NVIDIA Highlights Sarvam AI’s India-Built 100-Billion-Parameter AI Models Running on H100 GPUs

NVIDIA has publicly championed Sarvam AI, a Bengaluru-based artificial intelligence company, as a real-world example of what the chipmaker calls “sovereign AI at population scale.” The promotion came via NVIDIA’s official account on X, formerly Twitter, where the company posted that the concept “isn’t theory anymore, it’s shipping” — pointing to Sarvam AI’s work as the evidence.

The post describes Sarvam AI as building a “full-stack, ‘Made in India’ AI platform” that trains models with over 100 billion parameters across a cluster of more than 4,096 NVIDIA H100 Tensor Core GPUs. It also claims the platform “delivers millisecond-level, multilingual voice” services. These figures originate from NVIDIA’s marketing communication and have not been independently verified by an official standards body or peer-reviewed technical publication.

What Sarvam AI Is Actually Building

Sarvam AI describes itself as India’s full-stack sovereign AI platform, with a focus on building frontier-class models for population-scale use across India’s 22 officially recognised languages. Its product range spans speech-to-text, text-to-speech, translation, document digitisation, and conversational agents — all targeting Indic languages and the accents, scripts, and cultural contexts that go with them.

Co-founder Vivek Raghavan has described two distinct large language models in public interviews. The first is a 30-billion-parameter model optimised for conversational tasks in Indian languages. The second is a 105-billion-parameter model built for reasoning, function calling, and web-search-related tasks. That larger model is consistent in scale with the “100B+ parameter” language in NVIDIA’s post, though the precise mixture-of-experts architecture — where different sub-networks are selectively activated per input, enabling large parameter counts with relatively efficient compute — has not been laid out in formal public technical documentation.

Vivek Raghavan, co-founder of Sarvam AI, said: “We are building for India’s scale, India’s languages, and India’s needs — and we want to make sure that the infrastructure, the models, and the governance all sit within India.”

Sarvam AI reports raising around £32 million to date, according to venture and startup tracking databases, with its most recent round recorded as a Series A. That figure comes from market sources rather than an official government body, so should be treated accordingly.

The “Sovereign AI” Question

The term “sovereign AI” has become a fixture in both industry and policy discussions. It refers broadly to AI infrastructure, models, and data that are developed, hosted, and governed within a specific country — the idea being that a nation retains control over capabilities that are increasingly central to its economy and public services.

NVIDIA has been pushing this framing hard. The company positions its GPU clusters, and above all the H100, as the hardware foundation countries need to build domestic AI capacity. India’s central government has separately announced initiatives to support domestic AI development and digital public infrastructure, and Sarvam AI has aligned itself closely with that national agenda.

But the sovereign AI label has its critics. Some analysts argue that genuine sovereignty is difficult to claim when the underlying hardware — in this case, H100 GPUs designed and controlled by a US company — is itself a foreign dependency. Full technological autonomy, this argument goes, would require domestic chip design and fabrication alongside domestic models. It’s a tension NVIDIA’s own marketing doesn’t address.

Civil-society and digital-rights groups have also raised broader concerns about large-scale national AI deployments: how training data is collected, who audits it, and what safeguards exist against surveillance or bias at population scale. Sarvam AI has not published detailed responses to those questions in widely available public documentation.

Scale, Claims, and What Remains Unverified

The specific figures in NVIDIA’s post — 4,096-plus H100 GPUs, 100-billion-plus parameters, millisecond-level latency — are marketing claims from a corporate social media account. No independent benchmarking from an official standards body has been published to substantiate the latency figures. The GPU count and efficiency claims similarly await third-party verification.

That doesn’t make them wrong. Clusters of thousands of H100 GPUs are entirely standard at frontier AI labs and major cloud providers. And the parameter counts Raghavan has described in interviews are consistent with what NVIDIA’s post says. But the absence of peer-reviewed papers or external audits means these numbers sit in the category of company and industry disclosure rather than confirmed fact.

Sarvam AI says its APIs and tools are available to governments, enterprises, and developers wanting to integrate Indic language capabilities into their own services. The company has positioned its pricing as competitive with global providers, chiefly for Indian-language speech recognition and synthesis — an argument aimed squarely at public-sector and mass-market adoption within India.

What This Means for Kent Residents

Sarvam AI operates out of Bengaluru and targets Indian-language use cases; there is no direct operational presence in Kent, and no local public body — including Kent County Council, NHS Kent and Medway, or Kent Police — has announced any use of its products. The more relevant angle for Kent residents is the broader competitive pressure that platforms like Sarvam AI place on global AI providers: more domestic alternatives in major markets can influence pricing and capability improvements in the multinational tools that UK consumers and businesses do use day to day. For Kent residents or local organisations serving communities that include Indian-language speakers, improved multilingual AI tools — whether from Sarvam AI directly via API or indirectly through products that licence such capabilities — could eventually improve services in health communication, education, or customer support.