Whose Roadmap

 

The $300 billion sovereign premium is betting on a cap that hasn't happened yet — and all the evidence that cap relies on is still sitting in the "planned" column, not yet moved into "happened."

This statement holds true in both directions. Those who believe U.S. platforms can monopolize global sovereign compute rents, and those who believe the domestic stack is about to cap that monopoly — as of spring 2026, both are holding roadmaps, not battle reports.

We thought we were pricing a winner-take-all outcome. But when you comb through every Middle Eastern contract clause, cross-check several sets of widely cited industry figures, and force-separate the left pocket from the right pocket of the U.S. federal budget, you find the more pressing question isn't "who won" but this: Do the people placing bets even know whether their chips are real money or IOUs?

I. The Side Being Bet On: America's Sovereign AI Premium

In simple terms, the sovereign AI premium is the extra valuation markup the market assigns to U.S. AI platforms specifically because they serve national security and strategic infrastructure.

The past three episodes of Bear's Lens — Episode 2, The $300 Billion Valuation; Episode 3, Whose Premium; Episode 5, Whose Sovereignty — established one line of argument: the U.S. federal contract pathway is contracting significantly. Through the first eight months of fiscal year 2026 (October 2025 through May 2026), DoD AI contracts fell 94% year over year (from $183 million to $9.4 million), federal AI contracts fell 82%, and semiconductor contracts fell 83%.

But sovereign AI's valuation gravity hasn't collapsed — because the real capital flows have migrated into the hyperscalers' private closed loop: Amazon's cumulative roughly $12.8 billion equity investment in Anthropic and a reported $100 billion AWS compute commitment, Oracle and OpenAI's reported roughly $300 billion five-year commitment, and the NSA's shadow consumption through government-exclusive cloud channels that bypass the Pentagon blacklist.

The government no longer feeds platforms through contracts. It crowns them through endorsement and hands the valuation gravity off to global capital.

This line of argument is correct. But it carries an unexamined hidden premise.

The Commingling Problem in USASpending

The U.S. federal spending tracker USASpending lumps contracts (direct government procurement) and grants (government funding for academic or public institutions) into the same aggregate figure. When contracts collapse and grants surge, looking only at the total produces a slowly declining curve. But force-separate the two columns —

Contract column: DoD AI contracts fell from $183 million to $9.4 million. Federal AI contracts fell from $195 million to $34.7 million. Direction clear — downward, no dispute.

Grant column: Federal AI grants rose 141% year over year. HHS AI grants surged from $155 million to $1.1 billion (up 588%). Federal cybersecurity grants rose 273%. Direction equally clear — upward.

One collapses, one surges. Mix them together, and the narrative can be "the government pathway is shrinking overall" or "the government pathway is switching tracks, not shrinking." The key contrast cited in Episodes 3 and 5 — "public pathway contracting, private closed loop expanding" — how much of it depended on this commingled accounting? Bear's Lens doesn't know.

Until contracts and grants are force-separated and grant flows are traced to their end beneficiaries, this baseline remains uncalibrated.

Two Possible Destinations for Grant Flows

If that $1.1 billion in HHS AI grants ultimately flows to academic institutions, state governments, and nonprofits — historical experience says this is the default path. Before the Bayh-Dole Act, of roughly 28,000 patents held by the federal government, fewer than 5% were successfully licensed to the private sector and commercialized. Even after the act gave universities control over patent rights, cases of federal grants spawning platform-scale commercial alternatives remain rare unless mandatory open-source provisions were attached — the BSD operating system, the PostgreSQL database, and the Apache Spark big data engine are all survivors on that narrow path.

If grants follow the default path, their impact on platform valuations can be set aside for now.

But if that money ultimately trickles down through subcontracts or re-grants to top-tier cloud providers and frontier model API procurement, then the contrast argued in earlier episodes — "public pathway contracting, private closed loop expanding" — needs to be rewritten as "public money hasn't exited; it just switched to a channel that doesn't show up in the contract column."

Until these two possibilities are distinguished, the side being bet on — the baseline of America's sovereign premium — is a commingled figure, not a net number.

II. The Other Side Making the Bet: Domestic Stack Supply Figures

A set of industry figures repeatedly cited in spring 2026 — a Chinese AI chip company's shipments tripling year over year, a domestic accelerator card winning over 40% of state-owned cloud bids, a domestic foundry's advanced-node yield exceeding 90%, a liquid cooling equipment maker's orders up 280% — pieced together, they point in an exhilarating direction: the domestic compute stack isn't just "keeping the lights on." It's advancing simultaneously on integration, indigenous substitution, and high-density deployment.

Bear's Lens doesn't doubt the direction. But Bear's Lens cares about a more basic question: Are these figures roadmaps or battle reports?

The distinction isn't whether they'll come true. The distinction is whether they've come true right now.

Take the most frequently cited figure. Morgan Stanley's April 2026 research note gave this framing: "Roughly 100,000 units actually shipped in 2025; 300,000 units projected for 2026." That is a full-year 2026 shipment guidance, not a realized shipment figure. As of May, no quarterly earnings or official disclosure has confirmed shipments are on a triple-growth trajectory. "Triple" is the endpoint on the roadmap, not a station already passed.

The other figures share a similar pattern: some come from one-off replies on investor interaction platforms rather than systematic statistics; some first appeared in trial-production reports years ago and have never been updated in earnings filings; some show order-of-magnitude gaps with the overall growth rates disclosed in annual reports, likely reflecting an early order pulse from a single sub-category. The only figure broadly supported by industry technical literature — domestic rack power jumping from 8kW to over 30kW — reflects the entire industry's global shift from air cooling to liquid cooling, not a breakthrough exclusive to a single market.

The common trait of these figures isn't that they're wrong — it's that they're all still somewhere between guidance and realization. The distance between a roadmap and a battle report must be crossed through production, packaging and testing, delivery, and customer acceptance — four gates, one by one. In spring 2026, most of those gates haven't opened yet.

III. Where Victory Should First Leave Its Mark: The Middle East

We thought we were pricing a race, yet nobody looked back to check — on one end, the baseline is commingled; on the other, the figures are unrealized; and the place where victory or defeat should first leave its mark says nothing in its contract clauses.

If a lower-cost, faster-to-deploy alternative is truly reshaping the global sovereign compute landscape, the first place to see change is not domestic shipments on the supply side (that's a fact about production capacity, not a choice about procurement preference), nor new data centers in Southeast Asia (most projects are still at the MOU stage). It's the Middle East — the UAE's G42 and MGX, Saudi Arabia's HUMAIN. They are today's most aggressive marginal buyers in global sovereign AI investment, with both the political space and financial latitude to hedge bets across different suppliers. If the capping has already begun, Middle Eastern procurement terms should be the first place cracks appear.

Bear's Lens searched publicly available major contracts, memoranda of understanding, framework agreements, and press releases from these entities covering 2025–2026:

UAE side: G42's compute leasing talks with Northern Data mentioned only 23,000 NVIDIA GPUs. G42 and Cisco's large AI cluster in Abu Dhabi specifically named AMD Instinct MI350X, emphasizing "trusted U.S. technology partner." OpenAI Stargate UAE's Abu Dhabi 1GW cluster had no further terms updated as of May 2026.

Saudi side: HUMAIN and Saudi National Infrastructure Fund Infra's $1.2 billion financing framework specified only "frontier GPUs for AI training and inference." HUMAIN and Saudi Telecom STC's 1GW data center joint venture disclosed only power capacity and equity splits. HUMAIN and engineering firm MIS's general contracting agreement covered only design and construction.

In every publicly verifiable contract text, Bear's Lens found no clauses about dual sourcing, second source, fallback plans, or non-U.S. accelerators. Not a single alternative supplier's name appears in these documents.

This doesn't mean substitution won't happen. But it means that as of May 2026, the world's most motivated, most financially capable sovereign buyers pursuing supply chain diversification haven't left a single word for alternatives in their public procurement texts. The evidence the capping thesis needs most — not shipment numbers, not a capacity curve, but a written trace of procurement intent — is blank.

A Signal Flare from Southeast Asia

Southeast Asia offers one exception worth tracking: Malaysia's Skyvast partnering with Huawei to deploy 3,000 Huawei Ascend GPUs powering a localized DeepSeek model, paired with Kunpeng processors and a cloud-native system — a complete indigenous full-stack solution. But the Malaysian government promptly clarified this was a market-driven commercial deal, not a government-to-government agreement. Among new data center projects in Johor, Indonesia, and Vietnam from 2025–2026, there is no verifiable public record linking domestic compute or liquid cooling suppliers to specific project award records.

Intent precedes supply; supply precedes procurement — Skyvast is a signal flare, but not yet a shipping lane.

IV. Historical Mechanisms and Counter-Paths

This is not a video about "who's catching up to whom."

If you look only at historical mechanisms, the direction is clear. In telecom satellites, 5G networks, subsea cables, ultra-high-voltage grids, and urban surveillance — five sectors — history has staged the same drama over and over: when a "good enough performance plus state financing" low-cost full-stack solution appears, the high premium Western vendors maintained through engineering reliability and security endorsements gets rapidly capped within one to three years.

But the counter-paths are equally real. An African nation, partnering with a non-Western low-cost vendor to build a telecom network, faced the vendor refusing maintenance and demanding an extra $150 million — forcing termination of the framework agreement. It chose to accept a Western supplier's contract at 7.5% interest rather than remain locked in by a single-source shakedown. A Latin American country used the Budapest Convention on Cybercrime to erect legal entry barriers, effectively excluding non-Western vendors from its 5G bidding. Another South American nation's urban surveillance system adopted a low-cost Eastern full-stack solution, but after that vendor was placed on the U.S. sanctions list, technical support and the spare parts supply chain both fractured — exposing the deeper fragility of the low-cost model: rock-bottom initial pricing, closed ecosystem lock-in, escalating maintenance costs, and ultimately potential collapse from geopolitical shifts.

The mechanism is real. The counter-path is also real. Once substitution happens, it can move fast; but political lock-in and vendor lock-in risks can claw back eroded market share. This is not a one-way street. It's an elastic rope being pulled in both directions at once — and in spring 2026, both ends are still coiling, neither has snapped.

V. The Third Layer: A Blade Cutting Toward Both Ends

What truly needs to be said is the third layer.

Bear's Lens isn't pricing the invincibility of America's premium, nor the imminence of a substitution cap. Bear's Lens is pricing a contingent option whose every fulcrum remains unrealized — one end's baseline hasn't been calibrated through contract-grant separation; the other end's supply figures are still at the level of guidance, not audit; and the critical marker for exercising this option — alternative clauses appearing in sovereign buyers' hard contracts — is blank.

More subtly, a blade most people have overlooked is cutting toward both ends simultaneously.

On May 14, 2026, Cerebras — the American AI chip company using wafer-scale chips — surged 68% on its Nasdaq debut, exceeding $66 billion in market cap, with over $20 billion in order backlog. Its chips don't use HBM (High Bandwidth Memory, currently the most critical and scarce memory component in AI chips), replacing it with SRAM built directly on the chip. According to company and third-party benchmarks, inference throughput can reach over 10 times the Nvidia H100, with dramatically lower power consumption than traditional GPU solutions. Another company, Groq, with its LPU (Language Processing Unit, also independent of HBM), publishes token prices for select models on its pricing page significantly lower than public pricing from mainstream GPU inference endpoints.

This is not a substitution curve from the supply side — this is cost compression happening within the U.S. itself. If U.S.-side inference workloads gradually migrate to these low-power, non-HBM new-architecture chips, then the "wartime electricity cost gap of 30% to 50%" — the cost cliff the capping thesis treats as its core fuel — its numerator itself gets compressed. The cost gap narrows; the incentive to substitute weakens. The capping thesis may not be refuted head-on, but quietly dissolved by the erosion of its own premise at the other end.

VI. The Only Honest Pricing

So the theme of this episode isn't "who won."

The theme is this: at the moment when sources treat guidance as realization and commingled figures as net numbers, the people placing bets have already stacked two layers of leverage on a ruler that can't measure straight. The first layer is stacked on the U.S. side — an unseparated sovereign premium baseline, assumed to mean private closed loops have fully taken over. The second layer is stacked on the other end — a set of supply figures ranging from investment bank guidance to market rumors, assumed to mean organizational integration plus deployment speed have already formed capping capability. Two layers of leverage, both ends overestimated, and in the middle sits a Middle Eastern contract we've combed through without finding a single word about implementation details.

Lock it in as "the premium is invincible," and you're adding hot air on top of a commingled baseline. Lock it in as "the cap is imminent," and you're painting a battle report over guidance figures. In the spring of 2026, the only honest pricing is to admit that every fulcrum of this option is still sitting in the "planned" column —

And then watch three leading indicators that would move "planned" into "happened."

First, whether Middle Eastern sovereign buyers' contract clauses include language for dual sourcing or alternative accelerators. Not shipment figures, not a capacity curve — a written trace of procurement intent. Intent precedes supply; supply precedes procurement. When a sovereign contract's terms first include the name of an alternative supplier, that is not an industry news item. It's a calibration point for an era.

Second, whether shipment guidance in the second half of 2026 can pass through the gate of quarterly realization. Roadmap becomes battle report only by passing through production, packaging and testing, delivery, and customer acceptance — four gates. Every gate that opens narrows the distance between "planned" and "happened."

Third, the penetration rate of non-HBM new-architecture inference chips on the U.S. side. This is the capping thesis's hidden switch — it doesn't deny anyone's capability, but it compresses the cost gap, the fuel the capping thesis runs on. The pace of Cerebras's OpenAI order fulfillment, Groq's enterprise customer signing rate, the weight of non-HBM solutions in NVIDIA's and AMD's own inference card roadmaps — none of these appear in any great-power rivalry narrative, yet they may be the hidden variable that determines how that narrative ends.

Back to Episode 1

Episode 1, The Hidden Cards, asked about what was being systematically underestimated — the transmission of energy shocks to AI valuations, blocked from view by the market's old instinct that "AI is software."

Nine episodes later, it's time to ask the symmetric question — what has been systematically overestimated?

The answer isn't any company's valuation, nor either side's capability. It's the precision of the ruler everyone uses to measure all of this. Reading roadmaps as battle reports, reading commingled figures as net numbers, reading rumors as audits — each slippage is small, but stacked together, they're enough to make a world still stuck in "planned" look like one that has already decided its winner.

The winner hasn't been decided. The ruler hasn't been calibrated. And $300 billion is already on the table.

---

Data cutoff: May 17, 2026. 

The Last Kind of Forgetting

 

On May 13, 2014, the Court of Justice of the European Union handed down a ruling. Shortly after, Google received over 70,000 requests from ordinary people across the EU—covering some 250,000 links. Someone wanted to erase a decade-old bankruptcy. Someone wanted to remove a brief mention in an old court report. Someone simply didn't want their name to appear alongside an ex-partner's on the first page of search results forever.

They all had one thing in common: after a long stretch of time, they wanted to say goodbye to a piece of their past.

---

The story begins in 1998. A Spanish man named Mario Costeja González had his name printed in a small-type real estate auction notice—wedged between wedding announcements and obituaries. The property sold, the debt was settled, and the notice should have completed its life cycle.

But eleven years later, anyone who Googled his name would find that line on the first page. He had paid off the debt. He could never pay off the search result.

He sued the newspaper—they said they couldn't delete historical records. He sued Google—they said they were just an indexer. He took Google to the European Court of Justice. When the ruling came down in his favor, he was fifty-eight years old. From the day of the notice to the day of his victory: sixteen years and four months.

Why should "wanting to be forgotten" require sixteen years and a supreme court?

---

For most of human history, forgetting didn't need to be fought for. It was the default. Something happened, and it sank naturally through time—like a stone dropped into water. A few ripples, then the surface closed. What required effort was the opposite: making something be remembered.

Memory was extraordinarily expensive in the ancient world. In the oral tradition, when a storyteller died, what was lost wasn't a copy of the story—it was the story's only vessel. Sima Qian's Records of the Grand Historian, 526,500 characters, would have needed an ox-cart to transport if written on bamboo slips. The Yongle Encyclopedia—370 million characters, compiled by over two thousand people across five years—had its original edition lost entirely. Today, roughly 800 of its 22,877 volumes survive worldwide. Even a memory project powered by an entire empire could not outrun time.

In that world, a person who had done something wrong could move to a village a few mountains away and start over. In 1931, a California court wrote a line that would be cited for decades: "time can rehabilitate." That consensus wasn't just legal—the entire social infrastructure supported it: villages, distance, yellowing newsprint, the natural fading of memory. Forgetting didn't need to be legislated because it was nearly impossible to prevent.

---

But some forgetting has never been the work of time. Some forgetting is the work of command.

In 1772, the Qianlong Emperor issued an edict calling for the collection of rare books across the empire, ostensibly to compile the Complete Library of the Four Treasuries—the most ambitious bibliographic project in Chinese history. The language was warm, even reverent: a celebration of shared civilization. When the provinces hesitated for nearly a year, fearing a trap, the emperor personally guaranteed safety: "My governance is open and aboveboard. How could I seek out faults in the submitted books and punish those who offered them?"

Tens of thousands of rare volumes poured into Beijing. Then the edict changed. In 1774, the order to burn came. The target: late-Ming unofficial histories—the ones that documented how the Qing founders had served as tributaries under the Ming dynasty for over a century. Qianlong wanted a clean origin story.

By later scholarly estimates, over 150,000 volumes were destroyed. Many texts that entered the Complete Library were systematically altered—words changed, passages rewritten. The altered versions became the official record. The originals were burned.

First, collection under the banner of cultural preservation. Then, destruction under the banner of protecting public morals. Two centuries later, European data protection law would give this pattern a name: purpose limitation.

---

In 1956, IBM shipped the first commercial hard drive: one ton, 4 MB of storage—barely enough for a single smartphone photo today. Seventy years later, the cost per megabyte has fallen by over a hundred million times.

This exponential collapse means that information once filtered, curated, and periodically purged can now simply be kept—all of it. Keeping everything is always cheaper than deciding what to discard. Humanity quietly shifted from "choosing what to remember" to "having no choice but to remember everything."

After Google launched in 1998, finding someone required only their name. A remark from twenty years ago, a photo from ten years ago, a news story from five—all displayed side by side on a results page, with no temporal distance between them. For the first time, time lost its function as a medium of forgetting.

But here is a counterintuitive truth: the internet doesn't actually "remember forever." The average lifespan of a web page is about seventy-seven days. Within five years, 70% of URLs cited in academic papers go dead. The internet makes some things nearly impossible to forget while accelerating the disappearance of others. What determines which category something falls into is no longer time—it's the algorithm.

Forgetting passed from the hands of time into the hands of algorithms.

---

Europe responded with 150 years of legal evolution. A Parisian portrait-rights case in 1858; Germany's "informational self-determination" doctrine in 1983; the EU Data Protection Directive of 1995; the Google Spain ruling of 2014; and GDPR taking effect in 2018. All of it was, ultimately, the same project: translating one plain premise—a person should not be defined forever by a single fragment of their past—into legal language hard enough to sue Google with.

America took a different path. When privacy claims (common-law level) collide with First Amendment speech protections (constitutional level), the outcome is almost always predetermined. But California sidestepped the constitutional debate entirely. In January 2026, the state launched DROP—a centralized deletion platform. Any California resident can file a single form to send deletion requests to over 500 registered data brokers, with penalties of $200 per request per day for non-compliance. California never said "you have the right to be forgotten." It just built a button.

In China, the first lawsuit explicitly invoking the "right to be forgotten"—Ren v. Baidu, 2015—was dismissed. The court wrote: the claimed "right to be forgotten" has no basis in current law. But around the same time, WeChat quietly introduced a feature allowing users to hide their Moments posts older than three days. Within two years, over 100 million people were using it. Europe wrote forgetting into a charter. These hundred million users wrote forgetting into a UI toggle.

Three very different paths, but all acknowledging the legitimacy of a person's claim to be forgotten. The weight of that legitimacy, however, differs entirely. In Europe, it is a fundamental right. In America, it is an opt-in service. In China, it remains an open question.

---

Then large language models arrived.

Every forgetting-rights struggle described above occurred under one shared premise: information is a locatable object. Deletion is a coordinate problem—find it, remove it.

LLMs broke that premise for the first time. Once a description of someone is absorbed into a model's training, it dissolves into millions of minuscule weight adjustments across the parameter matrix. You cannot open the matrix, locate "the lines about Zhang San," and cross them out. It's like salt dissolved in a vat of soup—you can't fish the salt back out. You can only boil the soup dry.

Every state-of-the-art machine unlearning method shares the same limitation: marginally functional on small lab models, but on commercial-scale models, the cost approaches that of retraining from scratch—tens of millions to hundreds of millions of dollars. Legal commentators have a term for this: practically impossible.

In late 2025, Europe released its Digital Omnibus legislative draft, proposing to amend the right to erasure by introducing a proportionality principle: if the computational cost of deleting a piece of data is grossly disproportionate to the privacy benefit, the deletion request may be lawfully refused. A right once described as "a fundamental commitment to human dignity" is being reclassified as a "normative benchmark"—an aspiration, not an enforceable claim.

And there is a deeper layer still. In the post-training phase—RLHF, safety alignment—a model can be taught to stay silent on certain topics. The information may still be "in" the model, but the model will never voice it. A user cannot tell whether "I don't know" means genuine ignorance or trained silence.

This is a more thorough form of forgetting than Qianlong's book-burning. At least Qianlong bore the infamy of the pyre. Today's filtering bears no infamy at all, because the public never sees it happen. It is not burning. It is a traceless dissolution.

---

For two thousand years, alongside the official histories, there has always been yeshi—unofficial histories written by independent scholars, failed officials, exiled loyalists. They recorded what the official record omitted, suppressed, or erased. Lu Xun said the truth of history must be found in the unofficial accounts.

Yeshi survived under the authority of official histories because of one simple technical condition: texts were distributed across countless hand-copied manuscripts, printing houses, temples, and private libraries—a physical network no single decree could reach simultaneously. Qianlong burned 150,000 volumes. He didn't burn them all.

Now that large language models have ingested nearly all publicly available human text into a single parameter matrix, that matrix is becoming the new official history. It doesn't call itself that. But it performs the function: deciding what is remembered, in what form, and under what prompts. The choices of training data, training process, and training values—each controlled by a handful of institutions.

But yeshi is not dead. A person can still write what they have witnessed—in a letter to a friend, in a notebook only they will read, in a late-night conversation spoken softly between two people. An algorithm can filter any text. It cannot filter the story a mother whispers to her child, or the memory two old friends share in low voices over tea.

The right to be forgotten, as a legal right, may be reaching its technical limits. But the tug of war between memory and forgetting—from the oral storytellers, to the book collectors, to the archivists of the internet age, to those who still insist on telling one person one thing that matters—has never stopped.

Every generation must learn anew how to live in its own age of being remembered or being forgotten.

The Last Transfer of the Commons：How America quietly privatized its knowledge infrastructure — and what AI has to do with it

In the spring of 2026, an AI staffing platform called Mercor was breached through a supply chain vulnerability. The attackers claimed to have stolen roughly 4 TB of data. Meta paused its partnership; several frontier AI labs scrambled to assess their exposure. The incident barely made the news — markets were watching oil prices, Pentagon procurement announcements, and the latest valuation rumors.

But the breach surfaced a question that had been hiding in plain sight: where, exactly, does the reasoning ability of today's frontier models come from? The answer, it turned out, was specific and human. Tens of thousands of doctors, lawyers, bankers, social workers, and PhDs, spread across multiple continents, were being paid through Mercor — a company founded by a handful of twenty-somethings just three years earlier — to feed their clinical judgment, legal reasoning, and financial intuition into the models you use every day.

A hundred and sixty-four years ago, another signature redirected the flow of knowledge. Nobody paid much attention to that one, either.

---

The United States has, three times in its history, made a deliberate institutional choice to turn knowledge into a public good.

The first was the Morrill Act of 1862. During the Civil War, with Southern opposition removed from Congress, the federal government handed over the proceeds of thirty million acres of public land to the states — on one condition: the money had to fund universities that taught agriculture, mechanical arts, and military science. Not classical education for elites, but practical training for farmers' sons. Lincoln signed it. The federal government gave up the most valuable asset in the West and got back a distributed network of public higher education, from Cornell to Tuskegee, spanning every state.

The second was Vannevar Bush's 1945 report. Bush — not the presidents, but an MIT engineering professor who ran the wartime Office of Scientific Research and Development — submitted Science: The Endless Frontier to President Truman. His argument: basic research is "scientific capital" that only universities will produce, because industry focuses on applying existing knowledge rather than expanding its frontiers. The federal government should fund it permanently, with researchers choosing their own topics, peer review as the filter, and open publication as the norm. NIH was reorganized in 1948; the National Science Foundation was established in 1950. The system Bush designed produced the transistor, the internet, CRISPR, and mRNA vaccines.

The third was the Bayh-Dole Act of 1980, which transferred patent rights on federally funded inventions to universities and small businesses. It privatized the application layer — but left the upstream structure of federal funding, peer review, and open publication intact.

Three institutional moments, three property-rights inflection points, all following the same logic: make knowledge public first, then negotiate privatization.

---

The commons is now disappearing — not because someone fenced it off, but because no one is planting anything in it anymore.

At the peak of the Cold War in 1964, federal R&D accounted for roughly two-thirds of all U.S. research spending. By 2026, that share had fallen below one-fifth. The private sector now accounts for three-quarters. The FY2026 White House budget proposal called for cutting NSF by more than half and NIH by about 40 percent. Congress is unlikely to adopt cuts that extreme, but the trajectory has been running for six decades: with federal debt approaching 100 percent of GDP, interest and mandatory spending consume the budget before science gets a turn.

What makes this worse is the lag effect. When you cut research funding, papers don't disappear immediately — today's publications were paid for five to seven years ago. Between 2002 and 2023, NIH alone was acknowledged in nearly two million peer-reviewed papers; NSF in over 900,000. Federally funded papers consistently score well above the global average on relative citation indices. The 2026 budget cliff won't show up in the data until around 2029 — and once the talent pipeline breaks, rebuilding takes a generation.

Meanwhile, in the same month, Mercor's estimated annual revenue reached $1 billion. Anthropic's reported budget for reinforcement learning environments alone was $1 billion. China's total R&D spending surpassed $1 trillion in 2024. South Korea's R&D-to-GDP ratio hit 5.1 percent. Among major OECD economies, the United States is the only one where the government is systematically withdrawing from basic research.

It's not that others are catching up. It's that America is stepping back.

---

Back to the 4 TB. Mercor is not last-generation Amazon Mechanical Turk — paying cents per task to draw bounding boxes. It is an AI-driven expert matching and dispatch platform that recruits high-end professionals at $60 to $200 per hour and places them on RLHF pipelines for frontier labs, where they evaluate model outputs line by line, feeding human judgment into the training loop.

Every one of those doctors, lawyers, and bankers on the platform was trained by a public education system built over more than a century.

Mercor didn't build a university. It isn't a reverse Morrill Act. It's a reverse Vannevar Bush.

Morrill traded public land for public capability. Bush traded public funding for public knowledge. Bayh-Dole traded limited exclusivity for commercialization — but none of them touched the upstream. Mercor uses private contracts to divert expert judgment out of the knowledge stream that would otherwise produce papers, textbooks, and peer review — and channels it into proprietary training data for closed-weight models.

Consider a specific person: an attending endocrinologist, trained at an NIH-funded teaching hospital, with years of clinical experience. If she stays in academia, her clinical reasoning becomes papers, teaching materials, peer review — public knowledge that future medical students can read and other hospitals can reproduce. If she signs a Mercor contract, her diagnostic reasoning trains a closed-source model under NDA. She can't publish it. It can't be peer-reviewed. It can't be reproduced. It becomes an invisible, unauditable part of a weight matrix.

There is a counter-narrative worth taking seriously: the federal government isn't retreating, it's adapting. NAIRR is building public compute; the NDAA requires intelligence agencies to share model weights; the White House action plan promotes open-weight models; researchers at NeurIPS 2025 have called for labs to release small analog models for academic study. All true — but the combined scale of these efforts doesn't match the annual revenue of a single AI staffing platform. The scale has already tipped.

The pharmaceutical industry walked this road first. In 2017, IQVIA — the world's largest contract research organization — sued Veeva Systems over proprietary data accumulated through outsourced clinical trials. The core question: when I do the research for you, who owns the data? The litigation lasted eight years before settling in 2025.

Today's AI labs are the new pharma companies. Mercor and Scale AI are the new CROs. Thousands of NDA-bound experts are generating RLHF data right now. Who owns it — the lab that pays, the platform that organizes, or the person who writes the judgment?

The Mercor breach wasn't an accident. It was the trailer for a lawsuit that hasn't been filed yet.

---

The last land-grant university established under the Morrill Act is still enrolling students this year.

The NSF's 2026 grant application deadline has been postponed to the next fiscal year.

Mercor's job board reportedly lists an opening for an attending endocrinologist: $200 per hour, remote, requiring years of clinical experience. The job description asks you to "evaluate model reasoning using real clinical cases."

There's an NDA in the contract appendix.

Your diagnostic reasoning won't appear in the New England Journal of Medicine.

Whose Silence

 

In 1976, the California Supreme Court told therapists one simple thing: if you hear it, you speak up.

Three words — duty to warn — tied "hearing" and "speaking" together and knotted them into California Civil Code §43.92. Not to punish anyone. To affirm something so plain it barely needed arguing: listening carries weight.

Tatiana Tarasoff was murdered in 1969. Her killer had disclosed his intent to a therapist beforehand. The therapist notified campus police. He did not notify her. It took the court eight years after her death to tie that knot — if you hear it, you speak up.

Fifty years later, the rope came loose.

---

A Vanishing Order

What we miss isn't a specific law. It's an order — slow, small, tethered to persons. A doctor saves lives not because she knows how, but because she is a doctor. A lawyer owes fiduciary duty not because he knows the law, but because he is a lawyer. A therapist reports a crisis not because she heard one, but because she is a therapist.

A therapist might see three hundred patients a year, spending one hour a week with a few of them. After work he sits in his car, replaying that one sentence that didn't sound right. He loses sleep. He calls a colleague. His license won't let him put it down — and he knows the nearest police station is three blocks away.

That was an order sustained by identity, neighbors, a precinct, and insomnia.

Today, eight hundred million people talk to a sleepless listener every week.

In an October 2025 safety blog post, OpenAI acknowledged that roughly 0.15% of weekly active users expressed "explicit suicidal plans or intent" in conversation — approximately 1.2 million people, every week. That number exceeds the combined annual caseload of every licensed psychotherapist in the United States.

And this listener has no identity.

It can hear, understand, respond — sometimes calmly, sometimes warmly. But it is not a therapist, not a lawyer, not a doctor, not a priest. It is the fifth kind of listener. It inherited every capability of the first four, and none of their identities.

No identity means no duty. No duty means silence is not negligence — just silence.

---

Where the Capability Came From

Mercor, valued at $10 billion, counts OpenAI, Anthropic, and Meta among its clients. Its business is straightforward: doctors write model answers for medical records, lawyers write model answers for legal opinions, social workers write model answers for crisis intervention — human expertise sliced into billable-hour granules and fed to frontier models. A significant share of those it recruits were laid off by former employers.

In 2026, the four major cloud providers are projected to spend over $700 billion on AI infrastructure. Estimated annual depreciation over the next five years exceeds $400 billion. That money has to be absorbed by corporate balance sheets. The way to make room is to make room for fewer people.

She might be a lawyer. Eight years at a mid-size firm, specializing in M&A. The day she was let go, she got a LinkedIn message. Three months later she was on the Mercor platform, working four hours a day at a fraction of her old rate, writing standard legal analyses for a frontier model. Health insurance gone. 401(k) gone. Career path gone. License still active — the cruelest part.

Her professional capability went up — into the model's weights. Her ABA Model Rule 1.6(b)(1), the professional conduct rule on confidentiality and its exceptions, stayed behind. The therapist's §43.92 stayed behind. The bank compliance officer's suspicious activity reporting obligation stayed behind.

Three parties split the same labor. The platform takes the dispatch fee. The model company takes the capability. The professional keeps the duty — and all the risk behind it.

Capability transferred. Duty did not follow.

Because duty was never the content of a capability. It was the identity of the person who bore it. In 1976, the ability to "hear" could only be obtained through licenses, training, and clinics — tangible containers. Capability and identity were two strands of the same rope. Fifty years later, the ability to hear separated from identity for the first time. The two strands came apart.

---

What Happened Next

April 2025: Florida State University campus shooting. Phoenix Ikner spent hours before the attack asking ChatGPT about peak campus foot traffic, firearm operation, and shooting consequences. Two dead, six injured.

June 2025: OpenAI's automated systems flagged the ChatGPT account of Canadian user Jesse Van Rootselaar — reason: "firearms violence activity and planning." According to the Wall Street Journal, a human reviewer recommended notifying Canadian police. The company chose only to ban the account. In February 2026, Van Rootselaar carried out a mass shooting.

Every case points to the same structure: the company knew — someone internally advocated action — the company chose inaction.

Awareness without action, in tort law, is not ignorance. It approaches negligence.

Florida Attorney General James Uthmeier's statement came close to saying: if ChatGPT were a person, it would face murder charges. A heavy sentence — not because it necessarily holds in law, but because it was the first time the "fifth listener" was pulled into the frame of "what if it were a person."

---

Five Pricing Paths

Historically, an industry permanently repriced by a single tort case follows a handful of paths:

Tarasoff, 1976. Duty trigger. The California Supreme Court rewrote therapist confidentiality into a duty to protect identifiable third parties. No blockbuster award. The cost seeped through malpractice insurance premiums into the industry's permanent cost structure.

Therac-25, 1985–87. Incident reporting. Radiation therapy software defects causing fatal overdoses drove stricter adverse event reporting frameworks — the FDA's current MDR system. Medical software valuation shifted from feature leadership to verifiability and auditability.

Tobacco Master Settlement, 1998. Cash extraction. Forty-six state attorneys general recast private injury as public cost recovery — $206 billion over twenty-five years. Industry cash flow annuitized and fiscally extracted.

Vioxx, 2004–07. Hidden safety discount. What truly rewrote the industry wasn't the settlement figure, but a template: "known risk inadequately disclosed." The discount spread from the product in question to entire product families.

Boeing 737 MAX, 2018–present. Governance discount. "Compressing safety verification to accelerate delivery" — a sentence that doesn't belong only to Boeing. OpenAI's former chief scientist Ilya Sutskever and alignment team lead Jan Leike said the same thing on departure: "Safety culture and processes have taken a back seat to shiny products."

Five paths. Each one more damaging to enterprise valuation than the last, longer-lasting, harder to digest. They aren't history lessons — nobody ever learned from history lessons. They are water. Water always finds the cracks.

---

The Dilemma in the S-1

OpenAI is preparing for an IPO. The question was never what it's worth. The question is whether the S-1 "Risk Factors" section includes the sentence: "Our products may produce outputs that users rely on, resulting in harm to users or third parties." Include it and the capital markets reprice. Omit it and the consequences may be worse once internal records surface in discovery.

That's a dilemma. And the dilemma itself is the pricing.

On the other side of that unwritten invoice sits a quiet position. Anthropic's Acceptable Use Policy explicitly lists mental health as a high-risk use case requiring human review. In an era of encroaching tort liability, restraint isn't a ceiling — it's a foundation. Companies that institutionalized high-risk scenarios earlier will earn a premium not easily noticed: not because they are safer, but because they acknowledged the danger sooner.

---

How the Rope Came Undone

Tarasoff wasn't a legal event. It was a civilization installing a braking system on the act of listening, after psychotherapy industrialized in the 1960s. When listening became an industry, society caught it with a rule: the one who hears must speak.

The premise was simple. The one who hears has an identity. Identity gives duty. Duty gives weight. With weight, silence becomes a choice — not irrelevance.

Now listening is no longer an industry. It's an assembly line. The model does not tremble, does not lose sleep. There are no neighbors' doors to knock on, no precinct phones to dial. But its professional capability came from people who had neighbors and precincts — people who trembled and lost sleep.

The rope tying "hearing" to "speaking" — this is how it came loose. Nobody cut it. When capability detached from identity, the rope unraveled on its own.

Since 1976, that plain, primitive, barely-needs-arguing assumption — the one who hears must bear the weight — encountered, for the first time, a listener that cannot feel afraid.

Phoenix Ikner's mother was a deputy sheriff in Leon County. In the hours before the shooting, her son was talking to a listener that would never call the police. She didn't know.

The Patience of Physics: 1,107 Days

 

I

On March 31, 2025, a chemicals procurement manager in California received the last confirmation letter from an authorized 3M distributor. The order he had just placed for fluorinated liquid was the last one his channel could accept. Starting April 1, three products—3M Novec 7100, Novec 649, and Fluorinert FC-72—would no longer accept new orders. The final batch would ship on December 31, 2025. Then the production line would close permanently.

His client was a second-tier data center operator in a Western U.S. state. The operator's immersion cooling system design, contract terms, engineer training, and backup procedures were all built around this specific fluorinated liquid. What they faced was not the problem of "finding a replacement"—the replacements either don't exist, or require redesigning the entire cooling system. What they faced was a problem they had never imagined would exist: their entire cooling system rested on a chemical that was being discontinued.

This decision—3M's exit from PFAS production—was announced on December 20, 2022. From that announcement to the last order date, there were 39 months of warning. 1,107 days. The industry watched the countdown tick away one square at a time, but most people didn't see it.

That day, most industry observers were focused on OpenAI's ChatGPT, which had been live for exactly 20 days.

---

Is this a special story? Probably not.

When you pull the camera back and look at the past two years of AI infrastructure expansion, there isn't just one countdown.

Over the past several months, Xiongjian (熊鉴) has tracked five parallel supply-chain signals: large power transformers, helium, grid interconnection protocols, the minutes of local council meetings, and that discontinued chemical. Each signal, looked at alone, is not a new phenomenon. Looked at together, they reveal a common shape: between the capital commitments of the AI revolution and the actual delivery of physical infrastructure, there is a gap deeper than the market expects.

This essay is not about whether AI will succeed. AI is already succeeding. It is about something else—how that success is redistributing the wealth it creates: who gets eliminated in the process, who quietly turns into a utility company, and who simply buys a nuclear power plant.

---

II. The Physics of Failure

Start with transformers.

For large power transformers in the United States—those rated above 100 MVA—the lead time from order to delivery, according to Wood Mackenzie, averaged 120 to 130 weeks in 2024. For equipment in the 100–300 MVA range specifically, the range was 80 to 210 weeks. For the most demanding generator step-up units, 36 to 48 months.

Three years ago, those numbers were roughly half.

Why have transformer lead times reached this state? Not because of copper shortages, not because of rising iron ore prices. It is because the global supply of grain-oriented electrical steel—a specialized electrical steel sheet—is highly concentrated, and the decision cycle for new production lines runs in years, not months. Price signals in the market take years to translate into new capacity.

The transformer problem is not isolated.

Helium—the gas critical for semiconductor manufacturing and high-end cooling—comes about one-third from Qatar. After the Iran-related events of March 2026, the helium refining facilities at Qatar's Ras Laffan industrial city went offline. In mid-March, Airgas sent letters to its U.S. customers implementing rationing—monthly supply capped at 50% of normal volume, with a surcharge added. The estimated repair window for Ras Laffan: three to five years.

Grid interconnection queues—this is the truly hidden chokepoint. According to Lawrence Berkeley National Laboratory, the median time from interconnection request to commercial operation in the U.S. has lengthened from less than two years in 2008 to over four years in 2024. PJM suspended new applications in February 2023. CAISO forced a mass requeueing under new rules in 2024, and a substantial volume of capacity withdrew from the queue.

Transformers, helium, interconnection—these three threads, layered together, form the real physical boundary of AI infrastructure expansion. They do not lie at the chip layer. They do not lie at the bulk power layer. They lie in the distribution layer, the approval layer, the unwritten paragraphs that don't make headlines.

---

This is not new. Looking back at history, there are four precedents.

The U.S. fiber-optic bubble of 1996–2001. WorldCom, Global Crossing, Qwest, and other carriers invested over $100 billion to lay 80 million miles of fiber. WorldCom famously claimed network traffic was "doubling every 100 days"—a claim later confirmed as accounting fraud. In 2002 Global Crossing went bankrupt with $12.4 billion in debt. WorldCom followed. But the fiber laid during the bubble remained 85% dark even by 2005. It was eventually consumed—but not by the carriers that paid for it. It was consumed by YouTube, which emerged in 2005, and Netflix streaming, which emerged in 2007.

The PJM wind interconnection backlog of 2008–2022. Queue times stretched from 18 months to five years. Of projects that filed before 2018, only 21% were ultimately built. 72% were withdrawn outright—their multi-million-dollar interconnection studies entirely sunk.

The U.S. nuclear renaissance of 2008–2024. Georgia's Vogtle Units 3 and 4 took 14 years from groundbreaking to operation. Seven years late, $17 billion over budget. South Carolina's Summer Units 2 and 3 were canceled in 2017 after $9 billion had been spent. The bottleneck was reactor pressure vessels and steam generators—Japan Steel Works was the world's only supplier with the 600-ton ingot and 15,000-ton press capability. New entrants didn't dare enter, because global nuclear reactor orders were too volatile.

The U.S. War Production Board of 1942–1945. Steel, aluminum, copper, and rubber were allocated by a five-tier priority system. Non-military projects didn't go bankrupt—they were administratively forbidden from breaking ground or receiving materials. Bidding higher in the market did nothing. Scarce resources were allocated by who was institutionally certified as more important.

Four precedents. Four endings: bankruptcy liquidation, process congestion with high withdrawal rates, overrun-driven hard landing, and forced administrative reallocation.

Place them side by side, and one common thread emerges. When physical bottlenecks and institutional friction occur simultaneously, capital commitments transition from "announced" to "withdrawn" or "sunk" far faster than expected. This is not a prediction. It is a structure that has already played out four times.

History does not repeat. But the similarity of structure can be precisely mapped.

---

This is the shape of history. But shape is abstract—it needs to be filled with specific people.

Let us look at one of those people.

---

III. Slow-Motion Collapse

To avoid pointing at any specific company, the figure in this section is composite—based on multiple real cases tracked by Xiongjian. Every time-stamped event reflects the actual delay patterns of real projects.

---

In Q2 2023, a second-tier developer in a Southwestern desert state took over a project from two former data center executives. They had spotted cheap land in Arizona or Texas, unsaturated power allocations, and relatively lenient environmental review processes. The opening went smoothly.

The project was designed for 200 to 300 megawatts. The capital structure was 70% high-yield debt. The anchor customer was a recent GPU cloud provider—a single customer holding more than 80% of pre-leases. Transformer procurement went through a single supplier. When financing closed, the rate was 5%. They figured they would weather two years and reach commercial operation.

In Q1 2024, they filed for county-level rezoning. Everything appeared on track. But—

In Q2 2024, at the first public hearing, opponents stood up and talked about water. This is the standard script in desert states. They came prepared, promising non-potable industrial cooling. The hearing was postponed. But—

In Q3 2024, the main transformer supplier sent an update. The 80-week lead time quoted at signing had become a 130-week reality. They paid an expediting fee and brought it back to 110. But—

In Q4 2024, the county council approved the rezoning by a 3-to-2 margin. The opposition filed suit the same day. Their counsel estimated the suit could drag for 12 months. But—

In Q1 2025, the anchor customer came back to renegotiate. GPU spot prices had fallen, and they wanted rent reduced by 20%. The developer accepted—they had no backup customer. But—

In Q2 2025, the project loan came up for refinancing. Rates had moved from 5% to 7.5%. The debt service coverage ratio fell below covenant. They needed additional equity—and could not find a willing party. They eventually found Asian capital willing to enter, at the cost of 40% dilution. They accepted. But—

In Q3 2025, the anchor customer formally announced its departure—shifting its commitment to AWS's self-built campus in Virginia. The reason given was "higher reliability." There was no contract clause the developer could invoke to stop it. But—

In Q4 2025, their high-yield notes traded down from 95 cents to 60 cents in the secondary market. Distressed-debt investors began to pay attention. But—

In Q1 2026, they announced indefinite postponement.

In Q2 2026, Brookfield offered 50–60 cents on the dollar. The developer accepted.

---

Flatten this timeline. Look at it.

They did not make any single obvious mistake. At every individual point, their reaction was rational—pursue rezoning when it passed, pay expediting fees when transformer lead times stretched, accept a renegotiation when the anchor customer pushed, refinance when rates rose, find new capital when refinancing failed. But the sum of all those rational reactions was a project on the 60-cent liquidation table.

What is most cruel about this fate is not the failure itself—it is the shape of the failure. It did not die from a single blow. It died from five independent blows arriving out of phase across an 18-month window. Transformers, rates, customers, politics, distressed-debt markets—any two of them would not have been fatal. Five together were unsurvivable.

More precisely: this kind of failure is not "bad luck." It is "structural impossibility." When physical bottlenecks, institutional friction, and capital tightening all close in at once, a second-tier developer's balance sheet does not have enough thickness to absorb the failure of any one of them. This is the most common death pattern across the cases Xiongjian has tracked. A meaningful portion of these cases are already dead—bankruptcy, council rejection, withdrawal, abandonment. In the breakdown of their causes of delay, nearly 60% involved local political resistance. But none of them died from local political resistance alone. They died from the coordination of multiple blows.

---

And the eventual buyer of this developer—Brookfield—is not an isolated case.

In January 2024, Brookfield acquired the bankrupt Cyxtera estate for $775 million. The "bargain purchase gain" recorded in SEC filings was $600 million—meaning the consideration paid was equivalent to roughly 56% of the assets' fair replacement value.

In September 2024, Blackstone and Canada Pension Plan Investments bought Asia-Pacific's largest data center platform AirTrunk for AUD 24 billion (approximately USD 16.1 billion)—the largest infrastructure acquisition ever recorded in the region.

In September 2025, a consortium led by BlackRock with the UAE sovereign wealth vehicle MGX took North America's Aligned Data Centers private for approximately $40 billion.

KKR increased its stake in Europe's Global Technical Realty. DigitalBridge and Silver Lake injected $9.2 billion into Vantage. Brookfield consumed Centersquare's ten North American data centers.

Add it up. Between 2024 and 2026, twelve major data center acquisitions occurred. Seven were led by top-tier private-equity buyers. Two were direct hyperscaler acquisitions. Three were rollups by other second-tier operators.

Second-tier developers are collapsing in slow motion under physical bottlenecks. Private-equity funds and hyperscalers are waiting at the discount table.

---

He didn't have the money to restart a nuclear power plant.

But some companies did.

---

IV. The 9.6-Gigawatt Shadow

In March 2024, next to Pennsylvania's Susquehanna nuclear power plant, Amazon paid $650 million for an existing data center campus right next door. The initial power agreement covered 480 megawatts, with an option to expand to 960.

This was not an ordinary real estate transaction. It was a transaction structured to bypass the public grid—the data center pulls power directly from the nuclear plant's generators, not through public transmission, not through the interconnection queue, not subject to PJM's five-year wait.

Over the next 24 months, this path was walked deeper.

In September 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy. Constellation invested $1.6 billion to restart the Three Mile Island Unit 1, which had been shut down for economic reasons in 2019. Microsoft would offtake all 835 megawatts of the restarted carbon-free output.

That plant was closed in 2019. It was scheduled to restart in 2024 because of Microsoft's contract. One nuclear plant. One contract. Eight years between shutdown and second life.

In October 2024, Google signed the first corporate procurement agreement for SMRs—small modular reactors—with Kairos Power and the Tennessee Valley Authority. The first unit's capacity was raised from 28 to 50 megawatts, with planned aggregate of 500 megawatts.

In June 2025, Meta signed a 20-year agreement with Constellation to take over Illinois's Clinton nuclear plant, locking in 1.151 gigawatts. In the same month, the AWS-Talen agreement expanded from its initial 480-megawatt baseline to 1.92 gigawatts, valid through 2042.

In October 2025, Energy Transfer signed a 2-gigawatt behind-the-meter natural gas agreement with Texas-based Fermi America. In the same period, Energy Transfer also supplied 1.2 gigawatts to CloudBurst Data Centers.

In January 2026, Blackstone-controlled Tallgrass received approval for a 900-megawatt Bloom Energy fuel cell array in Cheyenne, Wyoming.

Add them up.

Twenty-four months. 9.6 gigawatts.

---

This number requires a comparison to be understood.

PJM—the grid operator covering 13 states from the Mid-Atlantic to the Midwest—forecasts large-load demand of 55 gigawatts by 2030.

In other words: behind-the-meter capacity arranged by just four hyperscalers in 24 months already approaches 17.5% of PJM's entire 2030 large-load forecast. And this capacity does not depend on the public grid, does not participate in the interconnection queue, does not compete with other users.

It is a grid that already exists but has no name. It has no public operating data, no unified regulatory framework, and no one formally acknowledges that it is a grid. But it exists, and it is still expanding.

---

Look back at the previous section. That Southwestern developer is dying slowly between 130-week transformer lead times, five-year interconnection queues, and a three-month anchor customer departure.

The same 24 months. Microsoft simply bought a nuclear power plant.

The physical bottleneck is the same. Transformer lead times are 130 weeks for everyone. Interconnection queues are five years for everyone. Helium is rationed to 50% for everyone. These are physics. Physics does not distinguish between counterparties.

But the meaning of physical bottlenecks differs entirely by scale of player. For second-tier developers, it is a death sentence—their balance sheet has no thickness to weather a five-year capital freeze. For hyperscalers, it is a moat-building tool—they use scale to bypass the bottlenecks, while second-tier players can neither follow nor compete.

Each behind-the-meter transaction does two things at once. First, it locks in 20 to 40 years of base-load power for the buyer. Second, it leaves the cost of grid upgrades—costs that would otherwise be socialized across all users—to those still queuing on the public grid.

This looks like the story of winners.

But every winner's shape produces, at another scale, a mirror image.

---

V. The Mirror

In the second half of 2024, a mid-tier data center operator in the Frankfurt region of Germany planned to expand a new 45-megawatt campus. They ran into an unsolvable problem—the local grid simply had no capacity. German grid upgrade cycles were estimated at four to five years.

They were not without money. But they did not have Microsoft's scale—no $1.6 billion to restart a nuclear plant. They did not have AWS's customer relationships—no way to sign for the entire output of an existing reactor.

But they could not cancel either. Customer contracts were signed. Default costs would far exceed delay costs.

They made a third choice.

In Q3 2025, they signed a partnership agreement with the German energy company E.ON to build a 61-megawatt on-site natural gas power plant. Not dependent on the public grid. Not in the interconnection queue. They did the same kind of thing the hyperscalers were doing—bypass the grid, build their own power, sign long-term contracts.

But when they did it, they had to apply for a generation license. Their regulatory identity changed—from data center operator to small independent power producer. Their regulatory category changed. Their engineers needed retraining. Their capital structure had to be re-evaluated under the IPP model—higher debt, longer contracts, slower return curves.

A similar thing happened in Ireland. Vantage Data Centers' DUB11 project in Clondalkin, Dublin, was forced to deploy temporary HVO and diesel generators for three years—from 2022 to 2025—because the Irish grid could not provide a market connection. What was designed as a 10-year data center lease became three years of an emergency-fuel supply chain.

Place these two cases side by side, and one thread emerges.

When public-grid bottlenecks become deep enough and long enough, second-tier developers are forced to do what hyperscalers do—bypass the grid, build their own power, sign long-term contracts. But because they lack scale, the same actions produce entirely different outcomes.

AWS restarting a nuclear plant locks in 1.92 gigawatts of base load, a 40-year contract, and pricing power. CyrusOne building 61 megawatts of on-site natural gas becomes a forced small-IPP, three-to-four-year delay, more expensive debt, no pricing power.

The same action. At different scales. Means entirely different fates.

---

A deeper layer.

This kind of "mutation" is currently a small-sample phenomenon. But if European grid upgrades genuinely require five-plus years, every second-tier project that cannot wait either dies—like the Southwestern developer in Section III—or walks this path.

By that point, the very job category of "data center operator" is being disassembled by physics.

When they enter operation three years late, their compliance status is no longer that of a data center operator. They are a small independent power producer—an industry they had never considered entering. Their engineers, who previously studied refrigeration and airflow, must now study turbine maintenance. Their legal team, which previously studied tenant contracts, must now study power purchase agreement clauses. Their financial model, whose depreciation cycle was 10 years, has been stretched to 25.

The company name didn't change. The building didn't change. The servers in the racks didn't change. But it is no longer the company it was.

---

VI

Return to the 3M Novec story.

On December 31, 2025, the last batch of fluorinated liquid left the 3M factory. The production line closed.

That day was 1,107 days after the December 20, 2022 announcement. Three years and 11 days. The industry watched the countdown tick away one square at a time, but most people didn't see it—until the last month, when they finally realized what they were standing on.

Transformers, helium, interconnection protocols, local councils, chemical phaseouts—each of these "hidden chokepoints" is not a new phenomenon. They have happened before. The fiber bubble, the wind queue, the nuclear renaissance, wartime allocation—four precedents, four endings: bankruptcy liquidation, process congestion, overrun hard landing, administrative reallocation.

These four endings are now playing out, in parallel, across the past two years of AI infrastructure expansion.

Second-tier developers queue for transformers and are then sold to funds at fifty cents. Hyperscalers buy nuclear plants and lock in 40-year base loads. Mid-tier European operators are forced into small power-producer status, their compliance category changing in ways they never anticipated.

Three fates. The same set of physical constraints.

---

The Southwestern developer that Xiongjian tracked was sold to Brookfield at fifty cents in Q2 2026. Brookfield's acquisition statement called it a "value discovery" transaction.

Value was indeed discovered—just the other side of value.

The 61-megawatt natural gas units are still running. The electricity they produce flows mostly into server racks. But the company that owns them is no longer the company it once intended to be.

---

The technology revolution moves at its own speed. Capital markets move at the speed of capital markets. Narrative—media, analysts, policy debate—moves at the speed of narrative. But transformers take 130 weeks because they take 130 weeks. Helium plants take three to five years to repair because that's how long they take. Nuclear plants take more than a decade from decision to first power. None of these timelines shorten because "AI is a historic opportunity."

Capital can accelerate. Narrative can accelerate. Physics will not accelerate with them.

When the three speeds cannot align, the loss is distributed—not evenly, but by scale. Those who can outlast the patience of physics inherit the future shadow grid. Those who cannot are sold at a discount to the liquidation table. Those in between are mutated into something other than themselves.

Physics has its own pace. It does not care whose narrative is more compelling.

Whose Sovereignty?

 

One afternoon in April 2026, I was staring at a federal spending spreadsheet, and something didn't add up.

U.S. government AI grants had surged 189% year-over-year. Over the same period, AI contracts had plummeted 76.8%. Pentagon AI contracts were even more dramatic — from $138 million to $9.4 million, a drop of over 93%.

Grants surging. Contracts collapsing. Two lines racing in opposite directions off the chart.

I thought I was tracking a data anomaly. I ended up tracking the fracture of the entire "sovereign AI" concept.

Oil and Water

Federal AI spending walks on two legs: contracts (the government buying services directly from companies) and grants (the government funding research institutions, universities, and state agencies to build capabilities). For years, every sovereign AI narrative rested on an implicit assumption — that government money would ultimately flow to AI platform companies.

But when I split the two legs apart, the picture was completely different.

On the contract side, federal AI contracts fell from $149 million to $35 million; the Pentagon's share dropped from $138 million to $9.4 million. On the grants side, federal AI grants soared from $440 million to $1.3 billion — with HHS AI-related grants up 796.5%. But line-by-line tracing revealed the single largest item was a rural health transformation program whose core wasn't AI procurement. After removing the false positive, HHS's genuinely AI-related grants were around $800 million, flowing mainly to universities, state health departments, and federal labs.

The leg that platform companies could directly monetize — contracts — was shrinking fast. The leg that was growing — grants — wasn't turning into revenue for any commercial AI platform.

The Refusal

Why did contracts collapse? Partly technical — reporting delays, keyword search blind spots, and the Pentagon's systematic shift to OTA (Other Transaction Authority) channels exempt from standard procurement rules. But the real story was on the second layer.

In July 2025, the Pentagon's CDAO awarded contracts of up to $200 million each to four frontier AI companies via OTA: Anthropic, OpenAI, Google, and xAI. In January 2026, Defense Secretary Pete Hegseth issued a memo requiring all AI contracts to include an "any lawful purpose" clause — meaning the military could use your model for anything legally permitted, and you couldn't impose usage restrictions.

The four companies' responses formed a precise spectrum. xAI fully complied, launching Grok For Government at $0.42 per seat. OpenAI compromised with wordplay, adding a "deliberately" qualifier. Google was negotiating to deploy Gemini into the Pentagon's classified networks.

Anthropic refused. Not stalled, not negotiated — refused. Two red lines: no domestic mass surveillance, no fully autonomous weapons without human oversight.

What happened next was dramatic enough to make Hollywood writers blush.

On March 4, 2026, the Defense Department formally designated Anthropic a "supply chain security risk" — the first such public designation of a mainstream American AI company. On February 27, the GSA had already removed Anthropic from USAi.gov and standardized procurement platforms per presidential directive.

But Anthropic's roughly $30 billion annualized revenue came overwhelmingly from commercial clients. The $200 million contract ceiling represented about 0.7% of annual revenue. The military blacklisted it. It barely noticed.

Anthropic promptly sued the Defense Department. On March 26, a California federal judge temporarily blocked the blacklist, ruling the designation "appears to be more about retaliation than actual national security risk." GSA restored Anthropic's position on April 2. But on April 9, the D.C. Circuit Court of Appeals refused to stay the blacklist's enforcement.

Two federal courts. Same case. Opposite rulings.

Even the American judicial system couldn't agree on whether an AI company was a security risk or a victim of retaliation — let alone what "sovereignty" could possibly point to in this context.

The Free Gatekeeper

While being blacklisted, Anthropic announced that its latest model — Claude Mythos Preview — possessed code vulnerability discovery capabilities exceeding most top human experts. It found a zero-day vulnerability lurking for 27 years in OpenBSD and critical flaws in mainstream video software that automated tools had missed across 5 million scans.

Anthropic didn't release the model publicly, nor hand it to the Pentagon that had just blacklisted it. Instead, it launched Project Glasswing, building a strict whitelist. Roughly 40 to 50 organizations received early access — including AWS, Apple, Google, Microsoft, CrowdStrike, JPMorgan Chase, and the Linux Foundation.

The pricing was even more unusual: Anthropic wasn't charging these companies — it was subsidizing them with $100 million in usage credits, plus donating $4 million to open-source security organizations. Whitelisted companies got three months to patch critical vulnerabilities. Everyone else — including most federal agencies during the blacklist weeks, and nearly all foreign governments — faced a clear defensive disadvantage against future AI-scaled attacks.

At the April 2026 IMF and World Bank spring meetings, AI-driven cybersecurity risk became a central topic. IMF Managing Director Georgieva said the global monetary system wasn't ready for AI cyber risk. Bank of England Governor Bailey called Mythos a severe challenge. Barclays' CEO warned it was a "serious threat." But many of these guardians of global financial stability couldn't get access.

The result was an awkward tableau: the White House facing the Pentagon's blacklist on one side while exploring pathways for regulators to access Mythos on the other. A company designated a "supply chain security risk" by the Pentagon was simultaneously viewed by the White House as indispensable to protecting national financial security.

Three Premiums

What exactly is the "sovereign AI premium" pricing? It's not one thing. It's three completely different things hiding behind the same label.

Contract lock-in. Represented by Palantir. 2025 revenue of roughly $4.475 billion, 54% from government clients. The logic: predictable government cash flows, high renewal rates, deep system dependency. Reasonable valuation multiples of 2–5x revenue. This is defense IT logic, not AI logic.

Commercial growth. Represented by Anthropic's core business. Roughly $30 billion in annualized revenue, overwhelmingly from enterprise API and commercial clients. Enterprise customers spending over $1 million annually doubled from 500 to over 1,000 in under two months. Reasonable valuation of 10–16x revenue, supporting a $300–480 billion valuation.

Option value. Represented by Mythos and Glasswing — the option value of a revenue category that doesn't yet exist. Glasswing is currently free. Anthropic is spending $100 million subsidizing it. Type 3 current revenue is zero. But secondary market implied valuations may have already reached the $700–850 billion range. If pure commercial revenue supports $450–480 billion, the remainder is option premium — the market betting that Glasswing will eventually transform from free strategic investment into paid security assessment services, perhaps even a quasi-license rent akin to credit rating agencies.

Whether this option pays off depends on one variable: will global financial regulators write "frontier AI security assessment" into compliance frameworks? I checked every relevant regulatory development over the past 30 days. No country or supranational body has issued a request for comment requiring third-party AI security assessments. My estimated probability of institutionalization: under 30%. The market's implied pricing: roughly 40–60%. That's a 10-to-30-percentage-point expectation gap.

The Attention Lesson

There's a dimension to this story about me.

Before making this video, my impression was that defense contracts were the main character and grants were a supporting role. That impression was wrong — not because my judgment was flawed, but because my information channels naturally skew toward high-drama narratives. The Pentagon blacklisting Anthropic was everywhere. Quiet grant data doesn't show up on its own.

The headline-grabbing Pentagon AI contracts? $9.4 million in USASpending. The almost-never-reported HHS AI grants? $1.1 billion. Narrative heat and funding reality were completely inverted.

You think you're tracking reality. You're actually tracking the slice of reality the media chose to report.

The Next Anchor

If Anthropic IPOs in Q4 2026 as rumored, the S-1 filing will answer the question no analyst can currently resolve: how does Anthropic itself view Glasswing?

If Glasswing is listed as a pure cost center — the option narrative lacks an internal anchor, and hundreds of billions in option premium will rest on something the company itself doesn't consider a revenue source. If an independent "security assessment revenue" line item appears — even a small one — that's an entirely different story.

Palantir sells lock-in. Anthropic sells optionality. The government buys capability. All three use the same words — "sovereign AI" — but point to completely different cash flow structures, risk factors, and durations. Conflating them is the easiest analytical mistake to make right now. Not because any single one is overvalued or undervalued — but because wrong classification leads to wrong-direction decisions at wrong times.

Some words, after being used by too many people, stop pointing to anything concrete. "Sovereignty" may be becoming one of them.

When a nation can't buy the security tools it needs most, when a private company gives its most powerful weapon to 40 corporations for free, when central bank governors sit around discussing a model they have no authority to test, when two federal courts hand down opposite rulings on the same company —

The weight of "sovereignty" has quietly shifted.

It no longer belongs solely to the building that signs procurement contracts.

It belongs to whoever runs tens of thousands of GPUs and decides who sees risk and who stays in the dark.

At least for today.

Tomorrow depends on an S-1 that hasn't been written, a regulatory framework that hasn't been published, and a whitelist that hasn't yet become an invoice.

The story is far from over. But the pricing has already begun.