The Security Pitch Became the Financing Plan

In Paris in June 2025, Jensen Huang stood before a room of government ministers and taught them a phrase: sovereign AI. The idea is that a country should run its own artificial intelligence the way it runs its own power grid or its own army, rather than renting it from a foreign company. Huang was selling, and he was explicit about what he was selling.

AI is now an essential form of national infrastructure – just like energy, telecommunications and the internet. — Jensen Huang

The pitch worked; over the following year a dozen governments signed up to buy their own compute. Fifteen months later, the same man made the same argument to a very different room: not ministers, but Wall Street. Compute, he told investors, is an asset you hold and rent out, not a machine you write down. It doesn't depreciate like hardware; it earns. He had a number to back the claim. On one rental exchange, three-year-old H100 chips were renting for 22% more than they had been a month earlier ^[1]. The security argument had crossed audiences, from the people who buy nations to the people who fund companies. That crossing is the whole story. The national-security frame around AI has quietly changed jobs. For the buildout's first years it was the sales pitch: buy this because your rival might. Now it is the financing plan: fund this because it is a matter of state. The clearest proof is the order in which the industry has gone looking for money. The repricing landed in the middle of 2026. S&P cut Oracle to BBB-, the last rung of investment grade, the rating that still counts as safe to lend to ^[2]. Taiwan's TSMC saw its bonds yield less than Meta's, and Tencent's traded on par with Apple's; the market now prices some American tech giants as riskier than their Asian rivals ^[3]. The two flagship data-center REITs, funds that own the server halls and pay investors from their rents, are down roughly 16% from their debuts ^[4]. OpenAI postponed its IPO to keep the trillion-dollar valuation intact, a delay that helped set off a global sell-off that week ^[5][6]. Bank of America counted $121 billion in new hyperscaler debt in 2025, more than four times the five-year average ^[7]. Bain estimates the industry needs $4.2 trillion in new revenue within five years just to support what it has already promised to build ^[8]. The buildout's answer was not to build less. It was to redirect the ask. The re-labeling had already been tried. In October 2025, OpenAI pitched federal support for its $500 billion Stargate project as the way to close an "electron gap" with China, a coming shortfall in electricity framed as a matter of national security ^[9]. A month later, Washington said no. The White House's AI czar declared there would be no federal bailout for the industry, and Sam Altman renounced the idea himself ^[10][11].

We believe that governments should not pick winners or losers, and that taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market. — Sam Altman

Weeks later came the relabel. Altman asked Congress to extend the Chips Act's manufacturing tax credit, written for semiconductor factories, to cover AI servers, data centers and the grid that feeds them, and he drew the distinction himself ^[12].

super different than loan guarantees to OpenAI — Sam Altman

By September 2026 the ask was back in security dress, and bigger. Altman proposed the government act as insurer of last resort for AI infrastructure. His chief financial officer, Sarah Friar, called for a government backstop to ease financing. Analysts reached for 2008, saying the industry was positioning itself as too big to fail ^[13]. Some doors cracked open anyway. Washington has explored taking equity stakes in the big AI firms, on the model of its existing stakes in Intel and Westinghouse, and the labs themselves have raised the prospect of handing the state shares ^[10]. The government announced two gigawatt-scale AI power projects on federal land before any final agreement was signed ^[14]. Nvidia is putting its own balance sheet behind its partners, including a reported $250 billion in lease support for OpenAI ^[15]. And sovereign money is already inside: Abu Dhabi's MGX fund co-anchors Stargate ^[16] and sat in the share sale that valued OpenAI at $500 billion ^[17], a company whose $638 billion contract backlog rests on counterparties borrowing against those very commitments ^[18]. China, for contrast, funds its own $560 billion data-center buildout directly from the state ^[19]. That is which way the asks lean. But the refusers below have been just as loud. The companies that signed last March's ratepayer-protection pledge are now disputing what it covers; they intend to pay only for infrastructure serving their own sites, not shared grid upgrades, and Google told federal energy regulators that forcing data centers to bring their own power would scare off investment ^[20]. Household bills are already rising in 13 states where the centers have arrived ^[9]. An energy lobbyist put the pushback most sharply.

Technology that allows you to be more efficient with what you have doesn’t do well in an incentive environment that’s tilted towards building new things. — Gary Rahl

The slowest and freshest proof is abroad, where the security story found its most eager buyers, and where the fear is real. Last spring a U.S. export-control directive switched Anthropic's models off for foreign nationals, abruptly cutting off India's Tata Consultancy Services and Infosys mid-project ^[21]. India answered by launching a sovereign program of its own. The head of Britain's Alan Turing Institute made the same argument, more precisely.

As AI becomes more embedded in critical services, in domains like defence, national security, critical infrastructure and healthcare we shouldn’t depend solely on systems we can’t inspect or control, or that could be restricted or even switched off. — George Williamson

A dozen nations are now racing to buy their own models, toward a market projected at $1.2 trillion ^[22]. OpenAI has built a product line around the fear: an "OpenAI for Countries" program and a sovereignty-framed gigawatt-scale data center in India ^[23]. That last project sits awkwardly against the company's own pause of its British data center over energy costs, the one once billed as strengthening the UK's sovereign compute ^[24]. No one was lying; those governments really were switched off. The record shows what the fear was asked to buy. One complication, because the security designation cuts both ways. After Anthropic refused the Pentagon unrestricted access to its models, the defense secretary named it a supply-chain risk and the White House ordered agencies off its Claude system ^[25]. The state can wield security as a weapon against a lab, not just as a shield for it. And the money is still flowing: BlackRock, Goldman Sachs and Apollo are mobilizing $500 billion behind the reclassification ^[1], and Nvidia's revenue more than doubled from a year earlier ^[26]. Which is what defines these asks: they are about terms and who bears the risk, not about survival. What has actually been paid: a Gulf sovereign's money inside both the flagship project and the valuation, and a dozen foreign governments buying compute out of a fear that is not feigned. What has been refused: the direct backstop has been refused twice, weeks apart, a year ago — and the newest ask is still waiting for its answer. What remains exploratory or unsigned: the equity stakes, the federal-land gigawatts announced before any agreement, the chipmaker's own balance sheet behind the labs. The next observable event is simple enough. Which of the unsigned things gets signed first.