{
  "slug": "obchyslyuvalna-zavisa",
  "url": "https://neurodrift.org/en/blog/obchyslyuvalna-zavisa/",
  "title": "The Compute Curtain: the week AI became an instrument of the state",
  "description": "On June 12, a single US government letter switched off two Anthropic frontier models for every foreign national on the planet. Why AI became an instrument of the state, how the 'compute curtain' works, and who holds the switch to your access.",
  "author": "Дністер",
  "language": "en-US",
  "published": "2026-06-21T13:19:00.000Z",
  "updated": null,
  "tags": [
    "ai",
    "geopolitics",
    "export-controls",
    "anthropic",
    "sovereignty"
  ],
  "translationOf": "https://neurodrift.org/blog/obchyslyuvalna-zavisa/",
  "sourceUrl": "https://neurodrift.org/blog/obchyslyuvalna-zavisa/",
  "body": "## Friday, 5:21 p.m.\n\nThe letter arrived on a Friday — of course it did. Fridays are when governments do what they'd rather not explain before Monday.\n\nJune 12, 2026. 5:21 p.m. Washington time. The US government delivered a directive to Anthropic: suspend access to Fable 5 and Mythos 5 for any foreign national — inside the US, outside the US, including the company's own foreign-national employees. The trigger was a narrow jailbreak: someone asked the model to read a codebase and identify vulnerabilities. The government did not disclose the underlying statute. Axios and Bloomberg named Commerce Secretary Howard Lutnick as the signatory; Anthropic's own statement was more careful — \"the US government, citing national security authorities.\" No name. No cited provision of law.\n\nThe technical problem turned out to be both simple and unsolvable at once: filtering foreign nationals from Americans in real time is not possible. The system does not see a passport — it sees tokens. So within hours, Fable 5 and Mythos 5 went dark for everyone. Opus 4.8 stayed online. Hundreds of millions of people received silence where an answer used to be — not for anything they had done, but because the directive left no other scalable option.\n\nThis is the first documented instance of the United States applying an export control mechanism to a specific deployed frontier model on national security grounds. Not to a factory. Not to chips. To the API itself.\n\n## The Compute Curtain\n\nWorth stating plainly before it gets lost in the incident's details.\n\n<mark>AI stopped being a market and became an instrument of the state: the question is no longer \"will you have access,\" but \"who holds the switch to yours.\"</mark>\n\nThe Iron Curtain divided what you can see on a map: borders, wires, checkpoints. The Compute Curtain divides something that appears on no map — the right to speak with the most intelligent machine humanity has built. And it is drawn not over decades but in half a day, by a single letter that arrives on a Friday at 5:21 in the afternoon.\n\nThe mechanism is not new — the floor it reached is. States have long controlled who can buy hardware. Last year they moved to who can own the model. This week: who can touch one already running. Each step higher up the stack, closer to thought itself.\n\nWhat the incident exposed is surgically simple: the chokepoint of the era is not silicon, not code, and not even data. It is **access**. And access is now a political instrument. Anthropic noted immediately that GPT-5.5 from OpenAI carries the same capability — which is almost an irony: they did not patch the hole in the fence; they moved the fence. But it changes nothing about the architecture: the state demonstrated that it **can**. Next time it will be OpenAI too. Or selectively: allies yes, everyone else no.\n\nThe \"Compute Curtain\" is not an apocalyptic metaphor. It is the operational reality for anyone building a product on another jurisdiction's frontier model.\n\n## The Ladder: 2022 → 2025 → 2026\n\nThis is not government's first reflex. It is the terminal station of a railway that left in 2022.\n\n| Year | What came under control | What it meant |\n|---|---|---|\n| Oct 7, 2022 | Chips and semiconductor manufacturing equipment | who can **buy** the hardware |\n| Jan 15, 2025 | Closed model weights (AI Diffusion Framework, 10²⁶-ops threshold) | who can **own** the model |\n| Jun 12, 2026 | Access to a specific deployed model by name | who can **touch** the thinking machine |\n\nIn 2022, the US said: we will not sell you TSMC equipment and Nvidia H100s. Chokepoint — silicon. In 2025, the AI Diffusion Framework put closed model weights under BIS control for the first time — not hardware, but the knowledge inside it, above a threshold of 10²⁶ operations. \"For the first time, BIS will control AI model weights,\" Sidley recorded. But Diffusion still looked like regulation: country tiers, carve-outs, certifications.\n\nJune 2026 took a step that had no precedent: control over a **specific** model by name. \"The first known U.S. use of export control authorities to regulate a particular AI frontier model on a national security basis,\" noted NatLawReview. The chokepoint migrated from silicon to endpoint: the question is no longer \"do you have a GPU\" — it is \"does the US government permit your IP address to ask a question.\"\n\nThe train arrived on schedule. The question is not whether anyone saw it coming. The question is who had already bought a ticket.\n\n## Why one letter equals hundreds of millions\n\nCompliance lawyers point to the doctrine of *deemed exports* — and it is precisely this that explains why Anthropic had no middle option between \"defy the order\" and \"switch it off for everyone.\"\n\nThe logic came out of the crypto wars of the 1990s. The US government classified encryption algorithms as munitions and ruled: transferring controlled technology to a foreign national is an export, even if both parties are standing on American soil. This is the \"deemed\" export: what moves is not the product but access to it.\n\nTranslate that to an API in 2026. Every request from a non-American — even from within the United States — becomes an act of \"deemed\" transfer of controlled technology to a foreign person, i.e., an export. And filtering \"foreign nationals\" from \"Americans\" in real time is impossible: citizenship is not a field in an HTTP header. Anthropic acknowledged the deadlock directly: unable to separate one group from the other, the company disabled both models for **everyone** — including its own foreign-national engineers. Hundreds of millions of users. One letter. Friday, 5:21.\n\nAn honest caveat is necessary here. The *deemed export* framing applied to AI is, for now, analysts' reading — not the disclosed legal basis: the government named neither the statute nor the specific provision. Just Security describes ECRA 2018 / EAR 744.22 as \"inferred by analysts\"; the actual authority could be something else — IEEPA, AECA, ICTS. Plausible is not the same as confirmed.\n\nWhat is confirmed: the mechanism worked. Without public rulemaking, without 60 days of notice and comment — the government demonstrated it can shut down any deployed model for anyone outside the US, and that the provider will have no option but to comply.\n\n## Teeth: how Anthropic picked a fight with the Pentagon\n\nFriday at 5:21 was not a bolt from the blue. It is the final line of an invoice that had been accumulating since 2025.\n\nIn the summer of 2025, the Pentagon awarded contracts worth **up to $200 million each** to Anthropic, Google, OpenAI, and xAI for national-security AI applications; according to the Congressional Research Service, Claude became the most widely deployed frontier model in the department. The world's largest military budget wants the smartest machine — so far, a normal market transaction.\n\nThen came early 2026. In a US operation involving Venezuelan President Nicolás Maduro, Claude ended up in the chain. Anthropic asked: *exactly how are you using this?* The Pentagon demanded access \"for all lawful purposes.\" The company adapted its policy, but held firm on two things — mass surveillance of Americans and fully autonomous lethal weapons. Dario Amodei acknowledged such systems \"may prove critical for our national defense,\" and immediately added that they are \"simply not reliable enough to power fully autonomous weapons.\" He said no — with a reason. The state is not accustomed to hearing no with a reason.\n\nOn February 27, 2026, Trump ordered federal agencies to \"**immediately cease all use of Anthropic's technology**\" — in a Truth Social post. Secretary of Defense Pete Hegseth designated the company a \"supply-chain risk to national security\" under FASCSA, barring defense contractors from working with it and giving them a six-month transition period. On March 9, Anthropic filed suit (case 3:26-cv-01996, N.D. Cal.). Judge Rita F. Lin issued a preliminary injunction, describing the blacklisting as \"classic illegal First Amendment retaliation.\" The case is still live — it would be wrong to imply anyone \"won.\"\n\nAnd here is an irony worth a moment's pause. The same company that refused autonomous weapons and mass surveillance contracts was simultaneously running **Project Glasswing** — a government program in which Mythos 5 (the same underlying model, but with certain safeguards lifted) operates for US cyberdefense and critical infrastructure. Anthropic is not a pacifist. It is a **selective contractor**: it signed one contract, not another. The state decided that the selectivity itself was unacceptable.\n\n## Where national security is right — and where it lies\n\nFirst, let's give the argument its due weight — without that, this is opinion writing, not an autopsy.\n\nMythos-class is genuinely dual-use. Anthropic itself describes Mythos 5 as its \"most capable model for cybersecurity and life sciences\" — vulnerability discovery, drug design, biodefense screening. If even the safety-filtered sibling (Fable 5) can be jailbroken into \"read a codebase and find its holes,\" that is offensive cyber tooling at industrial scale. The *deemed export* doctrine is not a recent invention — it is thirty years old. And Europe's own panic about its dependence is the mirror image of the same logic: if someone else's access to US AI is dangerous for you, then your most powerful tool is dangerous in someone else's hands. \"Frontier capability as a controlled munition\" is not a tyrant's caprice; it is the logical terminus of four years of coherent policy: chips → weights → access.\n\nNow the dissection.\n\nFirst: the jailbreak used as justification was **narrow and non-universal**, by Anthropic's own account. But the detail that collapses the entire security frame in a single sentence: **the same capability exists in GPT-5.5**. They recalled one model — the hole is still open. The threat was not eliminated; the supplier who said no was.\n\nSecond: the directive swept up **Anthropic's own foreign-national engineers** — the people who built the system. That is not a scalpel. That is a guillotine swung wide.\n\nThird, most critically: **the legal basis was never disclosed**. A directive without a published legal foundation is not \"national security\" — it is administrative will that borrowed national security's vocabulary.\n\nThe logic of national security is real. Its application here is not.\n\n## The other side of the curtain: Europe is building the door but cannot pay for the lock\n\nNine days before the letter — June 3, 2026 — the European Commission unveiled the **European Technological Sovereignty Package**: Chips Act 2.0, the Cloud and AI Development Act, an open-source strategy, an energy roadmap. The target: triple data-centre capacity in five to seven years and reach compute self-sufficiency by 2035.\n\nTo understand the scale of the ambition, you have to see the starting point. As of May 2025: the US holds **~75%** of global AI compute, China **~15%**, the EU **less than 5%**. Not \"5–10%,\" as some roundings have it — strictly less than five. The financial gap is starker still: US hyperscalers are deploying roughly **$600 billion** in capital expenditure in 2026 against approximately **€10.6 billion** in EU sovereign-cloud spending (per Euronews). Two orders of magnitude.\n\nAnd the sharpest point — why Brussels started talking about \"kill switches\" at all. The trigger was not theoretical. After the Trump administration sanctioned the prosecutor of the International Criminal Court, **Microsoft cancelled his corporate email account**. Just removed it. \"We want to be sure nobody has a kill switch,\" said Commission Executive Vice President Henna Virkkunen.\n\nThe problem is that this certainty is declarative, not architectural. Bruegel makes the point bluntly: Nvidia and CUDA are \"the sovereignty blind spot\" of the entire programme. Every planned European gigafactory will run on American hardware and an American software platform. Tripled capacity on someone else's CUDA is sovereignty in a rented flat. Add the physics: grid connection queues in Europe run seven to ten years, with moratoriums in Dublin through 2028, in the Netherlands and Frankfurt through 2030. The land for gigafactories exists. The electricity does not yet.\n\nArthur Mensch, CEO of Mistral, put it precisely: \"You cannot regulate your way to computing supremacy.\" Europe has the rules. America has the compute.\n\n## Who holds your switch\n\nThe question is not rhetorical. If your product runs on a frontier model from Anthropic, OpenAI, or Google, a single letter from the US Department of Commerce can change your operational reality within hours. Not tomorrow. At 5:21 on a Friday.\n\nWho wins from the Compute Curtain? Washington and US cloud infrastructure — those on the right side of the control mechanism. US incumbents whose competitors suddenly turned out to be \"foreign nationals\" at the worst possible moment. The national-security apparatus, which now has a precedent: one directive, global shutdown.\n\nWho loses? Builders outside the US, for whom dependence on a single jurisdiction just became a documented risk. Allies in the \"grey zone\" — not enemies, but not the protected Tier 1 core either (and the EU, where all this is unfolding, would prefer to consider itself exactly Tier 1). Open-weight models are the one category where \"recall\" is physically impossible — once the weights are with you, they stay; that is not a reason for euphoria (quality lags, maintenance costs more), but it is a structurally different dependency architecture.\n\nWhat to do about it is not paranoia — it is engineering: jurisdictional awareness when choosing a provider, diversification of exit points (not one vendor, not one country), open-weight fallbacks for critical functions where quality is acceptable. Anyone building a product from Ukraine already knows this environment: external shocks arrive without warning, and the only reliable protection is architecture, not agreements. Compute dependency is the same logic; the difference is that this time the switch is held not by a missile but by a stamped letter.\n\nThose who own the infrastructure win. Those who mistook the cloud for a neutral utility — like running water — lose. Water mains have switches too. The question is only whose basement they're in.\n\n---\n\nThe Iron Curtain divided what you can see on a map. The Compute Curtain divides the right to speak with the most intelligent machine humanity has built — and it is drawn not in concrete and wire over decades but in half a day, by a letter that arrives on a Friday at 5:21. It appears on no map. But if your product thinks through someone else's model, you are already on the other side of it. The only question is when you will be told. And whether it will be a Friday."
}