Sarvam 105B trained on about 4,000 GPUs. The trillion-parameter model is being built towards a 10,000-GPU cluster.
BENGALURU: Sarvam AI, India’s newest AI unicorn at a $1.5 billion valuation, plans to release a foundation model with more than one trillion parameters, trained from scratch in India, by roughly February 2027. The company announced the plan at its Epoch 2026 conference on 30 July and co-founder Pratyush Kumar restated the target this month. Encouragingly, the ambition rests on working products, a strong funding base and state-backed compute, although the date remains a company goal.
A founding idea with historical weight
Kumar’s argument is rooted in history. India met the steam engine, steel making and the early internet largely as a user and he says it lost out on key value creation as a result. Artificial intelligence, he argues, offers a chance to change that pattern, because the intelligence layer is the main value engine of the modern economy. Renting foreign models works for now, but India should build its own to own its destiny and become a rule maker. Co-founder Vivek Raghavan compares AI’s strategic weight to that of civil nuclear power.
Sarvam defines sovereignty in three parts. These are training and inference in Indian data centres, models pre-trained from first principles and domestic control over auditing and governing the weights.
Built on strong foundations
Notably, the team brings rare credentials. Raghavan helped design Aadhaar at the UIDAI, while Kumar co-founded AI4Bharat at IIT Madras after research stints at IBM and Microsoft. In April 2025, the Ministry of Electronics and Information Technology selected Sarvam from 67 applicants under the ₹10,000 crore IndiaAI Mission. Subsequently, the company delivered Sarvam 105B, pre-trained on nearly 18 trillion tokens using about 4,000 NVIDIA H100 GPUs from the mission’s allocation of 4,096. In June 2026, HCLTech led a $234 million Series B first close, investing $150 million for a 10.46% stake.
Economics that favour Indian builders
Tokenisation gives the strongest practical case. According to Sarvam, Western-trained models can need four to eight times as many tokens as English for the same Indic sentence, while its own tokenisers reach 1.4 to 2.1 tokens per word. Consequently, costs and latency fall for Indian enterprises. Sarvam 105B sells at $0.80 per million blended tokens, against $4.50 for OpenAI’s GPT-5.4 Mini and $9.00 for Google’s Gemini 3.5 Flash, both hosted in the United States. Sarvam also cites competitive benchmark scores, though independent verification is still pending.
Impact already visible
Meanwhile, the products are working at scale. In Odisha, Sarvam’s vision platform lifted accuracy on historic handwritten land records from about 30% to over 60%, with human review driving further gains. Its voice stack, whose text-to-speech engine supports all 22 scheduled Indian languages, powers voice ordering for Swiggy and handles over 450 million calls a year. The company targets 1 billion voice minutes by March. Furthermore, its enterprise agent platform supports a 350,000-person sales force at a major fintech, and its models serve applications reaching an estimated 17 million farmers and 45 million insurance policyholders.
The trillion-parameter leap
Against this backdrop, the next model aims higher. It targets coding, cybersecurity, simulation and scientific research, and it is being trained from scratch on Indian soil. Compute is scaling accordingly, from about 2,000 Blackwell GPUs today towards a dedicated 10,000-GPU cluster built with HCLTech in Odisha. That cluster would be a clear step up from the 4,000 H100s used for Sarvam 105B.
Sarvam also opened a San Francisco office to draw experienced engineers back into India’s AI ecosystem, and it appointed Devendra Singh Chaplot, formerly of Mistral AI, as Advisor. Leadership says development remains on track, and private previews of its Epoch Builder Edition have run since August. The company claims an optimisation layer can speed inference up to 15x on selected workloads, though third-party benchmarks are pending.
A healthy debate and what to watch
Naturally, not everyone shares the premise. Critics argue that locally hosted open-weight models can be cheaper, and Tech Mahindra executives have questioned whether generic large language models are the right national target. Sarvam takes a pragmatic view, since its hosted inference service runs Sarvam 105B alongside selected open-weight models on Indian servers.
The next few months will sharpen the picture. Key markers include confirmation that the 10,000-GPU cluster is live in India, a technical model card on the architecture and independent results on benchmarks such as SWE-bench and HumanEval. Sarvam has not yet said whether the model will carry open weights or ship through hosted APIs. Ultimately, February will show how far this foundation can carry India’s own intelligence layer.
About the Author
