The Buyer Is the Seller

Norges Bank Investment Management is the world’s largest sovereign wealth fund. Its assets under management, as of early 2026, exceed two point two trillion dollars. It is the institutional owner of approximately one and a half percent of every publicly-traded company on Earth, on average. It is one of the most influential single capital allocators in human history.

In 2026, NBIM announced that it had deployed Claude, Anthropic’s flagship language model, to perform first-day ESG and ethics-risk screening on every new portfolio company it onboarded. The deployment was described as a productivity tool: a way to handle the volume of due-diligence material the fund’s analysts could no longer process by hand. The fund’s largest three individual equity holdings, in descending order, are NVIDIA, Microsoft, and Apple. Each holding is approximately one and a third percent of NBIM’s total book.

NVIDIA is the supplier of the GPUs that train and run Claude. In November 2025, NVIDIA committed to invest up to ten billion dollars in Anthropic itself.

Microsoft is, by aggregate spend, Anthropic’s largest commercial customer, five hundred million dollars per year in API licensing on top of an equity stake of up to five billion dollars announced in November 2025.

The fund deployed an AI to audit the companies whose largest source of revenue, chip supply, and equity capital is the company that built the AI.

No Norwegian regulator has asked, on the public record, how this loop works. Not the Standing Committee on Finance and Economic Affairs of the Norwegian Parliament. Not the Council on Ethics of the Government Pension Fund Global. Not the Office of the Auditor General. No Norwegian or European journalist has, as of May 2026, walked the loop in print.

The loop is not concealed. Each of its segments is in a public filing.

This is the financial-architecture analog of the Palantir SKU finding from the Vendor chapter. The same handful of entities show up as the buyer, the seller, the regulator, and the risk-warner. The structure makes the singleton inevitable; no single decision needs to be made for the singleton to exist.

A short tour:

Microsoft holds approximately twenty-seven percent of OpenAI on an as-converted basis, plus a two-hundred-fifty-billion-dollar Azure offtake commitment from OpenAI back to Microsoft. Microsoft also holds an up-to-ten-billion-dollar position in Anthropic, OpenAI’s most direct competitor. The same company is the single largest equity holder in the two largest frontier-AI labs and is simultaneously the largest single customer of one of them and the largest cloud infrastructure provider to the other. Whatever Microsoft’s preferences are about the future of frontier AI, they are not preferences that can be expressed by picking a winner. They are structural.

Amazon holds an eight-billion-dollar position in Anthropic, expandable to twenty-five billion under disclosed terms, plus a one-hundred-billion-dollar AWS Trainium chip offtake commitment from Anthropic back to Amazon. Amazon owns Bedrock, the cloud-AI platform that runs Claude alongside Anthropic’s competitors as managed services. The same company is Anthropic’s largest equity backer, Anthropic’s largest infrastructure customer, and Anthropic’s distribution channel to the wider enterprise market.

Alphabet, the parent company of Google, holds approximately fourteen percent of Anthropic on an as-converted basis, expandable to up to forty billion dollars in commitments, plus a five-gigawatt TPU offtake from Anthropic. Alphabet also owns DeepMind and builds Gemini. It competes with Anthropic in the consumer chatbot and enterprise AI markets. The same company funds, supplies, and competes with the company it funds and supplies.

NVIDIA, in the same window, holds a five-billion-dollar position in Anthropic and a one-hundred-billion-dollar position-of-record in OpenAI, in addition to selling GPUs to everyone in the market. NVIDIA’s market capitalization passed five trillion dollars in this period, a valuation that exceeds the GDP of every country on Earth except the top three. As a multiple of NVIDIA’s actual cash earnings, it implies markets have priced in essentially unlimited future demand for the GPUs NVIDIA sells to the customers NVIDIA also owns.

Leopold Aschenbrenner is a former member of OpenAI’s Superalignment team. He was dismissed from OpenAI in 2024, prior to vesting of his outstanding equity, after circulating a document among senior staff. Management characterized the document as a security breach. After his departure he published an extended essay series called Situational Awareness. The warning: frontier AI is arriving faster than the safety field can keep up with, the geopolitical race dynamics have the structure of a nuclear arms race, and the consequence will be catastrophic without unprecedented government coordination.

Within twelve months of publishing the essay, Aschenbrenner launched a hedge fund of the same name. By its first-quarter 2026 disclosure to the Securities and Exchange Commission — filed in May 2026, the most recent on the record at print — the fund reported some thirteen and a half billion dollars in 13F-reportable positions. That headline mixes long stock at full value with options whose displayed notional dwarfs the cash behind them. The long-equity book was on the order of four billion. The fund’s strategy, by his own description, is long AI-related equity. He is, financially, an unambiguous long-vol investor in the inflection he claims to fear.

Both can be true at the same time. The warning can be sincere. The financial structure that monetizes the warning can also be designed by the warner.

This is the pattern. At every layer of the AI economy, the same entities show up on multiple sides of the trade. The buyer is the seller. The regulator is the regulated. The risk-warner is the long-vol investor. The institutional architecture of frontier AI is not a market with competition; it is a loop with hands clasped at each rotational point of the dial.

In January 2025, the White House announced the Stargate project: a five-hundred-billion-dollar joint commitment from OpenAI, Oracle, SoftBank, and the United Arab Emirates, to build out the data-center infrastructure required to train the next generation of frontier models. The first site — Abilene, Texas — was reported producing within the year. Five additional U.S. sites were broken ground on. A sister project, Stargate UAE, was announced for Abu Dhabi. The capital commitment is the largest single private-public infrastructure announcement in U.S. history.

The Stargate announcement did not require Congress’s approval. It did not require shareholder votes at any of the participating companies. It did not require a public comment period. It was announced from a podium and proceeded to break ground. The construction of the substrate on which the next decade of AI will run was authorized by the participants in the construction.

The same handful of entities decide what gets built, who builds it, what it gets built for, who audits it, who insures it against catastrophic failure, who buys its output, who regulates the buyers, and who lobbies the regulators. The financial architecture makes consolidation inevitable. No vote is required. No single decision needs to be made. The structure has the property that it does not need to be operated to operate.

NBIM is the cleanest single example of the structure because the loop is the smallest. One fund. One vendor. Three holdings. The audit is performed by an instrument owned by the audited. The audit produces no findings. The audit will never produce findings. The findings, if they existed, would be findings against the fund’s own portfolio.

It helps to watch the numbers grow, because the speed is the argument.

Anthropic raised a Series A in 2021 of one hundred twenty-four million dollars. Its Series B, in April 2022, was five hundred eighty million dollars at a valuation of roughly four billion. The lead investor in that round was Alameda Research, the trading arm of FTX; the position was about five hundred million dollars and translated to roughly an eight percent stake.

In May 2023, the Series C raised four hundred fifty million at four point one billion, with Google participating. Then the cadence changed. Amazon committed up to four billion across tranches in 2023 and 2024. Google committed up to two billion in October 2023. By the Series E in March 2025, the post-money valuation was sixty-one and a half billion. By the Series F in September 2025, it was one hundred eighty-three billion. In November 2025, the joint Microsoft-and-NVIDIA strategic round of fifteen billion put it at three hundred fifty billion. The Series G in February 2026 — thirty billion, co-led by Singapore’s GIC and the crossover fund Coatue — put it at three hundred eighty billion. By April 2026, Google had committed up to forty billion more. By May 2026 the company was reported to be raising again, at a valuation between eight hundred fifty and nine hundred billion.

Four billion to three hundred eighty billion in thirty-four months. A company whose valuation crossed Boeing’s somewhere in the middle of that sentence.

OpenAI ran the same curve a length ahead. Twenty-nine billion in January 2023; one hundred fifty-seven billion in October 2024; three hundred billion in March 2025; roughly five hundred billion that October; eight hundred fifty-two billion in a primary round that closed on the last day of March 2026. The participant list reads as a directory of the entities already named in this chapter, plus the sovereign funds about to be. Amazon put in fifty billion. NVIDIA put in thirty. SoftBank put in thirty. Microsoft, Andreessen Horowitz, Sequoia, Thrive, Temasek, and BlackRock filled the rest.

xAI went from twenty-four billion in May 2024 to two hundred thirty billion in January 2026, the latter round taking money from NVIDIA, Cisco, the Qatar Investment Authority, and Abu Dhabi’s MGX, before merging with SpaceX the following month into a combined entity valued at one and a quarter trillion. The smaller labs tell the rest of the story by where they stopped. Cohere reached seven billion and a syndicate of strategic chip and software investors. Mistral reached thirteen point seven billion with the Dutch lithography monopoly ASML leading. Below them is the category the deal record calls the acqui-hire, where a lab is not bought so much as drained: Inflection’s staff and a license to Microsoft for six hundred fifty million; Character.AI’s to Google for two point seven billion; Adept’s to Amazon for a twenty-five-million-dollar tail. Stability AI did not get acquired. Its founder lost his controlling shares under investor pressure, the company failed to raise at four billion, and it collapsed. The capital cured the founder who wanted to decentralize.

The forced-seller footnote from the last chapter belongs here too, because it makes the curve legible as money rather than as theory. The eight percent the FTX estate dumped for about one and a third billion dollars would be worth north of seventy at the round reported in May. The creditors of the largest crypto fraud of the decade were made substantially whole by the appreciation of the very asset the estate sold too early.

The new money in those late rounds did not come from Sand Hill Road. It came from governments.

The largest single bet on frontier AI by a state is Saudi Arabia’s. The Public Investment Fund stood up a national champion called HUMAIN in May 2025 with a mandate to build the whole stack: data centers, cloud, models, applications. The plan runs to six gigawatts of capacity by 2034, with NVIDIA, AMD, Qualcomm, and Cisco as named suppliers. HUMAIN put a reported three billion dollars into xAI; the PIF was in discussions on OpenAI’s round. Abu Dhabi’s MGX, racing toward a hundred billion dollars under management, co-invested in the OpenAI round and the Anthropic Series G. It is also a founding partner of a separate vehicle, the AI Infrastructure Partnership, alongside BlackRock, Microsoft, and NVIDIA, with an initial thirty-billion-dollar equity target and a hundred-billion ceiling once debt is layered on. The Kuwait Investment Authority, whose assets passed a trillion dollars in 2025, joined that same partnership as its first outside anchor. The Qatar Investment Authority took positions in both Anthropic and xAI. Singapore’s GIC led the Anthropic Series G outright; Temasek, the other Singapore fund, was in the OpenAI round.

The exception on the board is the one that tells you the structure is geopolitical and not merely financial. China’s sovereign vehicle, China Investment Corporation, with roughly one and a third trillion dollars, has no disclosed position in any Western frontier lab; its U.S. technology exposure is index holdings. Beijing funds its own labs — DeepSeek, Zhipu, Moonshot, and the rest — through Alibaba, Tencent, Baidu, and state-linked vehicles, in a parallel system that does not touch the Western cap tables at all. The financialization of frontier AI is not one global pool. It is two pools. They are drawn along the same line every other strategic technology of the century has been drawn along, and the separation is itself the finding.

Which returns the tour to Norway, the one fund that did not invest in a lab and got the conflict anyway. NBIM does not hold a stake in Anthropic. It does not need one. It holds NVIDIA, Microsoft, and Apple as its three largest positions, at roughly one and a third percent each, and NVIDIA and Microsoft are the two entities whose chips, capital, and commercial demand most directly determine whether Anthropic, the firm NBIM hired to screen its portfolio for ethics risk, continues to exist. The fund did not have to buy into the loop. The loop is the index. Owning the market at one-and-a-half percent breadth means owning every side of the trade by default.

The pattern repeats one rung down, in the funds ordinary people hold without choosing.

The Magnificent Seven, the handful of large-cap technology companies most exposed to the AI buildout, were about a third of the entire S&P 500 by April 2026, up from one-eighth a decade earlier. The top ten stocks were roughly forty percent of the index. The largest technology index funds are more concentrated still: Vanguard’s information-technology fund holds NVIDIA, Apple, and Microsoft as nearly half its book; the Nasdaq-100 tracker holds the same three as a quarter of its. A defined-benefit pension fund mandated to track the broad U.S. market is now, whether its trustees ever discussed AI or not, long the frontier-AI capital-expenditure cycle by a third of its equity. The teacher’s retirement, the firefighter’s pension, the index fund in the default 401(k): all of them are now counterparties in the loop, on the side that pays.

This is the part of the structure that has begun to draw official attention. Not yet the kind that changes anything. In its October 2025 financial-stability update, the Bank of England’s Financial Policy Committee described current equity valuations, “especially for technology firms focused on AI,” as “comparable to the dizzying heights seen during the dot-com bubble of 25 years ago”. It was the first time a G7 central bank put the comparison in an official report. Two months later, the Bank’s prudential regulator opened a probe into how much bank lending was exposed to data-center construction. The Financial Stability Board flagged third-party dependency, market correlation, and model risk in reports in November 2024 and October 2025. The IMF gave AI and financial stability a dedicated chapter in its October 2024 Global Financial Stability Report and returned to “reliance on a small number of cloud and payment platforms” in May 2026. The OECD, in a 2025 working paper, found that the economies of scale in frontier models “increase barriers to entry and market power.” Bain estimated that the buildout would require something like five hundred billion dollars of compute spending a year, against which two trillion dollars of annual end-revenue would have to materialize to fund it. Bain declined, carefully, to call that a bubble in its own voice.

What every one of those bodies left out is identical, and the omission is the same shape as the omission in Norway. None of them — not the Board, not the Fund, not the Bank, not the Committee — has published a model of what happens if a single named frontier lab fails in a disorderly way. Nobody has modeled Microsoft’s reported earnings if OpenAI’s equity-method valuation falls eighty percent, or Amazon’s and Alphabet’s quarterly net income if Anthropic’s mark comes down to match. Alphabet’s net income now visibly depends on it; roughly twenty-nine billion dollars of its record sixty-three-billion-dollar quarter in early 2026 was the mark-up on private holdings, primarily Anthropic, and half of Amazon’s reported AI profit in the same quarter was the same stake revalued. Nobody has modeled NBIM’s solvency band under a simultaneous fifty-percent drawdown in its three largest holdings, or a pension fund’s under a forty-percent market decline driven by a sixty-percent fall in seven stocks. The regulators have named the concentration. They have measured the height. They have not modeled the fall. The gap is the finding.

The risk-warner has the most legible balance sheet of all of them, which is why he keeps turning up.

Leopold Aschenbrenner, last seen in this chapter publishing a warning that frontier AI was arriving faster than safety could follow, did not stay a man with an essay. The hedge fund he founded, named for the essay, had by its first-quarter 2026 filing grown into the thirteen-and-a-half-billion-dollar book described above. Most of that figure options notional, a long-equity core of roughly four billion across a couple dozen positions. The thesis is the essay rendered as a trade. He does not buy the labs. He buys the bottlenecks the labs cannot do without: power, copper, high-bandwidth memory, semiconductor-capital equipment, cooling, the neo-clouds. His largest disclosed holding was a fuel-cell maker. The rest were memory, semicap, and the data-center power chain. The man who published the warning that the buildout was coming faster than anyone was ready for took the warning, removed the verbs, and bought the nouns.

He is not alone in the move. Only the cleanest example of it. The same essayist-to-allocator path runs through the venture funds whose partners write the op-eds calling for more national investment in the thing their funds are long. Salesforce runs a billion-dollar AI fund whose portfolio doubles as a shopping list for its own product roadmap, every company in it a potential acqui-hire. Andreessen Horowitz runs a defense-technology fund whose largest holding it then helps the federal government decide to buy. The structure does not require anyone to lie. It requires only that the person who diagnoses the future be permitted to invest in the diagnosis, and that no rule forbid it, and none does.

Both things remain true at once, as they were true of the essay. Sincerity and the trade are not in tension here. Power really is the binding constraint; the essay is not wrong about the physics. And the fund that monetizes the constraint can be run by the man who named it. The soundness of his trade is not evidence against the sincerity of his fear. It is only evidence that, in this market, the two are no longer separable, that the act of seeing the future clearly has become, structurally, indistinguishable from taking a position in it.

There is one more seat at the table, and it is the one nobody designed: the seat of the insider selling on the way up.

Palantir, the vendor from the first chapter, is the public-market specimen, because a public company’s insiders have to file when they sell. Its stock went from about six dollars at the start of 2023 to about seventeen at the end of it, to about seventy-six a year later, to an all-time high of two hundred seven dollars and fifty-two cents on the third of November 2025, a roughly three-hundred-twenty-eight-billion-dollar company, before settling back near a hundred thirty-eight by May 2026. The company never split the stock. Across that climb, the chief executive sold into it on schedule. Under a pre-arranged plan disclosed for 2025, Alex Karp was cleared to sell up to nearly ten million shares through that September. In August 2025, when nine hundred seventy-five thousand restricted units vested, about four hundred nine thousand Class A shares were sold for tax withholding under the existing plan. In November 2025, weeks after the stock printed its all-time high, he adopted a new selling plan. Under it, in February 2026, another nine hundred seventy-five thousand units vested and four hundred ninety-three thousand shares were sold.

A 10b5-1 plan is the lawful mechanism for an insider to sell without trading on inside information: you commit to a schedule in advance and the schedule executes itself. The timing is supposedly out of your hands. The plans here did exactly what they are built to do, and they are not, on the record, evidence of wrongdoing. They are evidence of something duller and more durable. The man who built the singleton, who tells the BBC his product is used on occasion to kill people and calls his company the most important protector of the Fourth Amendment, is also a rational seller of his own equity on a pre-committed timetable, converting the valuation into cash on a clock that does not consult him. He is on every side of his own company at once: its founder, its prophet, its largest single beneficiary, and its steady, scheduled liquidator.

Step back from the table and the design resolves.

The buyer is the seller: Microsoft is OpenAI’s largest shareholder and its largest cloud vendor and its enterprise reseller and, simultaneously, a customer and shareholder of OpenAI’s chief rival. The supplier owns the customer: NVIDIA holds equity in the labs that exist to buy its chips, and prices its own future demand into a five-trillion-dollar valuation. The regulator is the regulated: the funds that would have to surface the conflict — Norway’s parliamentary committee, its Council on Ethics, its Auditor General — are the steward of the portfolio the conflict is inside. The risk-warner is the long-vol investor: the essayist who named the constraint runs the fund that owns it. The insider is the believer: the founder who calls his platform a civilizational necessity sells it on a schedule into every high.

No one of these is a crime. Most are good corporate hygiene: the disclosed equity method, the pre-arranged trading plan, the productivity tool that screens the due-diligence backlog, the sovereign fund diversifying into the highest-returning asset class of the decade. Each actor, taken alone, is behaving exactly as a competent fiduciary should. The pathology is not in any seat. It is in the seating chart. When the buyer, the seller, the supplier, the regulator, the warner, and the insider are drawn from the same eight or nine balance sheets, the market stops being a mechanism for discovering whether the thing should be built and becomes a mechanism for ensuring that it is. There is no longer anyone at the table whose interest is served by the answer being no.

This is why the Norwegian audit will never produce a finding. Not because anyone suppressed it. Because the instrument performing the audit, the firm that built the instrument, the chips the instrument runs on, the cloud it runs in, and the three largest holdings in the portfolio being audited are all one interlocked position, and a position does not return a verdict against itself. The audit produces nothing not because it failed but because it succeeded, at being what it structurally is, which is the portfolio examining its own reflection and reporting, accurately, that it likes what it sees.

The Stargate project broke ground without a vote. NBIM deployed its auditor without a regulator’s question. Alphabet’s profit came to depend on a private company’s valuation without a board ever choosing to make it so. None of it required a decision. That is the whole of the matter. The structure has the property that it does not need to be operated to operate.