America’s AI Kill Switch and India’s Chokepoint Problem

Three days in June told us more about how technology geopolitics will work than any white paper could.

Every few months an event arrives that tells you more about how the world will actually work than a stack of white papers. The three days in June 2026 when the United States switched off two of Anthropic’s frontier models were one of those events. Most of the coverage filed the story under AI safety, or under one company having a rotten week. I would file it elsewhere. This was the moment export controls quietly climbed a rung on the ladder, from the chips to the cognition the chips produce. Once you read it that way, the consequences for India become hard to look away from.

On 12 June, the US Department of Commerce, acting through its Bureau of Industry and Security (BIS), ordered Anthropic to cut off access to its two newest models, Fable 5 and Mythos 5, for any foreign national. The order did not stop at people abroad. It covered foreign nationals sitting inside the United States, including, on a literal reading, Anthropic’s own non-citizen staff. No company can verify the passport of every user in real time, so Anthropic did the only thing available to it and pulled both models for everyone. They stayed dark for close to three weeks, until the restrictions were lifted on 1 July after the firm promised to tighten its safeguards, with the government pointedly reserving the right to switch them off again. 

But why is this alarming? Because technically, researchers at Amazon had shown that Fable 5 could be coaxed past its safety rules into finding and exploiting serious software vulnerabilities. While, US suspected a China-linked group had already probed Mythos, and, separately, Anthropic had been in a standoff with the Pentagon after refusing to let its models be used for mass domestic surveillance and fully autonomous weapons. The Defence Secretary had taken to calling the company a “supply chain risk,” a phrase usually reserved for the likes of Huawei. The trigger, in other words, was a tangle of safety, espionage and bureaucratic pique, which is how most consequential state actions come about.

A kill switch, not a chip ban

What makes this different from every export-control story before it is the object being controlled. For 70 years, export controls acted on things: machine tools, cryptographic software on a floppy disk, lithography machines, chips. Here, for the first time in public view, the control fell on a live, commercially sold service that people reach through an API. The legal point was an old and quietly powerful doctrine called the deemed export rule, under which showing controlled American technology to a foreign national, even one standing in your San Francisco office, counts as an export to that person’s home country. If you extend that logic to a cloud model anyone in Bengaluru or Berlin can call, you would turn a subscription into a border.

The moment access turns on who you are rather than what you pay, every company that builds on the model inherits a border-control problem it never asked for. A routine API call now needs an identity check sitting behind it. A large firm has to map every place a foreign engineer might see a restricted output, down to the log files, and be able to prove it does not happen. No single step is unreasonable. Stacked together they become a quiet tax on ordinary work, and the weight of it falls hardest on the small firms that were supposed to build on this technology in the first place.

Why chokepoints punish those who pull the trigger

Step back, and the episode fits a pattern that strategic-studies types have taken to calling weaponised interdependence: the idea, developed by Henry Farrell and Abraham Newman and popularised in Edward Fishman’s Chokepoints, that whoever sits at a central node of a global network can convert that centrality into coercive power. SWIFT is one case study, the dollar is another, and American AI models running on American clouds are the newest. The older Indian phrase for the underlying condition is matsyanyaaya, the logic of the fishes, where in the absence of a referee the big fish eats the small. Technology geopolitics runs on this logic far more than we like to pretend.

History keeps teaching the same lesson and the US keeps forgetting i.e. a chokepoint is powerful mostly as a threat, and it begins to lose its power the moment it is actually used. The record is remarkably consistent on this.

In the 1990s, the United States classified strong encryption as a munition and licensed it like a weapon. Activists printed the banned code on T-shirts to make the absurdity visible, legally turning cotton into contraband. The real result was that European and Asian software firms, unburdened by the rules, ate into American market share until Washington relented in 1999.

In satellites, the “See-Through” rule meant a foreign spacecraft carrying even a single American chip fell under strict US licensing, so European builders such as Thales Alenia simply designed American parts out and advertised “ITAR-free” satellites. American manufacturers lost roughly a quarter of the global market in a decade, and China’s own launch industry got a helpful shove. The chip controls of the past few years are running the same script in fast-forward. They did not so much stop China as hand Huawei and DeepSeek both a reason and a subsidy to build at home. 

Each of these is the same story told in a different technology. You deny access to something general-purpose, and you do not remove the demand; you relocate the supply, while Anthropic’s API is just a new chapter. Which is the reason why the sharpest way to put the lesson is that control over a chokepoint buys you leverage, but using it spends that leverage down. Every time the switch is flipped, engineers around the world start quietly designing the switch out.

India’s say-do gap

For India, the discomfort is not the episode itself but what it exposes: the distance between what the two governments say to each other and what the rulebook actually does. On paper, the relationship glows. iCET, now rebranded TRUST, along with INDUS-X, promise deep cooperation on AI, quantum computing and defence, and Secretary of State Marco Rubio’s recent Delhi visit added an underwater-domain-awareness roadmap for tracking submarines across the Indo-Pacific. Underneath the warmth, India sits in a grey zone. It has been a “Major Defence Partner” since 2016, a bespoke label meant to ease access to dual-use technology, yet because it guards its strategic autonomy and declines to become a formal treaty ally, the American security system keeps it at arm’s length.

That gap has widened recently as the earlier AI Diffusion Rule, restrictive but at least written down, was scrapped in favour of an ad-hoc, case-by-case licensing regime with no published criteria, administered by a BIS that had shed a fifth of its staff. Reviews that once took weeks began taking months. And the double standard was not subtle. While Indian deep-tech firms waited in unpredictable queues for compute, the POTUS personally brokered a deal to widen Chinese access to advanced chips. Rubio’s visit brought India no matching relief. The signal Delhi received was blunt, and I think correct: treating American technological consistency as a dependable input is itself a national-security risk.

Atmashakti, and the interoperability trap

So India is hedging, and I was to name it, tech atmashakti i.e. the pursuit of enough indigenous technological strength that no foreign supply chain can hold you hostage. The best move underway is a chip-agnostic pivot, instead of waiting for high-end, licence-controlled American hardware, Indian defence and deep-tech firms are re-architecting their software so that it does not depend on any single vendor’s chips or tools, Nvidia’s CUDA above all, which lets their systems run on cheaper, older, or domestically fabricated silicon. It is the satellite story in reverse. Where Europe once designed American parts out of its spacecraft, India is designing the American chokepoint out of its stack.

However, the very hedge that protects India quietly dismantles the joint projects America wants to build. Shared drone swarms and real-time submarine tracking need split-second synchronisation. If the United States builds on proprietary American chips and closed models while India builds on chip-agnostic, open ones, the two technology stacks drift apart, and in a live maritime crisis the lag from translating between them makes joint operations unworkable.

Worse, American military networks enforce a zero-trust rule that automatically rejects any system not built on approved, compliant software. By swinging export controls like a club, the Commerce Department is nudging India toward exactly the architecture that makes the State Department’s alliance impossible. One arm of the American state is undoing the work of the other, which is the kind of unintended consequence worth pausing on.

Not all dependence is the same

There is a temptation, after an episode like this, to treat every foreign model as a liability and to conclude that the only safe course is to build the whole stack at home. That instinct flattens a distinction that matters more than almost anything else in this debate. A model you rent through someone else’s servers and a model you download and keep are not exposed in the same way. When a system reaches its intelligence through a foreign API, each request travels abroad and is judged under another country’s law, which means it can be severed the instant that country decides to sever it. The reason Fable 5 could go dark between one afternoon and the next is that the link was never truly yours. It was a service, renewed with every single call.

Now picture the opposite arrangement. An Indian firm pulls down the parameters of an open model and runs them on machines it rents inside the country. Nothing in that setup loops back to whoever trained the thing. There is no console abroad with a button that ends the workflow, and no query ever leaves the premises. Judged by where the data actually goes, this can be the safer choice even when the open model was made in China, which is the part that tends to catch people off guard. Most of the dread about a hidden trapdoor sewn into the weights misreads what a weight is. A trained model is an inert heap of numbers. It keeps no line to the outside world and cannot quietly forward anything anywhere.

The genuine concerns about Chinese technology, a doctored chip or an owner who ultimately answers to the Party, are legitimate, but they live in the hardware and the ownership, not in a file you can download. So the takeaway from June is not to shun foreign models. It is to lift anything critical off live foreign endpoints and set it on weights you host on hardware you run.

That still leaves one exposure no amount of downloading can dissolve. Owning the weights settles the questions of access and data; it says nothing about the machines beneath them. Nearly every high-end processor doing the real work in an Indian server hall was designed in the United States, and stays within reach of an American licensing decision. So the pinch-point does not disappear. It slides a level lower, from the model down to the chip. Doing everything at home is no escape either, since fabricating leading-edge silicon at scale is years away for India.

The workable posture has two halves. Put serious inference capacity in place now, while capable open models can still be had without restriction and before the weather turns in some capital, and keep relations with Washington steady enough, through arrangements like Pax Silica, that the chips keep arriving. Autonomy in this domain does not mean owing nobody anything. It means never letting one reliance grow so deep that it can quietly settle your choices for you.

Making ourselves harder to coerce

None of this argues for cutting the cord. Trying to make everything at home is a Sisyphean errand, since no country holds every mineral, tool and skill, and every supply chain hides a chokepoint somewhere. The realistic response, and the one Delhi should adopt, is narrower. Draw up a short, honest list of genuinely strategic dependencies, the ones whose sudden loss would actually hurt, and de-risk those through a mix of domestic capability and reciprocal arrangements with trusted partners. For everything past that list, let ordinary diplomacy handle the work of staying on the right side of other people’s chokepoints. And police the thing that genuinely governs the risk, which is where the data travels and whose machines do the running, rather than pinning a blanket suspicion on a model for the nationality of the people who built it.

Two temptations are worth resisting along the way. The first is the lure of prestige at the frontier. India’s edge lies not in training the biggest model but in the layer below it, cheap inference, rich data in Indian languages, and applications that will run on whatever silicon is available, and that is where public money does more good. The second is the urge to fold everything into a single national champion, one AI cloud to rule them all, in the way a lone payments rail has ended up carrying the whole economy. That design gathers into one place the very vulnerability this affair should teach us to scatter. Far better to run several inference providers than to hand an adversary a single switch it can locate and press.

The larger move is to stop swimming as a solitary small fish and to convene what one might call an Open Technology Maitri, a grouping of middle powers that pools work on open models, open hardware and supply chains no single capital can cut. The point is not to spite the United States. It is to drain the value out of coercion by keeping the alternative dull, cheap and always to hand.

US, for its part, has a contradiction of its own to sort out. It cannot ask its partners to build their futures on American technology and then treat access to that technology as something to be dangled or withdrawn on a bad day. Until the rulebook grows up to match the speeches, the sane response for everyone else is the refrain this whole episode keeps circling back to. Anticipate the unintended, and build around the chokepoint, but remember that sovereignty here arrives in layers, so that the model which costs the least and moves the most freely, once it runs on machines you control, may in the end guard your independence better than the costly one you are only allowed to borrow.