India’s AI Dream Falls Short of a Million GPUs

The global AI race ultimately boils down to how much compute labs own.

While India’s flagship IndiaAI Mission has deployed 38,000 high-end GPUs to democratise access for startups, researchers, and students, leading Western companies, particularly AI giants like OpenAI, xAI, and Anthropic, command vastly larger private clusters numbering in the hundreds of thousands to millions. 

The question then arises whether India has enough compute to build frontier models or even run inference at scale to serve its customers. For comparison, GPT-4, which was launched back in March 2023, was trained using 25,000 NVIDIA A100 GPUs. 

To address these constraints, the government is ramping up GPU deployment. In February, The Economic Times reported that the government is set to add nearly 25,000 GPUs. According to the report, nine companies have cleared the technical stage of the fourth round of the IndiaAI Mission’s GPU tender.

Among them, Yotta has proposed offering 17,000 NVIDIA B300 GPUs. Cyfuture India has bid for 1,024 NVIDIA B200 GPUs in this round. Meanwhile, RackBank Datacenters has proposed a mix of 1,024 NVIDIA B300, B200, and H200 GPUs.

As of February, of the total committed capacity, 22,787 GPUs, or 56.21 %, have already been allocated through the IndiaAI Compute Portal. 

At the IndiaAI Impact Summit, startups like Sarvam AI and BharatGen launched indigenous models trained on these GPUs.

Sarvam AI, which launched its Indus chatbot, reportedly has access to around 4,096 NVIDIA H100 GPUs for model development, while BharatGen has a compute footprint of over 13,000 GPUs. 

The Stark Difference

The compute difference between India and the West is hard to ignore. OpenAI CEO Sam Altman last year said that the company expects to have over one million GPUs online.

Overall, the ChatGPT creator has committed over $1 trillion in compute deals, including roughly $250 billion with Microsoft as its primary cloud partner, $300 billion with Oracle for Stargate data centre, $38 billion with AWS cloud, $22.4 billion with CoreWeave, and a $100 billion commitment from NVIDIA in the form of GPUs.

Amazon said that OpenAI will consume 2 gigawatts of its custom chip, Trainium capacity, through AWS infrastructure to support Stateful Runtime Environment, Frontier, and other advanced workloads.

OpenAI is reportedly developing a new model, internally codenamed Spud, though the company has not made any official announcement. Notably, the AI giant recently shut down its video generation model, Sora, to prioritise compute resources for core, revenue-generating products.

Similarly, Anthropic is among the most heavily resourced AI labs globally. The company relies on a multi-cloud setup spanning Google’s TPUs, Amazon’s Trainium chips, and NVIDIA GPUs.

Most recently, the AI giant secured roughly 3.5 gigawatts of next-generation TPU capacity through its partnerships with Google and Broadcom, with deployments expected from 2027 through 2031. Last year, Anthropic announced that it would deploy one million Google TPUs. Now, it has already trained its most powerful model yet, Mythos. 

Meanwhile, according to a report by Epoch AI, Google accounts for about one quarter of global cumulative AI compute capacity as of Q4 2025, making it the largest single owner. Notably, the company relies primarily on its in-house TPU chips rather than GPUs from NVIDIA.

At the same time, Elon Musk’s xAI is estimated to have over 100,000 GPUs in operation, with newer reports suggesting its Colossus cluster has expanded to around 200,000 GPUs, which the company used to run reinforcement learning for Grok 4.

 In January, xAI said it was expanding its compute lead with its Colossus I and II supercomputers, targeting over one million H100 GPU equivalents by the end of 2025, and added that Grok-5 is currently in training. xAI’s Colossus 2 cluster is training seven models, ranging from one trillion parameters to ten trillion ones, including a new video model. 

More GPUs Coming 

Rahul Agarwalla, Managing Partner at SenseAI Ventures, tells AIM that frontier models are now trained on clusters of around 100,000 GPUs. 

“Whether it is xAI, Meta, or OpenAI, all of them have larger clusters on which they’re doing training. Each of these companies has more GPUs than our entire IndiaAI mission,” he says. 

According to SenseAI Ventures’ State of AI Report 2026, in India, around 78% of funding is directed toward application-based AI startups, while foundational AI research accounts for just 2%.

However, efforts are being made to close the gap. 

Indian neo cloud startup Yotta announced earlier this year that it would deploy 20,736 liquid-cooled NVIDIA Blackwell Ultra GPUs, which are expected to go live by August 2026.

Similarly, Neysa has also signed a deal with NVIDIA to bring in 20,000 GPUs. The company expects NVIDIA’s latest Blackwell GPUs, including the B300s, to make up a major part of the expansion.

“Our Blackwells are already on the way. The IndiaAI Mission is a source of utilisation for us as well. We need those GPUs as quickly as possible to meet a lot of those requirements,” Karan Kirpalani, Chief Product Officer at Neysa, told AIM in an earlier interaction.

Besides the neo cloud startups, several other data centre players are emerging in India.

Bharti Airtel recently announced a $1 billion investment in its data centre subsidiary Nxtra Data Limited, led by Alpha Wave Global, along with existing investors Carlyle and Anchorage Capital. The proceeds will be used to expand Nxtra’s data centre network across India and support growing demand for AI and cloud infrastructure. The company plans to scale capacity from about 300 MW to 1 GW in the coming years.

The Adani Group is also stepping up when it comes to investments in AI infrastructure, outlining plans to pump up to $100 billion into building renewable-powered, AI-ready data centres by 2035. The group is simultaneously exploring partnerships with global technology firms such as Meta and Google, while expanding its operations in India with new data centre projects, including a planned facility in Bhubaneswar.

For India, the IndiaAI Mission is a foundational move toward self-reliance. Yet, without accelerated GPU access and sustained multi-billion-dollar investments, India risks remaining a fast follower rather than a leader in frontier AI.