AI and digital platforms: Follow one offering through the chain
Model builders and apps selling AI as assistants, APIs and copilots, not hardware
Follow one offering through the chain
Sarvam's Indian-language AI model, built in India
Follow one answer in an Indian language back through the app, the model, public data, subsidised GPUs and the chips beneath.
Chips designed in America, made in Taiwan
80-90%Nvidia share of AI accelerator revenueEvery answer is computed on graphics processors, mostly Nvidia's, designed in California and fabricated and packaged by TSMC in Taiwan. Analysts estimate Nvidia took 80-90% of AI accelerator revenue in 2025. Sarvam's models were built on Nvidia hardware hosted in India, so the costliest and most concentrated part of this chain sits abroad.1
Constraint Supply depends on one foundry's advanced packaging capacity in Taiwan, and US export controls decide which countries may buy the top chips. Si SiliconCu CopperAu GoldHe HeliumESG Electronic specialty gasesSemiconductorsSubsidised GPUs on Indian clouds
₹65/hrsubsidised IndiaAI GPU rateThe chips sit in Indian data centres, such as Yotta's Shakti Cloud, rented out by the hour. Under the IndiaAI Mission more than 38,000 GPUs had been onboarded by February 2026, offered at about ₹65 an hour, nearly a third of the global average, the government says. In April 2025 the government said it would allocate Sarvam 4,000 GPUs for six months.2,3,4
Constraint Data centres used about 415 TWh, 1.5% of world electricity, in 2024; grid connections, transformers and cooling water limit new campuses. Cu CopperH₂O WaterTc Thermal coalDC Data centresLFP Lithium iron phosphate batteriesData centres and cloudElectrical equipmentComputers and electronicsIndian-language data, gathered and shared
7,541datasets on AIKosh, February 2026A model is only as good as its text and speech. The web holds far less material in Indian languages than in English, so public platforms help: the government's AIKosh listed 7,541 datasets and 273 models across 20 sectors by February 2026, and Bhashini supports more than 36 languages. Labs then clean, label and add licensed and synthetic data.2
Constraint Scarce high-quality text in many Indian languages, and unsettled copyright rules on training data, limit what models can learn. Media and entertainmentSoftware and IT servicesEducation and researchPublic money beside venture capital
₹10,372crIndiaAI Mission outlay, five yearsTraining runs cost more than most Indian start-ups can raise. The IndiaAI Mission has an outlay of ₹10,371.92 crore over five years, approved in March 2024, and picked 12 foundation-model builders, including Sarvam and IIT Bombay's BharatGen, from more than 500 proposals. Venture investors fund the rest, but on a far smaller scale than in the US.2,5
Constraint The US attracted US$ 285.9 billion of AI investment in 2025; India's five-year mission totals about US$ 1.2 billion (calculated). Finance: banking, insurance, marketsA 105-billion-parameter model from Bengaluru
105bnparameters in Sarvam's larger modelSarvam, a Bengaluru start-up, trained its own models rather than adapting a foreign one. At the India AI Impact Summit in New Delhi in February 2026 it showed a 30-billion-parameter model for real-time conversation and a 105-billion-parameter model handling up to 128,000 tokens, both supporting all 22 scheduled Indian languages and built with IndiaAI, Yotta and Nvidia support.6,5
Constraint Frontier labs spend far more: OpenAI's compute bill was about US$ 16.3 billion in 2025, so Indian models must compete on language and cost. DC Data centresData centres and cloudEducation and researchWired into banks and public services
$439bnworld AI services spending, 2025India's biggest AI strength is fitting models into other organisations' systems. IT services firms such as TCS and Infosys, and model makers themselves, connect models to documents, call centres and government portals, adding retrieval, security and testing. Gartner put worldwide spending on AI services at about US$ 439 billion in 2025, a field where Indian IT firms compete strongly.7
Constraint Clients can switch integrators within weeks, and AI coding tools may shrink the headcount-based billing on which Indian IT firms earn. Software and IT servicesVikram answers in your language
22scheduled Indian languages supportedThe user meets the chain as a chatbot or voice assistant. Sarvam showed Vikram, a voice-first multilingual chatbot, at the February 2026 summit, plus a model it says reads documents in Indian scripts with over 84% accuracy (company claim). Every question runs the model again on rented GPUs, so serving millions of users is a running cost, not a one-off.6
Constraint Every answer consumes inference compute, so without cheap GPUs and power, free Indian-language assistants struggle to cover their costs. Software and IT servicesData centres and cloudFinance: banking, insurance, marketsWhat the chain shows
India can build its own language models on subsidised compute, but the chips, and most of the money, still come from abroad; its edge is services and languages.
Sources
- Nvidia AI accelerator market share 2024-2026 (analyst estimates) Silicon Analysts, 2026
- Democratising AI in India: expanding access to AI resources, skills and technology nationwide Press Information Bureau, Government of India, 2026
- Sarvam AI to build India's first homegrown GenAI platform in six months under IndiaAI Mission Entrepreneur India, 2025
- Energy and AI: Energy demand from AI International Energy Agency, 2025
- The 2026 AI Index Report: Economy Stanford HAI, 2026
- Sarvam AI unveils indigenous multilingual foundation models at India AI Impact Summit India Science and Technology (Government of India), 2026
- Gartner says worldwide AI spending will total $2.5 trillion in 2026 Gartner, 2026