Written by: Arindam Goswami5 min readOct 15, 2026 02:27 PM IST First published on: Oct 15, 2026 at 02:27 PM ISTThe Indian government recently told the Lok Sabha that the IndiaAI Mission had sanctioned 93 lakh hours of chip time for 237 projects. For comparison’s sake, consider this. When the Chinese lab DeepSeek built its V3 model, it needed about 2.8 million hours of similar chip time, a figure it published itself. India has already sanctioned about three times that.The comparison is rough. The chips are not all of one type, and sanctioned hours are not the same as hours used. The H800, which DeepSeek used, is Nvidia’s export version of the H100, with chip-to-chip bandwidth cut by about half, so an H800 hour is a fair stand-in for an H100 hour in raw compute. Still, the direction is clear. Sarvam, a Bengaluru start-up, was given 4,096 Nvidia H100 chips for six months. DeepSeek trained V3 on half that many chips, in under two months. On paper, Sarvam had far more to work with.AdvertisementAlso Read | A Dean’s dilemma in India’s AI EconomyTo be fair, Sarvam delivered something commendable. Its models were built from scratch in India, and the company says the larger one holds up against open models that OpenAI and Alibaba released in 2025. That is a good result. But it is also behind the best by quite a margin, and chips are not what can explain this difference.Renting more chips abroad is possible, but it depends on decisions in Washington. In 2025, Bloomberg reported that the US was drafting curbs on chip shipments to Malaysia and Thailand. Takshashila’s research on technology denial finds that new technology tends to spread within a few years, which limits how long such curbs last. Planning around them is still unwise.What is actually scarceBuilding a model is only the first stage. Then comes the job of making it useful: Answering a legal question correctly, advising a farmer, writing a medical note in Marathi and so on. These tasks require fewer chips, although it could be for more runs. For V3, it used less than 0.2 per cent of the total chip time. These jobs need people who know the subject, and tasks whose answers can be marked right or wrong.AdvertisementSuch a set of tasks, with a marking scheme, is what is called a reinforcement learning environment. One example is a bail application drafted by the model and checked against a list written by senior advocates. Another example could be a crop advisory checked by agricultural scientists.India has such people in courts, hospitals, farm universities, language institutions, etc. This work is valued too. Meta paid $14.3 billion for 49 per cent of Scale AI, a company that supplies this kind of data. And most importantly, chip export rules generally don’t apply to this kind of work.What India can doFor starters, India should stop counting chips and start asking for results. One large block of chips could be held back for month-long runs and given to the teams whose models score best on Indian-language and subject tests. Second, it is important to pay experts to write the tasks and marking schemes, publish them, and let any Indian lab use them. The costing for this should be small next to the mission’s Rs 10,372 crore.The best free models trail the paid leaders by only about four months, according to Epoch AI. India can use them as they are. Its own money is better spent on what nobody else will build: AI that understands Indian law, medicine, farming and languages, and proof that it does.you may likeMoreover, on hardware, it is a good strategy for India to bypass the tightly controlled chokepoints. Accelerators and lithography run the risk of being closed by export controls, quite often on a whim. But the packaging and assembly layers of the supply chain are fragmented and have no export controls generally pointed at them. India has a few operating packaging plants already. Deepening that, particularly towards memory packaging with Micron at Sanand, is a more winnable aim than a leading-edge fab.India could also convene the countries the US AI diffusion rule effectively invented before it was scrapped — those with capital and engineers but no advanced fab, from the Gulf to parts of Europe. These countries buy a lot of compute. If this can be combined with open-weight models, a good hedge one could get is to keep national mirrors of significant open-weight releases across these countries. That provides availability and clustering advantages.The writer is research analyst at The Takshashila Institution in their High-Technology Geopolitics Programme