AI firms getting GPU sops may see govt at the table – Crypto News

Four startups–Bengaluru-based Sarvam and Gnani.ai, Gurugram’s Soket AI Labs, and Noida’s Gan AI–have been shortlisted by New Delhi to use government-licensed GPUs. 

The proposed move, which signals a shift from a purely subsidy-driven approach to one that seeks the Indian government’s long-term participation in the country’s AI ecosystem, has sparked debate among founders and investors. While startup founders are happy with the assurance that comes with the government’s skin in the game, investors raised concerns about conflict of interest. 

A senior government official told Mint, requesting anonymity, that taking equity in the startups “is one of the many methods through which the government is considering rolling out the GPU access”. The ministry is currently in process of vetting these methods, and a final decision on how the funding will take place may be announced in the coming months.

“An equity stake model does not have to only be through sovereign funds—under National Quantum Mission, too, startup funding is being offered in exchange for equity stakes,” the official said. 

Also read | Sovereign silicon: India targets indigenous 2nm, Nvidia-level GPU by 2030

To be sure, the government earlier this year announced the $1.2-billion AI Mission to offer a subsidized approach to access to GPUs, which are the lifeblood of AI firms. Reason: buying GPUs is expensive–Nvidia’s chips cost up to $60,000 for a single GPU system–and beyond the capabilities of India’s fledgling AI startups. 

So, the government has roped in Hiranandani Group’s Yotta Data Services, Tata Communications and others to offer access to GPUs at their data centres at about one-fourth of the market rate, or as low as $0.8 per GPU per hour.

Mint’s email to Meity seeking comment from an official spokesperson did not receive responses until press time.

Opinions divided

Some investors feel the government’s move may create a conflict with both private investors as well as taxpayers. Harshjit Sethi, managing director of venture capital firm Peak XV, said that as a result of the Centre’s equity stake in these startups, “the return to the taxpayer will come from the appreciation of this equity stake, in addition to the strategic imperative that we as India should have our own AI models from scratch”. Peak XV led Sarvam’s seed funding round, and co-led its $41-million series-A funding round in December 2023.

Such a move could create conflict for private investors, who seek profitability in exchange for large-scale investments. Three other VC partners investing in deep-tech startups told Mint that such a move may raise questions of conflict, at least in the near-term.

Read this | Home-made chips: India’s GPUs set for year-end trials

“It’s important that the government’s involvement is structured in a way that doesn’t limit startups’ ability to pursue global commercial opportunities,” said Vishesh Rajaram, managing partner of early-stage deep-tech-only VC fund Speciale Invest. However, Rajaram added that “if done right, the government can play the role of a catalytic capital provider—nudging innovation towards public good”.

“The question would then be, should you seek to raise a large funding round and pursue a venture that gives you closer control of the venture, or do you build for public equities and scale up your business after that?” a senior industry executive and investor at an early-stage tech-focused VC fund said. “This is a question that will need to be answered by the founders, in order to judge if a government-owned stake could raise conflicts.” 

For others, though, the decision is to favour access to these pricey GPUs rather than protecting their stakes.

Abhishek Upperwal, founder and chief executive of Soket AI Labs, said that the startup is leaning towards offering an equity stake to the Centre for access to 2,000 GPUs to train its 120-billion-parameter large language model (LLM). “The rationale is that if the government is a stakeholder in the business, this allows us to reduce our dependence on venture capital funding—since in exchange for a minority stake, we’re getting to reduce our operating costs significantly.”

However, this could come at the cost of focusing on public utilities, rather than a commercial business model in the near term. “The government owning a stake is to largely ensure that public funds are used to train AI models for public utilities—here, the Centre will play the role of a mentor, and not a client,” Upperwal said.

Also read | No GPUs, no problem: Ziroh Labs can run AI models just with CPUs

Access to ample funds is what startups believe could push them to find the government’s ask of equity stakes reasonable. Industry estimates said that access to 2,000 GPUs for one year, which Upperwal’s Soket AI Labs has requested for, will amount to $2.5-10 million for one year depending on consumption, even at Meity’s subsidized procurement rates. Barring Sarvam, neither of the other three startups have raised such funding yet.

“Access to high capital funding from early-stage VCs in India is difficult—which is why we’re happy with an equity stake model with the government. It also ensures that we’re a part of the effort that the Centre is putting to develop the overall ecosystem, and we’ll play a part in this in the long run. We’ve been looking to raise funding for two years, but building a new foundational AI architecture requires lengthy research and development (R&D) investments—which most VCs in India are not keen toward,” Soket’s Upperwal said.

Industry veterans said that citizens will also likely need to find value in the AI applications in question, for the Centre’s equity investments to qualify as a justified use of public taxpayer’s money. “Questions would arise on the use of public funds for startup investments. Globally, sovereign funds invest in commercial ventures, and use profits for public benefits and utilities. In this case, the use of public funds would come in building India’s own AI ecosystem,” said Kashyap Kompella, AI analyst and author.

And read | Nvidia’s Blackwell GPU may reach Indian shores as early as October