Who Really Controls Frontier AI?

The shutdown of Claude Fable 5 was reported as a national-security story: the US government switched off the most capable public AI on earth, three days after launch, citing concerns it would not fully explain. That is the surface. Underneath it sits a tangle of money, rivalry and state power that is worth mapping in its own right — because the same configuration of forces will shape the next shutdown, and the one after that.

This piece is not a verdict on whether Washington was right; the companion piece takes up that question. This one steps back and maps the board: who rents compute from whom, which incentives pull in which direction, the rivalries stacked on top, who actually made the call, where China fits, and the scenarios that follow. Some of what follows is firmly documented; some is contested or speculative, and it is flagged as such. The aim is to see the whole shape, not to settle it.

The compute map: who rents from whom

Start with the physical layer, because almost nothing in the coverage did. The frontier labs do not, for the most part, own the data centres their models run on. They rent them — and increasingly, from each other.

In May 2026, Anthropic leased the entire Colossus 1 facility in Memphis from xAI — Elon Musk's AI company, now folded into SpaceX — for a reported $1.25 billion a month through to 2029: more than 300 megawatts of compute and around 220,000 GPUs, rented from a direct competitor. Google, which builds the rival Gemini models, agreed to pay SpaceX about $920 million a month for AI capacity from October. Behind Anthropic sits Amazon — its largest investor and primary training partner, with commitments in the tens of billions.

So the map already looks less like a race between independent labs and more like a set of overlapping landlords and tenants: Musk rents to the labs he competes with; Amazon and Google both fund Anthropic and field rival models; everyone depends on someone who is also, in some direction, a competitor.

AI-generated

The push and the pull

The interesting thing about this structure is that the incentives do not point cleanly in one direction. They pull both ways at once — which is exactly why the shutdown is hard to read as a simple story.

Pull one way: a lease like Anthropic's is a fixed commitment. The roughly $1.25 billion a month is owed whether Fable 5 is serving millions of users or sitting dark under a government order. So when the model went offline, Musk's revenue did not move, while one of his sharpest competitors lost its flagship product. On that reading, a frozen Anthropic is, if anything, mildly good for the landlord — paid either way, with a rival hobbled for free.

Pull the other way: a thriving Anthropic is also good for the landlord. More usage means more compute demand, more capacity leased, more rent — and a healthy tenant that renews. Musk has an interest in Anthropic succeeding (as a customer) and an interest in it stumbling (as a competitor) at the same time. Neither pull proves intent behind the shutdown. But it is worth naming that the company collecting the rent benefits in more than one scenario, which is unusual, and which makes "who gains here" a genuinely open question rather than a rhetorical one.

The rivalries stacked on top

On top of the money sits a set of personal and corporate rivalries that colour how each move is read.

OpenAI's Sam Altman and Elon Musk have feuded publicly for years, and Altman had already mocked Anthropic's safety-first posture as fear-based marketing — "we have built a bomb… we will sell you a bomb shelter." Anthropic and OpenAI are the two front-runners in the enterprise market; xAI's Grok competes with both; Google competes with all of them while funding one. The result is that no major lab can make a move — a warning, a launch, a lobbying push — without it being read through the lens of who it helps and who it hurts. The Fable 5 shutdown landed in that environment, where every actor has a reason to spin it.

Who actually made the call

"The government switched it off" hides a chain of specific people and choices worth tracing.

The order came from the Commerce Department — Secretary Howard Lutnick, in an export-control directive to Anthropic's CEO Dario Amodei. Export control is a deliberate instrument: the same machinery used for advanced chips and weapons, and it lands hardest on foreign access. It did not arrive in a vacuum. Months earlier the administration had designated Anthropic a "supply-chain risk" after the company refused to let its models be used for mass surveillance and autonomous lethal targeting — a dispute Anthropic took to federal court. And Musk, Anthropic's landlord and a competitor through xAI, has publicly reconciled with President Trump after their 2025 feud, even joining him on a trip to China.

None of that proves the decision was anything other than a clean-handed security call on the merits. But the board around it — the regulator who acted, a prior dispute with the same company, a landlord-competitor back in the President's favour — is the relevant context for anyone trying to understand how a call like this gets made, and who has standing to make it.

AI-generated

The China thread

There is a genuine national-security spine here, and the popular version of it is backwards, so it is worth getting right.

The rumour is that Anthropic is somehow built on Chinese technology — that the "foreign national" framing points to Chinese code inside Claude. It does not; Anthropic builds Claude itself. The real China story runs the other way. Anthropic has publicly alleged that Chinese labs — DeepSeek, Moonshot and MiniMax — used roughly 24,000 fake accounts to "distil" Claude, copying its capabilities out through the API. The fear in Washington is frontier capability leaking outward, to exactly those actors, which is why an export-control order aimed at foreigners was the chosen lever. Read that way, the shutdown is the US treating a commercial AI model the way it treats a sensitive dual-use technology — something to be kept on the American side of a line.

AI-generated

The theories, weighed

Several explanations are circulating. They are worth laying side by side, because the facts fit some better than others.

"Elon engineered a trap." The idea that Musk lured Anthropic into a costly lease and then helped pull the product. There is no evidence for it, and it is the weakest of the theories — the shutdown order came from Commerce, not from a landlord. What is true, and quieter, is that Musk's incentives are conflicted, which is enough to make the arrangement worth scrutinising without inventing a plot.

"It is the best marketing Anthropic could ask for." The claim that the whole episode burnishes Anthropic's safety brand. This does not hold up well either: the company lost its flagship product, its own foreign staff lost access, and reporting framed its safety warnings as having backfired. A genuine own-goal is hard to rebrand as a win.

"It is a move to slow Anthropic down." The suspicion that the action conveniently hobbles the front-runner while rivals are untouched. This one has more surface to it — only Anthropic's models were hit, even though Anthropic argues the capability in question exists in rival models too, and the company was already in a dispute with the administration. It does not require a conspiracy; selective enforcement can do the same work as one.

"It is a genuine security response." The straightforward reading: a model the company itself had warned about shipped publicly, a jailbreak claim surfaced, and the state moved fast. It is defensible on its own terms — the weakness is not the instinct but the absence of any visible standard, applied to one company.

The honest position is that more than one of these can be partly true at once, and the public record does not yet let anyone close the case.

The alternatives from here

Where the board moves next is the part still being written. A few branches are visible.

Access could be quietly restored — Anthropic has said it believes the order rests on a misunderstanding and is working to reverse it, which would make this a sharp but temporary shock. The next frontier model could ship with less fanfare — an Opus-class release framed as routine rather than as a dangerous capability, precisely to avoid drawing the same scrutiny. The episode could push the industry toward defined rules — Anthropic has called for FAA-style certification of frontier models, with published risk categories and thresholds, in place of discretion exercised by phone call. Or the precedent could widen, with similar orders reaching other labs once the mechanism has been used once. Which branch the board takes will say more about who really holds the controls than any single shutdown does.

The map, not the verdict

Step back and the shape is clear even where the motives are not. The most advanced AI on the planet runs in a rival's building, on a rival's terms, paid for whether it is switched on or off, funded by competitors, and subject to a government that can change its mind in an afternoon. No lab in this picture fully owns the frontier it is supposedly conquering: there is a landlord, an investor, and a government inside every model.

That is the map. It does not tell you who was right about Fable 5, and it is not meant to. It tells you that the next time a frontier model goes dark — or doesn't — the forces deciding it will be these: the compute leases, the cross-investments, the rivalries, and the handful of people in Washington with the authority to act. Worth watching, whichever way the board moves next.