India's AI Build Was Never a Reaction to Washington

The received story is clean: in June 2026, the Trump administration forced Anthropic to cut off foreign access to its most advanced models, India's IT giants lost the tools they had built their AI practices on, and New Delhi responded by accelerating a sovereign AI build. The ban as catalyst, the build as hedge. The timeline says otherwise. India launched its sovereign AI mission in September 2025, committing Rs 10,300 crore to indigenous large language models and naming eight partners including BharatGen, which is building a trillion-parameter model across 22 Indian languages. ^[1] In December, a government white paper warned against the concentration of compute power in "a few global firms and urban centres" and proposed AI infrastructure as a shared national resource. ^[2] By February 2026 — four months before the Anthropic ban — India had unveiled a $200 billion sovereign AI roadmap spanning five layers from applications to energy, ^[3] and Sarvam AI had launched indigenous models that outperformed Google Gemini 3 Pro and OpenAI ChatGPT on Indian-language document tasks. ^[4] The strategy was not a reaction. It was already ten months old. Once you read the chronology correctly, a pattern that looked like reaction becomes visible as design. The design is a deliberate straddle at every layer of the stack — domestic capacity held alongside foreign dependence, not in sequence but simultaneously, as if the point is not to eliminate the dependence but to calibrate it. Start at the model layer. India is building BharatGen's trillion-parameter sovereign LLM and has shortlisted twelve teams including Sarvam AI and IIT Bombay to develop indigenous foundation models. ^[5] At the same time, it joined the US-led Pax Silica chip alliance in February 2026 ^[6] and, just this month, the BRICS+ coalition developing a sovereign AI stack explicitly designed to "decouple from Western cloud infrastructure." ^[7] The same country sits in both rooms. Move down to chips. In May, C2i Semiconductors taped out an AI power chip conceived, architected, and verified entirely in India — a shift from design services to original chip development. ^[8] The same month, Intel and 3DGS committed $3.3 billion to an advanced packaging substrate facility in Odisha, using Intel's glass-core technology under the US-India TRUST Initiative. ^[9] The facility is explicitly packaging, not leading-edge logic fabrication — a distinction that matters. India is building chip-design capacity and packaging infrastructure while leaving the hardest layer, advanced-node fabrication, to partners. At the compute layer, the straddle is starkest. India has deployed 38,000 GPUs toward a target of 200,000, ^[10] and is developing sovereign GPUs with NVIDIA through a three-to-four-year plan. ^[11] NVIDIA is also building gigawatt-scale AI factories in Chennai and Mumbai through a joint venture with Larsen & Toubro, and partnering with India's payments infrastructure to build a native AI foundation model on NVIDIA's Nemotron architecture. ^[12] The sovereign GPU is being designed with the company whose hardware currently runs the entire operation. The procurement strategy makes the logic explicit.

We are not going to necessarily say that we will buy only NVIDIA GPUs. Our approach is whoever produces the chips — Maya Sundarakrishnan

At the data-center layer, the same dual track holds. Anthropic launched in-country Claude inference via AWS Bedrock this month, offering data localization and audit trails for India's regulated sectors. ^[13] Alongside it, Andhra Pradesh is building a 100-kilometer-radius AI data city in Visakhapatnam with $175 billion across 760 projects, backed by six approved nuclear plants and tax holidays through 2047. ^[14] Foreign-hosted inference and a domestic nuclear-powered data city, advancing in parallel. The intellectual frame for all of this is not autarky. Former Foreign Secretary Nirupama Rao defined the goal precisely.

The objective must be strategic agency: the ability to shape options, influence standards and make choices that others must take seriously. — Nirupama Rao

The India Narrative publication, a policy forum close to the government's thinking, explicitly argued for "managed interdependence with partners in Europe, Singapore, and the Gulf states" — not total technological autonomy. ^[15] Sarvam AI sharpened what the strategy is for after the Anthropic ban hit.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

The sovereign stack is real but incomplete — and the incompleteness is the point. India's ten approved semiconductor units are heavy on packaging and chip design, with no leading-edge logic fabrication comparable to TSMC or Samsung's advanced nodes. ^[16] The white paper's warning against compute concentration in a few global firms ^[2] sits alongside the fact that every GPU in the country today is foreign-made. India is building to the exact height from which it can bargain, not to the height from which it can stand alone. The Anthropic ban, when it came, was the proof case. In June 2026, after the Mythos AI model penetrated nearly all US classified systems within hours, Trump ordered Anthropic to suspend foreign national access to its Fable 5 and Mythos 5 models. ^[17] Tata Consultancy Services and Infosys lost access to the models they had integrated into their AI practices. ^[18] The vulnerability of dependence became concrete. What happened next is the evidence that the straddle is a strategy, not a contradiction. Two weeks after the ban, India and the US held a high-level strategic technology roundtable deepening cooperation "from chips to neural networks," framing the partnership as building "trusted, resilient technology ecosystems" to reduce dependencies on unreliable supply chains — meaning China, not Washington. ^[19] India did not retreat from US alignment. It held the line, deepened the partnership, and kept building its own stack on both tracks simultaneously. The strategy is not a hedge against any single power. It is a calibrated build of just enough domestic capacity at every layer — models, chips, compute, energy — to convert raw dependence into negotiating leverage. The Anthropic ban did not start this. It validated a bet India had already placed, ten months earlier, that in a world where AI access can be cut off overnight, the only durable position is to have enough of your own to make the terms of access a negotiation rather than an ultimatum. From both Washington and Beijing.