But it's not a casualty count or a market crash that should worry New Delhi this month. It's the number NITI Aayog loves to cite: AI could contribute almost 1.7 trillion dollars to the Indian economy by 2035. That is the prize. That's what the week has just demonstrated about the extent to which that reward is at stake, on equipment that the country does not make and does not have under its own jurisdiction.

On Friday, June 12, the United States Commerce Department issued Anthropic a letter and within hours the company took its two most powerful systems, Fable 5 and Mythos 5, offline for all foreign nationals on Earth. The order came in at 5:21 p.m. Eastern. These models had been put live 3 days ago. But by the time the engineers of Bengaluru and Gurugram logged in, most of the strong commercial AI that they touched had simply ceased responding. Anthropic apologised and said the disruption will be temporary, while its other Claude models would continue to be available. But all that didn't alter the message that spread like wildfire down the technology highways of India: The most ingenious gadget in the construction business is a switch, and the switch is in Washington.

The depth of the commercial entanglement is the story: see? Anthropic's second largest market is India. Its run-rate revenue on Claude had increased by twofold in about four months and the usage of the company's coding tool had increased fivefold since the beginning of the summer, which its chief executive had pointed out to the Prime Minister with pride. A day prior to the blackout, Tata Consultancy Services announced a strategy to train 50,000 workers on Claude and establish a new Anthropic practice, a type of anchor contract that transforms a vendor relationship into infrastructure. Now that deal is on ice. A nation that had staked its productivity narrative on a foreign platform found out, in one news cycle, that it wasn't a decision it could make.

The plot is one of those corporate thrillers and the cast is important. According to sources, Amazon researchers, who have invested about 13 billion dollars in Anthropic through an earlier $8 billion investment and a recent $5 billion infusion, reportedly discovered a method to circumvent Fable 5's protections against vulnerability checks to search for software weaknesses. Amazon's chief executive took it up with Washington; according to press coverage, he called Anthropic's founder, who didn't answer, then called in top officials, who finally gave him three calls with the founder, involving key cabinet members such as the secretaries of Treasury and Commerce. The Commerce secretary said the order was put in place because the models might ultimately be used by China, Russia or other countries of concern for military or intelligence purposes. Anthropic challenges the framing, saying the action was disproportionate and the problem was small, and it was not dramatically different from other problems in other systems. About 100 cybersecurity experts openly criticized the ban as an overreaction. The precedent, however, seems to have been set: this is the first time the United States has targeted an AI model not the chips that power it with an export control.

That is a distinction that should be seated with, as it alters the strategic arithmetic for all countries but the United States. The power to build is curtailed by chip controls. A model-level control limits the user to use, immediate and remote, irrespective of where the user is sitting and how much he or she has already paid. It is not a procurement risk that can be mitigated by a contract to the detriment of the nation wiring AI into banking, healthcare, agriculture and government services. It is a matter of sovereignty and it answers itself when the lights go dim. A power which has to be able to operate its vital facilities on a basis which it does not own is open to the politics of whoever does own it, and whatever the commercial relationship may appear on a favourable day. A disagreement that starts in a Potomac conference room permeates an office park outside Pune within the hour, in this hyper-connected world. The buffer that used to protect countries has been eliminated, and the only real padding is when you have the stack you need, and don't rent it, and don't hope that the terms will look right.

It is what all of India's own builders have been saying this week, but they have said it with a level of bluntness. The CEO of one of the nation's biggest software companies termed sovereign AI as an "ultimate weapon" that can no longer be taken for granted by the country. The CEO of Bengaluru-based Sarvam, which is developing models in Indian languages, cautioned against equating access with ownership, and that adoption is not synonymous with advantage. Others called for being resilient with multi-model configurations, open-weight fallbacks, and on-premise deployments, so that no foreign directive can take down a product in a single night. A leading investor again took up this call for an infusion of about 50,000 crore rupees, or some five billion dollars, to be spent annually in building the homegrown capability, as opposed to once and celebration. One frequently cited policy analyst summed up the entire situation in a memorable statement: “American AI is tied to American geopolitics.”

However, when placed in the context of the actual expenditure by India, the gap becomes visible. The IndiaAI Mission has invested approximately USD 1.25 billion in sovereign capability. The fund proposed for this year would be more than three times that much, year on year. In 2025, Indian startups raised approximately 10.5 billion dollars, placing them third in the world by investments after the United States and the United Kingdom, but most of that money was invested in applications that are built over foreign models, not the models themselves. The need is there, the engineering skills are there, and so are consumer and enterprise markets. The basis from which everything else rests is what's been borrowed, quietly and on a grand scale.

That is where there's a real difference of opinion among respectable people as to how to bridge that gap, and I wouldn't want to mince my words either. There is one camp that believes India "must" build its own frontier foundation model or be "permanently exposed". Another, which is well represented in the venture community, believes that the focus on training a model from scratch is the wrong fight: compute access, engineering talent is the key rather than money, and the quickest route is through adapting and governing strong open-weight systems. Within days, the market responded: open models from Cohere, China's Moonshot and Zhipu, and even India's own Sarvam, which has deployed systems with 30 billion and 100 billion parameters, respectively, that have been fine-tuned for 22 Indian languages, emerged as the next obvious choice for enterprises struggling to keep their products on the market. The camp believes that sovereignty at the data and application layer could be more important than even a trophy frontier model that relies on imported silicon.

That imported silicon is the fly in the ointment that restricts triumphalism. We do not yet manufacture cutting-edge chips here, India is still far behind the U.S. in data-centre capacity and the 20,000 new GPUs announced with great fanfare this year do not make up for the hundreds of thousands of chips that are missing. Such a model, trained on Indian data, by Indian team, in Indian language, by American chip, with American government terms that can be tweaked at any time by anyone upstream, is sovereign only by sufferance. The Anthropic affair was just a revealing of that dependency one level up from the typical hardware debate.

It's also a test of an international rulebook that doesn't exist yet. In Korea, which is one of Anthropic's fastest-growing markets, twelfth by usage, the same blackout shook boardrooms in Seoul, where an executive told reporters he was "very confident" access would be available again within days. Europe's "digital-sovereignty hawks" took notes. The Commerce secretary and Anthropic's founder were both scheduled to attend the G7 meeting in France, while working-level negotiators such as the national cyber director were almost daily in Washington attempting to work out a formula for compliance. The odds of “return to access” before July now sit at 57 percent, betting markets said, and they have now surpassed three in four by mid-July. Most likely the freeze will break. In August, it could be dismissed as a footnote, a three-day outage that the lawyers and engineers were able to patch quietly.

That would be the incorrect method of filing it. But the outage does not set a precedent. It was two years since India was celebrating its speedy adoption of the world's best tools and June was the uncomfortable corollary that adoption borrowed is adoption that can be recalled, without notice, without consultation, and without regard for the trillion-dollar economy being built on top of it. The two model names in the middle of this storm will be forgotten. The question that they want to make them ponder, and ponder hard, is whether a rising power can afford to think on machines it doesn't own, and whether the urgency that sparked this week will remain once the switch is flipped back on. Self-reliance here has become more than an economic choice. It's no longer a matter of choice as it's become a necessity for India to be strategic.

---

A single-voice column article, updated every 12-24 hours with the latest developments (Seoul press conference; G7 context in France; daily Commerce negotiations with the national cyber director; and Amazon's entire $13B investment). The strategic-autonomy lens you requested is analysis, not attribution, and runs through the body. All paraphrased from the spread of coverage including the CNBC, Globe and Mail, The Hill, Business Today, Analytics India, Korea JoongAng Daily, Kalshi and others so there is no copyright exposure.

The updated data graphic to accompany it is all home made, hence no licensing issues.For the lead image, the image should be a clean editorial fit (no branding or logos), and a data-centre server hall or a server rack with a circuit motif is a modern Indian piece of generic infrastructure imagery. Take any stock that your CMS is licensed for (Sify, Adani and Cyfuture all have usable data-centre frames), and then throw in a caption such as: India is conducting a large portion of its daily business using AI, most of which is on models they do not own.

Here are a couple of editing notes. The infographic now contains four data points: the adoption gap (41/26/19), the context of Amazon spending in India ($1.25B committed, the proposed $5B per year fund and $10.5B in 2025 startups funding), the split of the funds between Amazon, and the prediction-market odds of restoration. This puts you in a position to have two bar panels and two pie panels in a single image.