data away from cloud’s compute-and-storage-rental model back to in-house hardware, a model known as on-premises. This is why, in addition to creating a private cloud, DISA has modernized on-premises outfits, has accelerated development and simplified cloud management regardless of CSP, and is building a cloud mesh to bypass idiosyncrasies and improve cross-system data sharing.

Accommodating DoW’s wide variety of agencies and services forces DISA to build more agilely, lending flexibility when a customer’s cloud needs change. Especially for large enterprises that, like the DoW, have varied and indefinite cloud needs, this strategy makes awarding cloud contracts more like renewing a lease and less like restructuring your mortgage.

Second, contracting out essential IT services is always a calculated risk. Vendor lock-in is the downside of best-in-class tools, some of which are proprietary to a single CSP. If interoperability is to be a worthy goal, it cannot mean relying only on uncustomized, open-source tools. The least common denominator is the least cutting-edge offering. If we think of interoperability and portability as essentially supply chain optionality—how many vendors can provision a tool and how well it will work with tools elsewhere in the stack—it becomes clear how idolizing these concepts stifles innovation. Just think of where critical infrastructure would be if we demanded every component have a dozen potential suppliers.

So then, desirable interoperability must mean weighing the benefit of a proprietary tool with the reality of being locked in to a vendor. As the deputy CTO in charge of JWCC put it, the expectation that you can pick “this piece of a cloud and this piece of another and then jam them together is…a fallacy.” The ability to be cloud agnostic, to be in one cloud on Monday, another on Tuesday, or simultaneously in both by Wednesday, relies on DoW’s ability to modernize apps such that they are as portable as possible.

Perfect cloud agnosticism just doesn’t exist. If complete off-the-shelf interoperability is possible (which it likely isn’t), we are years away from it, regardless of the regulatory state or industry standards. Relying on an outside contractor to provide a critical service requires doing the heavy lifting beforehand to ensure apps and data can migrate elsewhere should the time come. It also requires recognizing that life in the cloud isn’t frictionless, a lesson that rings all the truer in classified environments where some of the greatest obstacles to interoperability—identity and access management and security—are all the harder.

Many seem to presume the burden of interoperability falls solely on cloud providers, who avoid it in an anti-competitive attempt to entrench themselves in the IT infrastructure of their customers. While standards may facilitate interoperability, we cannot forget the role that personal responsibility plays in setting yourself up for success.

Lesson #2: In Government Procurement, Purchasing Power and Continuous Contracts Encourage Competitive Cooperation

A thoughtful approach to multi-cloud procurement isn’t the only thing working in JWCC’s favor. Its massive purchasing power also contributes to its relative agility. In the first 18 months of contract awarding, JWCC inked $1 billion in contracts, instantly ranking it on par with cloud vendors’ biggest customers. In the 12 months that followed, it awarded $2 billion more. Egress fees are a good example of this purchasing power leverage at work. JWCC negotiated baseline discounts ranging from 35 to 100 percent for data egress, upon which further discounts can be offered for particular task orders.

JWCC’s steady stream of lucrative contracts encourages CSPs to create interoperable solutions. This may seem counterintuitive, as CSPs participate in a prisoner’s dilemma-type game where interoperability makes winning future contracts easier regardless of existing infrastructure, but each CSP benefits from locking in customers to their cloud by using proprietary tech and mutually unintelligible semantics. However, this is an infinitely repetitive prisoner’s dilemma, in which there are near infinite rounds of contracts. In this scenario, game theory suggests that CSPs will cooperate with each other, making their systems as interoperable with one another as possible so that everyone plays nice in the sandbox. (Cue John Nash telling his buddies not to all go after the blonde in A Beautiful Mind). All the while, third-party providers, which JWCC looks to increase in its next version of the program, are incentivized to create seamless migration across CSPs for a particular app or program.

This cooperative strategy assumes some things about the way the repeated contract bid game is set up. First, it assumes that for any given contract, CSPs can either cooperate by making their service as interoperable as technically feasible, or they can defect by purposefully making the service less than optimally interoperable. For instance, if a customer seeks a contract for a customized tool that would function best using proprietary technology, the winning CSP is not “defecting” if technical hurdles prevent seamless interoperability with other CSPs. Rather, the provider only defects if it purposefully obstructs interoperability.

This leads to the second assumption: all losing providers must be able to detect whether the winning CSP has cooperated or defected, and all losers will calibrate their strategy the next round based on the winner’s strategy this round. Lastly, it assumes the payoff for the current contract round is not so astronomically high compared to the payoff of future task orders that it nullifies the incentive to continue cooperating. If CSPs expect this round’s contract to be worth more than the present value of all future contracts, the cooperative strategy no longer makes sense.

In truth, these assumptions don’t perfectly reflect hyperscalers’ cloud provision for JWCC customers, but they hold sufficiently to encourage greater cooperation. The first assumption obviously holds. In fact, the defect strategy of making services purposefully less interoperable is the strategy regulators presume hyperscalers adopt in the broader market. The second assumption is trickier. JWCC contracts are negotiated, not open. At most, the government only has to reveal how much the total award amount is and explain why the losers lost. Even if later interactions with the winner’s systems make obvious a previous defection, the time lag doesn’t allow for direct retaliation against the defecting winner; it only harbors regret that you yourself are not defecting more often.

Unlike in traditional prisoner’s dilemma games, however, this one has a game master. If a CSP defects, the probability of winning future contracts diminishes. Its unnecessarily poor interoperability may “lock in” future contracts that touch specific systems, but it has tarnished its broader reputation. JWCC negotiators act as third- party arbitrators or mediators, both of which lead to greater cooperation. As for the last assumption, a rational actor has no reason to believe that future contracts won’t be just as lucrative as the present one. Winning big now in a way that locks in large future contracts is still shortsighted in light of the potentially bigger contracts from which you’ve now practically disqualified yourself.

Realistically, JWCC is not a perfect game master. As its financial operations metrics improve, it will be better positioned to sniff out the runaway costs that vendor lock-in can cause, but for now its ability to detect defection probably leaves something to be desired. What’s more, JWCC contract officers may have short memories or turn a blind eye to defection if the service quality is otherwise exceptional. On the whole, however, JWCC’s pooled purchasing power and iterative multi-vendor contract awarding encourage hyperscalers to cooperate so that they can maintain a good reputation for future contracts and benefit from the cooperation of others. Government agencies have something to learn here—in procurement, sometimes, it pays to keep paying.

Lesson #3: There’s a Right Way to Think about Efficiency

To say JWCC is successful is to imply it is more “efficient” than the system it replaced. That word has gotten a lot of attention from policy wonks lately. Between DOGE’s bureaucratic warpath and the breakout success of the Abundance movement, 2025 was the year when both the right and the left began reimagining what our government could do, if only it were more efficient. At the crux of this imaginative project is how exactly to define efficiency. JWCC is a good example of what American bureaucratic efficiency should look like precisely because it doesn’t aim to be as efficient as possible.

First, in the pursuit of resilience, JWCC accepts redundancy. One could argue that the entire military operating off a single cloud provider would be the most efficient in that it is the only way to ensure complete interoperability, but this is one of many reasons its predecessor program failed. Procuring from a single vendor may streamline things, but it also fosters mediocrity and vulnerability, whereas real efficiency is durable and versatile.

Imagine a race car capable of hitting 275 mph, significantly faster than the top speed on most NASCAR tracks. If hitting 275 causes it to overheat or break down, it would come in dead last every time. Similarly, if you build such a vehicle as cheaply as possible, its versatility is limited. There’s a reason an F1 car is wildly more expensive than a NASCAR stock model—the latter can only handle oval tracks with banked turns, the former can turn on a dime. The car best suited to its aim is the most efficient, even if it takes longer to build or costs more.

Second, JWCC entices, rather than manipulates, private companies’ participation. JWCC invites companies to compete in a way that serves the public interest without attaching political prerequisites or changing the conditions to participate partway through the process. When unelected officials’ administrative capacity is so great that it can pursue its own political ends, we’re in trouble. Perhaps this is what prominent mid-20th century liberal Senator Eugene McCarthy meant when he said, “An efficient bureaucracy is the greatest threat to liberty.”

Not only is this type of efficiency anathema to our democratic republic, it also isn’t efficient in the long run. If bureaucratic directives at the behest of the sitting executive are a hyper-efficient way to enact a political agenda, they are also imminently reversible.

In the modern whirlwind of extra-congressional action, bureaucrats marshaled by the last three presidents have changed everything from environmental goals to immigration policy to student loan forgiveness programs, only to have them changed back again roughly four years later. Those of us who have closely watched the disbursement of money (or lack thereof) under the BEAD Act are intimately familiar with the inefficiency of executive tampering. In 2024, a landmark Supreme Court decision overturned a longstanding precedent known as Chevron deference. In effect, this deference encouraged agencies to interpret their mandate according to their own political preferences and the sitting executive’s agenda. Its overturning all but guarantees bureaucratic delays as Congress further clarifies and updates statutes. Yet the Court’s decision also exhorts Congress, rather than unelected federal agencies, to steward the programs it has enacted and the taxpayer money with which it has been entrusted. This improves long-term efficiency: careening the ship side to side prevents it from moving forward. Efficiency demands having a clear mandate and sticking to it.

JWCC, then, is a successful government program not because it worships efficiency at the expense of innovation or resilience, nor because it uses its lucrative contracts to entangle companies in ongoing political fights. It is successful because it is clear eyed enough to know it can’t outsource project stewardship to contractors, it’s big enough not to get bullied by major corporations, and it is insulated enough from political meddling both within the Pentagon and at the White House.

Waging war in Iran and at domestic tech companies leaves little time for introspection, but the DoW would be well served by reflecting on its own success as it prepares to announce a JWCC Next, JWCC’s successor. It would seem many of the lessons JWCC offers have yet to fully sink in, even within the Pentagon.

If online chatter is any indication, egress fees are the boogeyman of compute, the bane of competition. These fees have come under scrutiny from regulators the world over, not least of which because there’s often a charge to move data to a competitor or back into your own on-premises IT infrastructure. While providers vary in their data transfer pricing, egress occurs any time data is sent around the world, transferred to different parts of the cloud, pulled to an external location, or summoned by an external IP address. Egress fees, to the lament of regulators, are multiples more expensive among the largest three cloud service providers (CSP), the hyperscalers of AWS, Microsoft Azure, and Google Cloud Platform.

But the panic makes a mountain out of a molehill. Egress fees distract from more earnest and harder to resolve industry issues, like interoperability, hyperscalers’ cross-market advantages, and inflated license fees for SaaS offerings on other providers’ platforms. Here, we outline a reproach of common regulatory proposals against egress fees, particularly those offered by the EU Data Act and the various European market studies that informed it.

Not All Egress Fees Are Made Equal: The Case Against the Outright Ban

Though invisible, data is heavy, at least when pushing massive amounts of it around. Egress fees are a way to make the businesses that take up the most bandwidth internalize the cost to the overall system. Should egress fees be outright banned, that cost would be redistributed among all customers, including small businesses whose infrequent data transferring is covered under hyperscalers’ “free tier.” In fact, over 90% of AWS customers pay nothing in egress fees. Yes, the egress bill of a video streaming company is expensive and confusing, but that’s the tradeoff for quality scalability.

Egress fees are also a way for some CSPs to signal that their comparative advantage lies in compute and storage rather than data transfer. In fact, one might reasonably argue that they advance competition by allowing smaller competitors and adjacent industries to capitalize on more efficient, less energy-intensive ways to move data. If an enterprise does not like its egress costs, it can use companies like Alluxio or Snowflake to compress files or cache the most commonly used data locally, drastically reducing egress costs while expanding the cloud industry ecosystem. Or an enterprise can choose to bypass the internet when transferring data, thereby taking advantage of the public cloud without draining shared resources by constantly retrieving data stored on the internet. In either case, egress fees encourage the biggest cloud customers to better steward their data in a way that keeps cost lower for the majority.

The Practical Difficulties of Exempting Fees When Switching to a Competitor

But what if an enterprise wants to leave its current provider by returning to on-premises or by opting for a smaller CSP that displaces egress costs elsewhere in its pricing structure? Making the bandwidth hogs internalize costs is one thing, but holding companies’ data hostage when they try to exit your service is practically the definition of anti-competitive. While this line of thinking produces more targeted regulation, it’s misguided. A study prepared for the European Commission found that the egress costs for fictitious customers to transfer out of all their data over the public internet amounted to less than half a percent of their annual cloud services cost. Even if eliminating egress fees did lower total switching costs, there remains the difficulty of distinguishing when a customer is switching to a competitor rather than transferring data as a normal course of business.

To comply with the EU Data Act, the hyperscalers now allow customers to apply for fee waivers on one-time data transfers to other providers or on-premises systems, so long as the customer removes all data and workloads. Google also suggests European customers “may be eligible” for free transfers to proven multi-cloud partners.

The 60 (Google and Microsoft) to 90 day (AWS) complete removal required for the waiver is not only a tight turnaround that could compromise security during migration, but is also completely out of touch with the market. Eighty to ninety percent of public cloud users operate on two or more providers. Assuming Google sets a standard in discerning when data transfers are phased exits from Google’s service and when they are egress typical of multi-cloud architecture, this still distracts from a larger concern.

The more scrupulous CSPs are in determining which data transfers qualify, the more valueless these fee waivers are and the more complaints regulators will get that they’ve not gone far enough in preventing vendor lock-in. The more generous CSPs are in granting waivers, the more these exemptions resemble an outright elimination of egress fees, with all the market inefficiencies and harm to smaller cloud customers that brings.

Offering Data Transfer at Cost Hurts Innovation

The EU Data Act mandates egress charges associated with multi-cloud use not exceed the cost incurred by CSPs. It further stipulates one-time data transfers be offered at cost for the next three years as a transition before eliminating all provider-switching related egress fees. Insofar as this attempts to appear more measured than a full ban, it is all the more nonsensical. The average company saves over 200% by using Infrastructure as a Service from cloud providers. This savings is closer to 270% for start ups and smaller companies, who have long been noted as the primary beneficiaries of the cloud’s offerings. The cloud is not without downsides, but its services greatly benefit customers. Free market societies encourage innovation by allowing service providers to compete and to profit from that competition. We eliminate all incentive for providers to differentiate products and improve services when we prevent profit. There is only an arbitrary difference in making CSPs offer data transfer at cost and asking them to offer compute or storage at cost. In fact, if the aim is to improve consumer welfare, there is no point in stopping at egress fees when, for even the most data-intensive cloud customers, they constitute just 6% of total cloud spending.

What’s more, AWS has passed on over 100% of the drop in egress costs over the last several years, meaning it has lowered prices slightly more than the cost of transferring data has decreased. Sure, it up-charges egress costs more than its competitors, but it is, after all, a business. Businesses that overcharge lose market share, which AWS has, albeit slowly, over the last few years.

Sticky But Not Stuck

The inconvenient truth for regulators is that even in an idealized, highly competitive cloud market, cloud provision is sticky, meaning that the very nature of the service being offered makes customers unlikely to switch often. Cloud providers are more like landlords than they are like cereal brands. You’ll probably swap Rice Krispies for Cheerios if the price of the former shoots up, but you may choose not to move even if your rent rises. It’s just too darn annoying. As one economic study found, 90% of cloud customers did not adopt a cheaper, exact replica product offered by the same provider even though it would have saved customers 22% of their total cloud spending. If customers refuse to switch to perfect substitutes offered by the same provider, where interoperability and data portability are nonissues, this tells us the market is simply inelastic. Before regulating allegedly anti-competitive practices, we need to consider how the nature of the market shapes customer preferences. Granted, a demand-inelastic market does naturally allow for greater abuse on the part of cloud providers. Luckily, the cloud is a commercial product, not a public service.

Public Cloud, Not Public Service

Cloud service providers are profit-maximizing companies whose main customers are other profit-maximizing companies. Everyday consumers interact with the cloud either through their workplace or indirectly when, say, watching a Netflix show or ordering GrubHub, without realizing their systems run on AWS. You may have many subscription services, but renting compute and storage is likely not one of them. This makes the cloud different from other digital services coming under scrutiny, like Apple’s walled garden or Google’s advertising empire, in that neither individual consumers nor small businesses are adversely affected by the practices of digital service giants. Anti-competitive practices don’t have to target the little guy in order to be sub-optimal, but don’t let regulators tug at your heartstrings on this one: in this arena, there is no David, only Goliaths.

On December 11, 2019, Twitter CEO Jack Dorsey announced a bold new effort to develop an open and decentralized standard for social media. It would be incubated within Twitter, but the ultimate goal of the project would be to “allow [Twitter] to access and contribute to a much larger corpus of public conversation, focus our efforts on building open recommendation algorithms which promote healthy conversation, and [force Twitter] to be far more innovative than in the past.” He named it Bluesky.

After Elon Musk bought Twitter in 2022, Bluesky was spun off and eventually launched an app of the same name. Early growth was healthy but manageable, driven largely by invitations. Friends invited friends; nerdy techies made up a large percentage of the early commentariat. Many were Twitter expatriots fed up with Musk’s management of their beloved social media platform—an echo of earlier decampments on the right, when those unhappy with the Dorsey regime fled to Parler and Gab. The sharp politicization of X during the 2024 election, however, led to a tipping point: whereas the right-wing networks had peaked around 2 to 4 million active users, Bluesky exploded to over 42 million. Suddenly, social media had balkanized: X was for right-wingers, and Bluesky had become Lib Twitter.

Now, many are starting to ask questions about Bluesky. Is the platform really an idealized version of X for the liberal intelligentsia, where science and facts get their due? Is the balkanization of social media actually a good thing for democracy? With Trump Derangement Syndrome running rampant, is Bluesky trapped in a doom loop of its own making?

But all these questions miss the forest for the trees. Its real significance lies in its architecture and ethos of openness. By building a platform where control is distributed—across users, algorithms, and servers—Bluesky offers a vision of social media that isn’t beholden to any single corporation or ideology. It has the potential to deliver something rare in today’s tech arena: user empowerment. But to get there, the platform will need to expand its network, welcome a broader public, and encourage outside innovation. If Bluesky succeeds, it won’t just be another Twitter clone; it could redefine how we connect and communicate on the internet.

Subscribe now

The X-odus

Bluesky’s genesis can be traced directly to Dorsey’s acknowledgement of Twitter’s failures: he saw that Twitter had become too centralized and vulnerable to corporate and government control. To get Twitter back to its roots as an open, decentralized system, Dorsey founded Bluesky. The idea was bold: instead of letting a single company control the public conversation, build a federated, user-controlled network; recreate Twitter as a protocol—a set of open standards that anyone can use and build off of—rather than as a closed platform.

In 2021, Twitter’s leadership brought on Jay Graber as CEO to spearhead the effort. The small but scrappy Bluesky team worked on building what would eventually become the Authenticated Transfer (AT) Protocol: the underlying architecture designed to enable a federated social web.

At its core, the AT Protocol establishes a standard format for identity, follows, and social data, allowing different apps and servers to interoperate​. In practical terms, this means a user could move their account from one server to another without losing their followers or content​. It also means no single provider owns your social identity. “There’s no one company that can decide what gets published; instead there is a marketplace of companies deciding what to carry to their audiences,” Bluesky explained. In other words, the network is meant to be decentralized by design: many independently operated communities connected by a common protocol, much like email. Users, not corporations, ultimately hold the reins.

All of this was going on behind the scenes at Twitter, with the goal of eventually integrating Twitter into the AT Protocol as its first client. Then Elon happened.

Musk’s acquisition of Twitter in late 2022 marked a dramatic shift for the platform, as he sought to reduce excessive content moderation and restore what he saw as a commitment to free speech. His changes—including reinstating some banned accounts, overhauling Twitter’s verification system, altering content moderation and curation policies, and restructuring the company’s workforce—were intended to make the platform more open and financially sustainable. Twitter also jettisoned the Bluesky team.

As fate would have it, the chaos happening in San Francisco turned out to be a blessing. The Bluesky team kept working on achieving its mission of creating a decentralized social media protocol as an independent organization. In its early days, the platform was tiny and exclusive—essentially a techie salon for those with invite codes. This closed beta phase beginning in early 2023 helped refine the technology, but it also shaped public perception: from the outside, Bluesky looked like an elitist clubhouse of techy cognoscenti. As one commentator quipped, Bluesky was little more than an “invite-only clone of Twitter” that was more “like a country club … where you can go and be with your [techie] peers.”

Bluesky opened to the public in February 2024 and grew to a few million users over the next few months. Then, in the immediate aftermath of the 2024 election, daily usage saw hockey stick growth as the app went from approximately 3 million to 30 million users overnight. People fleeing Musk’s X—primarily people with a progressive-liberal bent revolted by Elon Musk’s close relationship with President Trump—flocked to Bluesky in what has become known as the “X-odus.”

Celebrities, politicians, and the chronically online rushed to claim their usernames on Bluesky, lending the new social network an aura of being the next big thing. Given that most of these new users’ only commonality was a hatred of Donald Trump and Elon Musk, it was natural that the Bluesky discourse trended heavily toward anti-MAGA politics, shaping public perceptions of it as a progressive sanctuary rather than the more ideologically diverse network it aimed to become. High-profile progressive users including Alexandria Ocasio-Cortez, George Takei, Kara Swisher, and Bryan Tyler Cowen were influential in defining the discourse. It is no coincidence that The Lincoln Project, a vehemently anti-Trump political group, is one of the most followed Bluesky accounts.

This was partly by design: Bluesky’s content moderation approach is far more proactive than Musk’s laissez-faire ethos on X. The Bluesky app launched with robust tools to help users curate their experience with features like community block lists, content labels, and customizable feed algorithms​. Users could even subscribe to shared block lists to automatically mute entire groups (one popular block list targets accounts labeled as “MAGA supporters” en masse)​. However, this also reinforced the Lib Twitter label. The federated model was not yet fully in effect—most users were all in one big community—so if that community leaned liberal, there was no easy on-boarding of right-leaning or nonpolitical groups elsewhere.

Amid this hype, Bluesky’s deeper mission has been obscured. Fundamentally, the project has always been about building an open, federated social media ecosystem where users control their data and communities set their own rules. Indicative of this mission is Bluesky’s unofficial motto, mundus sine caesaribus, which roughly translates to “a world without emperors.”

But high-profile incidents have distracted from this mission. Nowhere is this more evident than in the Jesse Singal Incident.

When Jesse Singal, a self-described liberal journalist known for his opposition to youth gender transition, joined Bluesky, the backlash was immediate. Singal quickly became the most-blocked account on the platform, and a petition circulated urging Bluesky’s leadership to ban him in spite of the fact that he had not broken any platform policies. Deplatforming Singal rapidly turned into a progressive cause célèbre. Ultimately, Bluesky leadership declined to take action, adhering to its principle of user-controlled moderation rather than top-down enforcement. But the incident underscored an important tension. While Bluesky was designed as an open and decentralized network, its early user base was ideologically skewed, and incidents like this demonstrated that it was an inhospitable space for certain viewpoints. In sum, a perfect storm of user migration, moderation philosophy, and early adopter demographics earned Bluesky its reputation as Lib Twitter.

Looking Under the Hood

Despite its politically homogeneous start as Lib Twitter, Bluesky’s architecture holds the promise of far greater pluralism. The very goal of federation is to enable many communities with different norms and viewpoints to coexist, each operating its own server under the shared network. In theory, Bluesky could host a constellation of diverse forums—conservatives, liberals, centrists, Twilight stans, Catholics, furries, Joe Rogan fans, sadomasochists, etc.—all interoperating through the AT Protocol. Instead of one company’s rules setting the tone (as with Twitter or Facebook), each community can set its own moderation standards. Most importantly, each community is subject to the others’ rules when entering each other’s spaces; no spamming your political enemies.

The key innovation Bluesky is pursuing is algorithmic choice. Unlike traditional social platforms that impose one opaque algorithm on everyone, Bluesky is building a system where users can select (or even build) their own feed algorithms. “We want a future where you control what you see on social media,” CEO Jay Graber wrote. The goal is to enable a “marketplace of algorithms.”​ In practical terms, a Bluesky user might swipe between a purely chronological timeline, a community-curated feed, or an algorithm that downranks “hateful” content. Users’ choice over how their feeds work can be even more granular, switching from their puppy-pictures feed to their hyper-political news feed to their sports-news-only feed and back again in seconds, depending on their preference​.

By giving people this power, Bluesky aims to break the cycle of engagement-driven outrage. If you can opt out of the algorithm that amplifies the loudest and most inflammatory posts, the whole tenor of the online conversation can change. This user-centric approach to curation could foster healthier discourse and more cross-pollination of ideas than the enraging shouting match that characterized the Twitter experience.

Crucially, Bluesky’s decentralized architecture may offer a better path forward than earlier federated projects like Mastodon. Mastodon and the broader Fediverse—shorthand for the “federated universe” of independently hosted, interconnected social platforms and services that communicate via shared protocols—proved that decentralization can work at scale, but they also exposed pitfalls. The Fediverse experience is uneven and confusing for newcomers. For example, Mastodon’s email-like federation often leads to fragmented conversations: two users on different servers might see completely different reply threads on the same post, because each server only shows the replies it knows about​. Bluesky is trying to avoid those pitfalls by designing a more unified network structure on top of federation.

The AT Protocol includes a “Big Graph” service—essentially a network-wide index—to ensure a global view of content and consistent identity across servers​. That means a user on a small community server can still seamlessly follow, reply to, or search posts from the main Bluesky server (or any other server) with the system handling the synchronization behind the scenes. If successful, this model would combine the diversity of decentralization with the convenience of a centralized service. In short, Bluesky is decentralized without feeling disjointed. This could make it a far more compelling alternative to centralized platforms than earlier attempts.

More than merely cloning Twitter’s microblogging, Bluesky’s technology opens up new possibilities for social media innovation. The AT Protocol is open source and intended as a foundation that developers can build on​. From alternative Bluesky clients, to browser extensions that verify identities via domain names​, third-party apps and services have already started to appear built on top of the AT Protocol.

In the future, entirely new social experiences could be constructed atop the network. Imagine an Instagram or YouTube in which content flows through the same decentralized backend, meaning your followers, content, and identity travel with you. Startups could build novel social networking apps that could immediately be part of a larger interoperable social web. Some enthusiasts even talk about integrating Bluesky with other decentralized tech: using cryptocurrency wallets for identity verification or payments or linking Bluesky content with open publishing tools like WordPress. This extensibility hints at Bluesky’s broader significance: it is not a platform, but an ecosystem.

The full potential of an open social protocol will likely be discovered by outsiders tinkering with it in the same spirit that the early open web spawned unexpected innovations. Bluesky’s decentralization isn’t an end in itself; it’s a means to unleash creativity, competition, and diversity in social media.

Getting From X to B

For Bluesky to evolve from Lib Twitter into the open social revolution its founders and early adherents envision, several steps will be crucial. First, it must expand federation in practice and onboard a far more ideologically and globally diverse user base. This means rolling out support for independent servers and making it easy for all kinds of communities to set up their own spaces. Bluesky will need to actively encourage a wider range of groups to join, including moderates, conservatives, international users, and others who didn’t rush in during the X-odus. If federation remains limited, or if most users stay corralled on the one big bsky.social server, Bluesky risks remaining an insular bubble.

The team’s greatest challenge is to balance growth with its values. In practical terms, that could involve developing tools for inter-server moderation so that servers with lax policies don’t overwhelm others with spam. The goal should be to let a thousand communities bloom while maintaining some connective tissue and shared standards across the network for dealing with illegal content.

Second, Bluesky and others need to nurture a developer and enterprise ecosystem around the AT Protocol. Its success will hinge on others embracing it as an open standard. The project has made its code open source and has already seen independent developers craft custom feeds and apps​. This is a promising start, but the momentum must continue. When entrepreneurs have their pick of dozens of social media protocols, the value proposition of building on the AT Protocol must be crystal clear.

Bluesky should facilitate third-party innovation by providing stable APIs, documentation, and perhaps funding or hackathons for building on the protocol. From novel algorithmic feed providers to specialized community servers or moderation tools, independent tools will only enrich the overall network. If companies see opportunity in the AT Protocol, they might build social features into their own products that hook into Bluesky’s network, much like how different email clients and providers all interoperate on the email protocol. Imagine news organizations, universities, or interest groups running their own Bluesky servers for their members, or software companies adopting the protocol for internal social feeds.

Such buy-in won’t happen overnight, but rolling out a welcome mat for outside innovation is essential. This also means Bluesky should prepare for standards governance. Cultivating this kind of open ecosystem will help Bluesky fulfill its role as a public commons rather than just another app.

Finally, Bluesky must figure out a sustainable business model that preserves the platform’s independence and user-centric ethos. As a public benefit corporation, Bluesky has explicitly committed to prioritizing an open and decentralized social web over profit​. That noble stance doesn’t pay the server bills or developer salaries by itself. Thus far, Bluesky has been supported by investor funding and the goodwill of its backers. Long-term, it plans to avoid the advertising model used by traditional social media since locking in user data would contradict the whole premise of user ownership​. The team’s initial revenue experiments have instead focused on paid services that enhance the user experience. For example, Bluesky introduced a paid domain name service that allows users to purchase custom domain handles (e.g., yourname.com) more easily​.

Thousands of users have shown interest in using personal domains as their identity, which both generates revenue and advances the idea of user-controlled identity. This is a start, but additional revenue streams will be needed. Possibilities include premium features, subscription plans for power users, or offering hosting and support for communities that run their own servers (Bluesky-as-a-Service anyone?). Ironically, a good example on how to do that is Elon Musk’s re-commercialization of Twitter in this area.

A Bluesky that fulfills its promise could empower communities of all stripes to take charge of their online experience. It could spawn new social apps we can’t yet imagine, all interoperable and user-first by design. And it could demonstrate that decentralization online isn’t just a pipe dream or a playground for geeks, but a viable alternative to the status quo. In an era of disillusionment with social media, Bluesky’s open-at-the-core approach is a refreshing experiment in what the next chapter of the internet might look like​.

It is change, continuing change, inevitable change, that is the dominant factor in society today. No sensible decision can be made any longer without taking into account not only the world as it is, but the world as it will be—and naturally this means that there must be an accurate perception of the world as it will be. This, in turn, means that our statesmen, our businessmen, our everyman must take on a science fictional way of thinking, whether he likes it or not, or even whether he knows it or not. Only so can the deadly problems of today be solved.

—Isaac Asimov, My Own View, 1978

Our Goal

When Isaac Asimov published his first piece of science fiction in 1939, he wrote for an analog world; a world without transistors or tupperware. There were no atomic bombs or automatic transmissions. We had yet to discover DNA or duct tape. Throughout his life, Asimov bore witness to many of the fictions he and others envisioned becoming reality. But he also knew that we were only getting started.

We live in an era of progress and incomprehensible change. The course of the last couple of centuries demonstrated that innovation can radically reconfigure geopolitics, economics, and daily life in ways that are not always under our control. Needless to say, as progress continues to escalate, technology presents us with enormous promise but also potential perils. The challenge before us is to continue ever upward without losing our humanity.

Toward this end, Asimov’s call for a “science fictional way of thinking” is not a flight of fancy. It is, in part, a recognition that we cannot approach questions of technology and society solely with the present in mind. A mindset that fails to anticipate technological trajectories risks becoming obsolete before any idea can be implemented. Consider the dilemma posed by the internet: early pioneers of cyberspace saw it as a pure conduit of freedom, liberty, and equality. In the words of John Perry Barlow, the internet was a place “more humane and fair than the world your governments have made before.” The reality has proven far more complex as the internet has proven a source of liberation for some and a source of oppression for others.

Asimov was right. Change is the dominant factor in society and yesterday’s perception of today almost always proves to be inadequate. At the Foundation for American Innovation, we operationalize that insight by treating science-fictional thinking as a discipline rather than a vibe. To us, science fictional thinking is not daydreaming of alien worlds or lightsabers. It is a process of backcasting from multiple plausible futures to shape today’s policy options. We consider different scenarios to stress-test our assumptions and continuously scan emerging technologies for second-order effects. In short, we take Asimov’s charge to “see the world as it will be” and translate it into concrete strategy so our decisions are resilient to tomorrow’s shocks and expand the frontier for American innovation.

But looking to the future is not sufficient. As Asimov highlights in My Own View and other essays, science fictional thinking requires not only foresight but also hindsight. To him, science fiction is the product of observing the advancement of science and technology and in order to imagine a plethora of futures, as Asimov did, one must have a firm grounding in the past. Technology policy is no different. As a colleague, Robert Bellafiore, has written before, “it is precisely by looking backward that we prepare to move forward, by ‘remembering what had come before’ that we become capable of imagining what may come next.”

Just as the ancient Roman god Janus’ two faces looked both forwards and backwards in order to oversee beginnings, endings, and transitions, we aim to look to both the past and the future. It is only by doing so that we can hope to glean the foresight, imagination, and prudence to meet the challenges of inevitable change.

Our goal, then, is this: to do everything we can to make the future happen sooner.

Subscribe now

Our Work

What we do is to try and approach issues of public policy in a science fictional way of thinking. Much of the DC intelligentsia surrounding technology and telecommunications approaches issues beginning with economics, sociology, or common law. These are all useful lenses from which to approach issues of public policy and many of our colleagues and competitors do incredible work from such a vantage point. Where we differ is that we choose to approach issues of public policy through the lens of technology.

It is for this reason that we spend an inordinate amount of our time learning about technology. We are fortunate enough to have built an extensive network of techies and we are not shy about leaning on their expertise. Whether it is educating ourselves about the organization and development of ancient Roman road systems or conducting cutting edge research on the future of quantum computing, our team starts with the tech.

In analyzing technology, we take the approach that tech and telecomm—which are often thought of in policy circles as distinct from one another—are, in fact, inextricably linked. We consider the full tech stack as one. Platforms, cloud hosts, DNS servers, wireless transmitters, cable lines, and submarine cables all work together to bring us into the digital world and policies that are targeted at one piece of this stack have ripple effects throughout the rest. Therefore, it is a crucial part of our work to think beyond just the piece of tech being discussed but to also consider how that piece interacts with all the rest.

While we start with tech, we don’t stop there. Economics, philosophy, common law, and many other fields are crucial to our work. But in starting with tech, we have the advantage of being able to apply social science and law to tech rather than the other way around. But it would be foolish to not use all the tools at our disposal. If you’re curious about where we’re coming from, just take a glance at our Bookshop.

Our Future

The pace of change is not slowing, and the stakes of getting it wrong are rising. In the years ahead, the frontier will be defined as much by software as by the oftentimes invisible infrastructure of compute, connectivity, chips, and standards that determine what is possible, what is profitable, and what is permissible. Those who treat technology as an exogenous force—something that “happens” and then must be managed after the fact—will find themselves forever reacting to the last disruption. Our bet is that the next era of American strength will belong to the institutions that can see the world as it will be and act early.

That is why this Substack exists. It will be our workshop in public. Some posts will be speculative, structured, evidence-based imagining anchored in what engineers are building now. Others will be more historical, because durable policy is rarely born from amnesia.

The throughline for our work is that innovation is not fate. Technological progress can widen the space of human freedom and national flourishing, but only if it is paired with the imagination to anticipate second-order effects and the prudence to build guardrails without smothering the frontier. We are uninterested in navel-gazing nostalgia and equally uninterested in naïve techno-utopianism. We care about the hard middle. We care about institutional design that lets free societies capture the upside of change while keeping the center of gravity human.

The future will be built by people who can see several moves ahead, who can learn from what came before, and who refuse to outsource tomorrow to accident.

What’s in a Name?

The name of this publication, The Ansible, describes the work we are trying to do: communicate quickly across distance—between disciplines, institutions, and futures—without ignoring the fact that distance exists. The word comes to us from the science fiction of Ursula K. Le Guin. In her novels she first introduced the ansible as a device that allows information to be transmitted instantaneously across space. But the ansible is more than just a gadget, it is a philosophy of connection: the right information sent to the right place at the right time to shape what happens next.

Le Guin never treated the ansible as a trick to make plot problems disappear. In her Hainish Cycle novels, she made it the backbone of an interstellar social architecture. Her cosmos is full of worlds too far apart for physical connection, even at lightspeed. Yet these varying worlds are capable of coordination because information can arrive when it is needed. The ansible enables the Ekumen, a federation that persuades instead of conquering. Communication has become a tool for reciprocity. That intuition—connection without domination—animates our work on technology and public policy. We aim to connect the people who build systems with the people who govern them, bring evidence into rooms where decisions are made, and enable pluralism.

The name also clarifies what we mean by taking on Asimov’s call to a “science-fictional way of thinking.” Science fiction at its best is neither prophecy nor mere entertainment; it is a laboratory for stress-testing assumptions. It allows us to ask how today’s architecture behaves under tomorrow’s loads. That is the method we try to bring to questions of networks, platforms, devices, and public policy.

Extending the Metaphor

The ansible metaphor disciplines our practice in several ways. It reminds us that information can move faster than institutions. Statutes do not change overnight, but ideas and incentives can propagate quickly and change behavior long before the ink dries on a bill. Our job is to compress the distance between insight and implementation. Speed does not excuse sloppiness; it demands clarity, because analysis that arrives late or muddled is indistinguishable from silence.

It also warns us against centralization by convenience. Le Guin’s ansible allows governments and technologists to coordinate without homogenizing. That is a design principle we try to honor in technology policy. When we champion interoperability, open standards, and contestable markets, we are not romanticizing chaos; we are acknowledging that a complex, innovative society stays resilient when no single chokepoint can dictate outcomes. Experimentation at the edge, coupled with accountability for power at the core, is how we ought to govern technology without smothering it.

A third lesson concerns tradeoffs. Even an instantaneous link has limits: bandwidth, context, interpretation. Likewise, modern networks are miracles bounded by physics and incentives. Undersea cables can be cut; satellites can be jammed; software can be hacked; supply chains can be poisoned. If policy pretends that constraints do not exist, it will push systems into failure modes. Starting with engineering humility is not technocracy; it is prudence. From that base we can apply economics and align incentives; we can apply law to protect rights and enforce responsibility; we can apply political theory to keep public institutions sovereign.

The Ansible in Practice

What does this metaphor mean for this publication? First, it is a promise about translation. The ansible in Le Guin’s work connects worlds that would otherwise ignore or talk past one another. Our aim is to do the same across the tech stack and the policy stack. We convene builders and policymakers, operators and scholars, civil society and regulators. We try to get deeply technical without losing legibility, and to get philosophically serious without losing purchase on implementation.

Second, it is a promise about scope. The ansible is a communications device, but in Le Guin’s hands its implications span from physics to diplomacy. So, too, our beat ranges from radio access networks to First Amendment doctrine, from datacenter interconnects to export controls, from privacy engineering to administrative law. Posts will vary in form. Some will be deep dives and some will be short opinions. All will seek to synthesize the complexities of technology and public policy into useful and actionable information.

Third, it is a promise about time. The ansible makes it possible for distant actors to coordinate while coordination still matters. In policy, timing decides whether analysis changes outcomes or becomes a tidy footnote. We aim to publish while Congress, agencies, courts, companies, and cities can still use what we write. When the facts change, we will revisit our priors.

The name guards us, too, against a common failure mode in Washington: mistaking rhetoric for reach. Faster-than-light communication does not abolish politics. It gives capable institutions the chance to coordinate and act with less friction and more foresight. That is enough. Our work is to help those institutions—public and private—see around corners, understand constraints, and choose well.

What to Expect

What you can expect here, then, is writing with the builders in mind and the public in view. We will begin with the technology. We will integrate economics, law, and philosophy as tools rather than talismans. We will try to be timely without being faddish, principled without being doctrinaire, and ambitious without being naïve. We will take seriously the way choices at one layer of the stack ripple across the others. We will pay attention to incentive gradients, institutional competence, and constitutional limits. Most of all, we will never pretend that complexity is an excuse for paralysis.

The last reason we chose this moniker is that the ansible refuses the fantasy of omnipotence that sometimes clings to technology. Instant communication in Le Guin’s books never makes disagreement disappear. It never erases culture or preempts moral risk. It simply gives communities a better chance to meet their responsibilities with eyes wide open. That vision keeps us oriented toward the work that matters in tech policy right now: expanding capacity without stifling competition, securing critical infrastructure without freezing innovation, moderating at scale without deputizing private censors, globalizing supply chains while hedging against authoritarian leverage.

Calling this Substack The Ansible is our way of committing to that balance. We want to move the right ideas quickly to the people who can use them. We want to knit together the technical and the civic, the near-term constraint and the long-term horizon. If we do our job, this publication will not just talk about the future; it will help bring the best versions of it into being sooner.

Welcome to The Ansible. Be sure to subscribe to get all of our latest research, opinions, and musings. Help us make the future happen sooner by sharing this with at least one other person today.