How are India and Japan approaching sovereign AI?

Sovereign AI does not look the same everywhere. A country's language, industry and bank balance shape what it wants from AI. India and Japan show two clear answers, and both are useful for Canadians because neither country is trying to beat the American labs at everything.

India's priority is reach: many languages and many smaller users who need cheap access to chips. Japan's priority is industrial: it wants to keep up with the United States and China as AI moves into robots and other machines.

India: rent the chips, share them cheaply

India's government launched the IndiaAI Mission with Rs. 10,372 crore (a crore is 10 million, so that is 103.72 billion rupees). Its compute pillar puts GPUs on a shared portal and subsidizes access for startups, researchers and public bodies.

• Capacity. The government reports more than 38,000 GPUs onboarded, with 20,000 more announced at the AI Impact Summit in February 2026.
• Price. Users pay a subsidized ₹65 per GPU-hour.
• Models. More than 500 proposals came in, and 12 startups were selected in the first two phases, including Sarvam AI.
• Uptake. By March 2026, 190 projects were approved: 78 with government bodies, 46 with startups and small firms, 27 with researchers.

Sarvam shows what this can produce. It released two open models, 30B and 105B, trained from scratch in India on IndiaAI compute, using a design called mixture of experts, in which only part of the model works on each token. Weights are on Hugging Face and AI Kosh, India's public data platform. Sarvam reports leading results on Indian-language tests. Those are the company's own benchmarks, so independent testing is still worthwhile. BharatGen, a government-funded effort, targets 22 Indian languages too.

Here is what the subsidy means in practice, for a small team training on 16 GPUs for one week:

16 GPUs x 24 hours x 7 days = 2,688 GPU-hours
2,688 GPU-hours x Rs. 65 = Rs. 174,720

That is an affordable experiment for a funded startup, although a large model takes far more hours. Note the catch: the hours are subsidized, and building domestic chips is much slower work. The government's plan there is longer term, with processor design projects based on the open RISC-V architecture and 10 approved semiconductor manufacturing units, only some of them producing so far.

Japan: put money behind a national champion for machines

Japan took a more concentrated route. On June 30, 2026, the Ministry of Economy, Trade and Industry (METI) selected a group built around Noetora, a new company with SoftBank, Sony, NEC and Honda as its core members. It will work with the national research institute AIST, reports News On Japan.

• Money. About 390 billion yen in fiscal 2026, and about 1 trillion yen over five years, if plans hold.
• Goal. Core technology for physical AI, which lets robots and other machines operate on their own.
• Background. Late in 2025, Japan's Yomiuri Shimbun reported plans for a joint venture of about 10 companies aimed at a one-trillion-parameter model, with SoftBank discussing about 2 trillion yen for data centres. Those reports came before the selection, so treat the details with care, and note that the final focus emphasizes machines.

These figures come from press reports of METI's announcement.

Why machines? Honda and Sony among the core members hint at the reasoning: the bet plays to industries Japan already has, instead of a head-on chatbot race. It also leaves open whether the public gets a model it can actually use, since the Yomiuri described a model opened to Japanese companies rather than the general public.

Two approaches side by side

Currencies are not converted here because exchange rates and buying power differ too much for a neat comparison.

The dependency on foreign chips is the shared limit. An academic review of sovereign AI in Natural Language Processing calls chips the layer where countries inevitably cede sovereignty, noting that full self-sufficiency is hard even for the United States and China. Neither India nor Japan is free of this.

A fair question applies in both cases: are the public funds building lasting capability, or mostly subsidizing a few firms and foreign chip suppliers? It is too early to answer, especially for Japan, where the money has only begun to flow.

Can you try India's and Japan's sovereign AI work yourself?

For India, yes. Sarvam's weights can be downloaded, and the company documents how to run them with common tools such as Transformers and vLLM. They are large, so a rented GPU is more realistic than a laptop for the 105B model. Read the licence file before any commercial use, and test on your own material:

Answer this in Hindi, then give an English translation, then tell me which words you were least sure about:
Explain in three sentences how a small shop can keep customer records safely.

If you do not read Hindi, ask a speaker to judge the result. A polished answer you cannot check tells you little. If you know another language Sarvam supports, use that, and compare with a major assistant.

For Japan's Noetora, not yet. The group was selected only in June, and press reports describe no public model so far. Your practical option is to watch for a release, then look at who may use it and on what terms.