Sarvam’s $10-billion AI gamble faces the frontier test

“The gap is actually not as big as it looks. If we were to invest $10 billion today, we could reach the frontier by Independence Day next year,” Kumar said earlier this month. Such an investment would give Sarvam access to large-scale compute, advanced chips, infrastructure and the ability to hire talent. But catching up with models from OpenAI, Anthropic and Google involves more than assembling these resources, experts said.

“The challenge is therefore not merely capital, but research depth, compute availability, training efficiency and ability to attract a very small pool of world-class researchers,” according to Jaspreet Bindra, co-founder and CEO of AI&Beyond. Sarvam could plausibly get close to today’s frontier within a year, he said, but matching the frontier of that time would be considerably harder because global AI labs would have moved ahead.

This makes the availability of frontier-level researchers another critical variable. Pramod Gosavi, senior principal at Bay Area-based VC firm Blumberg Capital, said Sarvam would struggle to build a credible frontier model without experienced AI researchers leading the training effort. He pointed to the research depth at OpenAI, Anthropic and Google’s DeepMind and said Sarvam would need to attract researchers from abroad, as Chinese AI companies have done.

But not everyone sees talent as the primary constraint. Pratyush Choudhury, co-founder at early-stage AI VC fund Activate, said compute and data remain bigger limitations. “Once you have enough compute, you will attract the right talent with ease,” he said. Researchers, he added, are unlikely to stay if they lack the infrastructure needed to experiment and train increasingly large models.

Data itself is also becoming a more complicated advantage. For foundation models, it is not only the volume of material available on the Internet that matters, but also usage data showing how people interact with AI systems. Such data can be used to fine-tune models and improve their performance, Choudhury said.

The question, therefore, is whether Sarvam needs to win the general-purpose AI race at all. Experts said the company’s more defensible opportunity could lie in areas where global models remain less suited to Indian requirements. These include multilingual and voice applications, government and citizen services, regulated industries requiring data sovereignty, and enterprise systems that need to understand Indian documents, workflows and cultural contexts.

“I don’t think that Sarvam needs to ‘beat’ GPT or Gemini at everything,” Bindra said. Sarvam already works across 22 Indian languages and is expanding enterprise deployments beyond pilots. Kyndryl India CTO Sreekrishnan Venkateswaran said Sarvam and BharatGen models were increasingly being used by clients requiring Indic-language capabilities in sectors such as banking, financial services and insurance and telecom.

This could make sovereign AI a bigger commercial opportunity than trying to build the world’s best general-purpose model. “If Sarvam can make AI cheaper, open, sovereign, and multilingual, that is a defensible gap,” Bindra said.

Sarvam has also brought in Devendra Chaplot, a founding member of Mistral AI and Thinking Machines Lab, as an adviser for its planned trillion-parameter frontier model. The company’s biggest model is still months away, leaving the outcome uncertain.

The frontier race, however, remains relatively young. “A couple of years ago, we were completely confident that India couldn’t be a player in the AI foundation layer,” Rishi Bal, CEO of BharatGen, said. The narrative has now shifted to whether India can compete at the global frontier. For Sarvam, the bigger test will be whether $10 billion can buy not just the infrastructure to reach that frontier, but the research capability and data advantage needed to stay there.