India Discovered It Does Not Control the Switch of Its Own Digital Economy
Late on a Friday afternoon. A press release from Anthropic lands in the inboxes of its global partners with the neutral, contained tone of a system maintenance notification. The text announces that the Fable 5 and Mythos 5 models are suspended for all foreign nationals, including the company's own employees who do not hold American citizenship. The cause: a directive from the United States government invoking national security concerns linked to an alleged jailbreak vulnerability.
The timing could not have been more eloquent. Hours earlier, Anthropic had publicly celebrated its partnership with Tata Consultancy Services to accelerate the adoption of artificial intelligence in Indian enterprises. India, which both Anthropic and OpenAI describe as their second-largest market after the United States, had just discovered something that its founders, investors, and officials preferred to keep in the realm of abstraction: access to the tools that underpin a large portion of its technological ambitions can be shut down with a call from Washington, without prior hearing and without a defined restoration timeline.
What followed was not merely a reaction of indignation. It was the beginning of a public and accelerated audit of the design of the technological strategy of a country that has spent years building on foundations it does not own.
The Dependency Nobody Wanted to Name
India has spent more than a decade positioning itself as a powerhouse of technological services. Its developer base, the density of its startup ecosystem, and the weight of its major IT firms such as Infosys, Wipro, and TCS made it a mandatory destination for any technology company with global ambitions. Anthropic and OpenAI opened offices, hired local talent, signed alliances with system integrators, and described the country as a central market for their expansion.
The problem with that model is that all of the value infrastructure rested on foundational models developed, trained, and governed in California. India consumed the final product, integrated it into applications, distributed it to enterprises, and built specialised value layers on top of it. But it controlled none of the decisions that define how powerful that product is, nor when it ceases to be available.
That is not technological dependency in the abstract sense. It is geopolitical supply risk operating at the software layer, something for which the majority of Indian organisations had neither coverage nor a contingency plan. The Anthropic episode made it concrete in less than 48 hours.
Vijay Rayapati, co-founder of Atomicwork, articulated the operational consequence with precision: if access to the most advanced models is filtered by citizenship, companies with distributed teams split between engineers in Bengaluru and product managers in San Francisco are structurally disadvantaged compared to firms whose teams are entirely American. This is not a minor disadvantage. In industries where development cycles are measured in weeks and the capability differential between models translates directly into iteration speed, unequal access to tools becomes a cumulative competitive disadvantage.
Prasanto Roy, a technology policy expert based in New Delhi, was more direct about the systemic implications. The comparison he used was not with another episode in the technology sector. It was with Russia's exclusion from the SWIFT system following the invasion of Ukraine: a foreign policy measure that instantly reshaped the financial architecture of a country. His thesis carries weight because it points to the correct pattern: export restrictions on artificial intelligence models operate with the same logic as controls on critical infrastructure, and until now India had chosen not to treat its exposure to that logic as a problem of strategic design.
The Ecosystem That Built on the Layer It Did Not Build
There is a thread running through the entire Indian reaction to the Anthropic episode that deserves examination without condescension or excessive optimism: India's artificial intelligence ecosystem bet almost entirely on the application layer and specialised its value in adapting third-party models to local contexts, without seriously building the foundational layer that grants access to that adaptation in the first place.
That was not necessarily a mistaken decision in terms of capital efficiency. Training a frontier foundational model costs, according to reasonable industry estimates, anywhere between hundreds of millions and several billions of dollars, depending on the approach. For the majority of actors in the Indian ecosystem, that investment had no individual economic justification. Building on existing models and concentrating on applications allowed for the generation of real value with manageable budgets.
The problem is not the decision itself. The problem is that this decision was never accompanied by a supply risk mitigation strategy. There was no serious development of domestic backup alternatives, no public investment at the scale that the strategic role of that dependency demanded, and no systematic incentives for enterprises to diversify their foundational model providers.
Sarvam, one of the few Indian laboratories that advanced toward its own open-source models, represents the exception that proves the rule. Krutrim, which began with foundational ambitions, pivoted toward cloud infrastructure and AI services when it encountered the cost and capability realities that path demands. The rest of the ecosystem, including initiatives such as Avataar AI with its video generation model, operates on top of third-party models and adds value at the layer of cultural adaptation, speed, or price. That has genuine merit, but it does not resolve the vulnerability that became visible on Friday night.
Sridhar Vembu, founder of Zoho, reacted with a statement that sounds not like political rhetoric but like an architectural diagnosis: "technology is the ultimate weapon." His recommendation that Indian organisations adopt smaller models, both Indian ones and open-source models from other geographies, points toward a strategy of provider diversification at the foundational layer. The proposal from T. V. Mohandas Pai, former Infosys executive, was more ambitious in scale: an annual fund of 500 billion rupees for artificial intelligence and deep technology, plus a credit guarantee programme of 2 trillion rupees for computing infrastructure, hardware, and semiconductors. For reference: the IndiaAI Mission approved in 2024 contemplates 103 billion rupees distributed over five years. The gap between what currently exists and what Pai proposes is of an order of magnitude.
Hemant Mohapatra, partner at Lightspeed, introduced the necessary nuance: capital is not the only bottleneck. Talent, access to compute, and the capacity for sustained execution are equally decisive in building models that are competitive at a global level. That is the kind of argument that unsettles simple plans. Technological sovereignty is not built on public budget alone; it is built through an architecture of incentives, capacity building, and the accumulation of learning that takes years. India has some of those ingredients, but it does not have them assembled in a way that produces foundational capability.
When the System Design Reveals the Risk That Success Concealed
What makes this episode interesting from a design perspective is not Washington's decision nor Anthropic's response. It is the architecture of dependency that was exposed when both decisions collided with the reality of the Indian market.
For years, the relationship between India and the major American artificial intelligence platforms operated with the logic of a mutually beneficial alliance. India contributed talent, adoption at scale, and a rapidly growing market. The companies provided access to the most powerful models and the possibility of building on top of them. That relationship generated genuine value in both directions and explains why Anthropic and OpenAI prioritised India as their second market after the United States.
The problem with that model is structural: in any architecture where one party provides the layer that no one else can replicate in the short term, the party that consumes that layer has dependency without genuine negotiating capacity when the provider faces external restrictions. The size of the market does not matter, nor does the volume of the commercial relationship, nor the solidity of the alliances signed with TCS or Infosys. When a government directive arrives, the size of the second-largest market does not stop the suspension.
That does not make Anthropic a bad-faith actor nor the United States government an adversary of India. What it reveals is that the design of India's technological strategy assumed that commercial logic would protect access, and that assumption turned out to be incomplete. The absence of a credible alternative plan is not a moral failure but an architectural flaw: nobody designed the system thinking about what happens when the switch is in someone else's hands.
The reaction of sector leaders in the 48 hours that followed carries the tone of people discovering that the building they inhabit has no emergency exit. Not because nobody knew that exit might be needed, but because building alternatives required accepting that present success did not guarantee future access. And that is the moment when the exploitation of the present becomes a trap: when the dependency is so deeply integrated into the business model that imagining the system without it feels like imagining collapse rather than precaution.
Technological Sovereignty Is Not Resolved With Budget but With Prior Design
The Indian debate on sovereignty in artificial intelligence did not begin on that Friday. It existed before, with less urgency and a smaller audience. What the Anthropic episode did was transform it into a conversation with immediate operational consequences, visible to founders, investors, corporate CIOs, and technology policy officials simultaneously.
That simultaneity has value. It also carries a risk: that the response amounts to an emergency plan rather than a systemic redesign. Emergency plans fund what is urgent. Systemic redesigns build capabilities that reduce the probability of urgency recurring.
The difference between the two is not merely one of budgetary scale. It is one of the sequence of decisions. Funding foundational models without first having resolved the problem of specialised talent and sustained computing capacity produces investment that does not scale. Diversifying model providers without having built the organisational processes to evaluate and migrate between them produces a dispersion of resources. Declaring technological sovereignty as a national objective without having designed the governance mechanisms that align private incentives with public objectives produces policy documents that do not change real behaviours.
India has genuine capabilities to build a different position in artificial intelligence. It has technical talent in quantity, a domestic market that generates data and unique cultural contexts, and a track record of scaling digital infrastructure at unprecedented speed and cost, as demonstrated with UPI and Aadhaar. What it lacks is not declared willingness nor the budget that could eventually be allocated. What it lacks is the prior design that converts those capabilities into an architecture of resilience before the switch is activated, not after.
The Anthropic episode is a diagnosis, not a catastrophe. But diagnoses have a shelf life. If the reaction is consumed in the debate over how many billions should be allocated to the AI fund and does not produce changes in how Indian organisations design their relationship with foundational model providers, the next access cutoff will find the same system, with a different model name and the same absence of an emergency exit.
A country that has spent years being the second-largest market for tools it does not control does not have a problem of vision. It has a design problem that confused access with ownership, and market size with negotiating power. Those two errors, taken together, are precisely the kind of fault line that remains invisible until someone turns off the lights.
