India just produced its first sovereign AI unicorn, and the investor list reads like a who’s who of enterprise AI infrastructure: Nvidia, HCLTech, Bessemer Venture Partners, Khosla Ventures, and Peak XV. Sarvam AI closed its $300 million Series B at a $1.5 billion valuation this week, with Nvidia completing a $74 million second tranche alongside Glade Brook Capital, Gaja Capital, and Activate. For enterprise AI leaders watching the sovereign AI space, this is not just a funding story. It is a blueprint.
What Sarvam Built and How
Sarvam’s pitch is deceptively simple: India needs AI that works in Indian languages, runs on Indian infrastructure, and stays under Indian data sovereignty. Executing on that requires the full stack, not just a fine-tune of someone else’s model.
The company has trained two foundation models from scratch entirely in India, using compute provided under the government’s IndiaAI Mission:
Sarvam 105B is a Mixture-of-Experts model with 105 billion total parameters and approximately 10.3 billion active parameters per query. Context window is 128K tokens. It runs on H100 infrastructure and is optimized with custom MLA (Multi-head Latent Attention) kernels, vocabulary parallelism, and disaggregated serving.
Sarvam 30B is the edge-optimized sibling: 32 billion total parameters, 2.4 billion active, 65K context window, designed to run on consumer GPU hardware and power real-time conversational deployments.
Both models carry an Apache 2.0 license, meaning enterprises can fine-tune, deploy, and redistribute commercially without seeking further permission. Both are on Hugging Face.
Benchmarks: Where Sarvam Competes and Where It Does Not
The benchmark picture is honest. Sarvam 105B wins on specific agentic tasks but does not beat every frontier model on every measure.
| Benchmark | Sarvam 105B | DeepSeek R1 | Gemini 2.5 Flash | o4-mini |
|---|---|---|---|---|
| Math500 | 98.6 | 97.5 | 72.0 | 97.6 |
| AIME25 (w/ tools) | 88.3 (96.7) | 87.5 | 72.0 | 92.7 |
| GPQA Diamond | 78.7 | 81.0 | 82.8 | 81.4 |
| Live Code Bench v6 | 71.7 | 73.3 | 61.9 | 80.2 |
| MMLU Pro | 81.7 | 85.0 | 82.0 | 81.9 |
| BrowseComp | 49.5 | 3.2 | 20.0 | 28.3 |
| Tau2 Bench (avg.) | 68.3 | 62.0 | 49.7 | 65.9 |
The standouts are BrowseComp (49.5 vs. 3.2 for DeepSeek R1, indicating strong grounded web search capability) and Tau2 Bench (68.3, the highest among compared models), which measures long-horizon agentic reasoning and task completion. These are the benchmarks that matter most for enterprise agentic workflows.
The weaknesses are real: TerminalBench Hard scores are low (1.5% for 105B), and hallucination rates are higher than frontier models with much larger training budgets. Models trained with Nvidia Blackwell clusters and 10x more compute will tend to outperform on breadth. Sarvam’s leadership in agentic web tasks and multi-step reasoning is therefore genuinely competitive, not merely an artifact of cherry-picking.
The Investor Thesis
Nvidia’s equity participation tells the structural story more than any press release. Nvidia supplied the H100 GPUs, contributed NeMo libraries and Nemotron datasets, and now owns a slice of the outcome. That alignment of interests turns an infrastructure sale into an ongoing partnership: Sarvam gets early access to Blackwell capacity (plans to scale from 2,000 to 10,000 Blackwell GPUs), and Nvidia gets a showcase deployment that near-linearly scaled training across 4,096+ H100s.
HCLTech’s $150 million lead investment is the enterprise distribution angle. HCLTech brings 200,000-plus delivery engineers and client relationships across the Fortune 2000. The company’s sovereign data center in Odisha provides the physical infrastructure for on-premise enterprise deployments in regulated sectors. Combined with Sarvam’s model stack, this is an integrated offer: foundation model plus forward-deployed engineers plus sovereign compute, all under one roof.
Bessemer, Khosla, Peak XV, and Lightspeed fill out the cap table with traditional venture capital. The round’s structure, with a strategic anchor from the country’s largest IT services company and a compute anchor from the world’s dominant AI chipmaker, is not accidental. It reflects a model for sovereign AI funding that other nations are watching closely.
Enterprise Deployments Today
The production deployments matter as much as the benchmarks. Sarvam is not a research lab waiting for scale. It is already running at scale:
The company processes over 2 million AI interactions per day across its platform, with API call volume tripling over the prior three months to roughly 10 million daily calls. Nearly 1 million developers have used the platform.
Tata Capital and Infosys use Sarvam agents for KYC, sales enablement, and customer support workflows delivered over telephony and WhatsApp in local Indian languages.
UIDAI, India’s national digital identity authority, uses Sarvam for Aadhaar, delivering real-time voice-based identity verification and fraud alerts in local languages to India’s entire population of 1.4 billion.
One leading fintech deployed a sales enablement platform for a 350,000-person salesforce using Sarvam models, reporting 3x performance gains.
Sarvam Vision, the company’s document intelligence model, is digitizing more than 35 million insurance forms and legacy land records. The speech pipeline transcribes more than 500,000 hours of audio monthly.
What This Means for Enterprise AI Strategy
Pratyush Kumar, co-founder of Sarvam, framed the sovereign argument plainly in a CNBC interview: “India absolutely needs to have an alternative. It’s not clear when export controls may come in. Today, there are open Chinese models. Maybe next year there aren’t.”
That framing applies beyond India. Any enterprise or government operating under data residency constraints, export control risk, or multi-language requirements is exposed to the same fragility: dependence on a small number of non-local foundation model providers. If those providers restrict access, change pricing significantly, or become unavailable due to geopolitical events, the enterprise AI stack breaks.
Sarvam’s approach addresses that risk through three choices that every enterprise team building a sovereign strategy should study. It also reinforces a pattern visible in the Microsoft and Mistral sovereign AI partnership and the Korean government’s backing of SKT’s A.X K2: large enterprises and governments are no longer content to rely entirely on US frontier labs for mission-critical AI infrastructure.
Train from scratch, not just fine-tune. The IndiaAI Mission compute subsidy made this economically viable. For enterprises without national backing, the equivalent is a strong cloud partner deal or consortium compute arrangement. The Sarvam models released under Apache 2.0 are a starting point for teams that want to train derivatives.
Pair open weights with forward-deployed humans. The HCLTech partnership is not just distribution. It is the implementation layer that moves models from benchmarks to production workflows. Enterprises buying foundation models without equivalent implementation depth tend to plateau at pilot stage, a pattern Enera covers in depth in our analysis of the enterprise AI deployment gap.
Target the multimodal and agentic layer from day one. Sarvam’s BrowseComp leadership and Tau2 results suggest deliberate investment in agentic tasks rather than pure language benchmarks. Vision, speech, and long-context document processing are all part of the initial release. This reflects where enterprise value actually sits: not in chat, but in automated workflows that cross modalities.
The Road Ahead
Sarvam has opened a San Francisco office and a Bay Area research lab, signaling that the next generation of models will be built with US research talent alongside the India-based team. An IBM partnership is in place. A 1-trillion-parameter model is explicitly on the roadmap.
The IndiaAI Mission has already designated Sarvam as the developer of India’s sovereign foundation model, a commitment that provides both guaranteed compute and a government deployment channel spanning defense, public services, and regulated industries.
For enterprise AI leaders, the signal from this round is not that India built a good model. It is that sovereign AI is fundable, deployable, and competitive on the benchmarks that matter for enterprise agentic workflows. The Sarvam playbook, full-stack training, government compute access, open-weight Apache 2.0 release, strategic enterprise distribution partner, and Nvidia infrastructure anchor, is a replicable template that is already being studied by AI programs in the EU, Southeast Asia, and the Middle East.
Enterprises weighing open-weight model strategy have a new high-water mark to benchmark against. And enterprises navigating export control and data sovereignty questions have their most credible case study yet that building locally competitive AI is achievable without waiting for a US or Chinese frontier lab to solve it for you.
Sarvam’s primary investor announcement and model benchmarks are available at the Sarvam Series B post and the Sarvam 105B/30B model release. Nvidia’s technical partnership overview is published in the NVIDIA case study.