IndiaAI Mission at 38,000 GPUs: India's AI Cloud Arrives

India's AI Compute Build-Out Reaches Scale

When India's government launched the IndiaAI Mission in early 2024 with a total budget of Rs 10,300 crore, the initial stated target under the Compute Capacity pillar was 10,000 publicly accessible GPUs. By mid-2026, the mission has deployed more than 38,000 GPUs — nearly four times the original target — accessible to registered Indian startups, academic researchers, and government agencies at a heavily subsidised rate of Rs 65 per GPU-hour. A third procurement tender is currently under execution, adding approximately 3,850 additional processing units including 1,050 Google Trillium TPUs — the first time non-NVIDIA AI accelerators have entered the IndiaAI infrastructure stack. The government's stated target is 100,000 publicly accessible GPUs by December 2026.

From 10,000 to 38,000: The Scale-Up

The IndiaAI Mission Compute Capacity programme's mandate was to create shared AI infrastructure accessible to Indian organisations that could not afford market rates for GPU cloud computing, which in 2024 ranged from approximately Rs 150 to Rs 200 per GPU-hour from commercial cloud providers. The subsidised rate of Rs 65 per GPU-hour represents approximately a 60 per cent discount to market pricing for H100-class GPU capacity, making training and inference workloads significantly more accessible for startups and research teams operating on constrained budgets.

The ramp from 10,000 to over 38,000 deployed GPUs happened through two main procurement rounds. The first round allocated capacity across providers including Yotta Data Services, L&T, and E2E Networks, distributed across data centres in multiple Indian cities. The second round, announced by Union Minister Ashwini Vaishnaw at the AI Impact Summit in February 2026, added 20,000 more processing units and introduced a target of reaching 54,000 GPUs in the near term before scaling to 100,000 by December 2026.

Yotta Shakti Cloud: Blackwell Ultra at Scale

The single largest infrastructure deployment within the IndiaAI compute programme is Yotta Data Services' Shakti Cloud, built on more than 20,000 NVIDIA Blackwell Ultra GPUs. Blackwell Ultra is NVIDIA's current data centre GPU generation, designed for large-scale training runs and high-throughput inference on models above 70 billion parameters. It offers substantially higher memory bandwidth and flops per unit than the H100 generation, and supports NVIDIA's NVLink interconnect for multi-GPU training at scale.

The Shakti Cloud deployment makes India one of a small number of countries with significant Blackwell Ultra capacity in a nationally accessible compute programme — comparable in ambition, if not yet in absolute scale, to the AI compute investments made by the United States, EU, and UAE through their respective national AI programmes. For teams building frontier-scale models for Indian language, agricultural, or healthcare applications, Blackwell Ultra access at Rs 65 per hour creates training economics that were simply not feasible when the IndiaAI Mission launched.

The Third Tender: Google Trillium TPUs Enter the Stack

The third IndiaAI procurement tender is notable for one architectural departure from the previous two rounds: the inclusion of 1,050 Google Trillium TPUs alongside approximately 2,800 additional GPU units from NVIDIA and other suppliers. This is the first time the IndiaAI Compute programme has included non-NVIDIA AI accelerators.

Trillium is Google's sixth-generation tensor processing unit, optimised for inference workloads and transformer-based model architectures. Its per-unit economics for large-batch inference on models in the 7B to 70B parameter range are competitive with H100-class GPUs, and in some configurations — particularly sustained high-throughput inference rather than training — Trillium offers lower cost per token. The inclusion of Trillium in the IndiaAI stack signals that the mission is diversifying its accelerator portfolio rather than depending entirely on the NVIDIA supply chain, which has faced constrained availability globally since 2023.

For teams building Indian-language inference pipelines, Trillium-class compute within the IndiaAI programme could offer a meaningfully different cost structure from GPU-based inference — worth evaluating once the third tender capacity becomes accessible.

Who Can Access This Compute and at What Cost

The IndiaAI Compute programme is open to three categories of users: DPIIT-registered Indian startups, academic and research institutions, and government agencies. Access requires registration at indiaai.gov.in and submission of a compute proposal describing the intended workload. Allocation is reviewed on a rolling basis. Access is not available to large commercial technology companies or to non-Indian entities.

The subsidised rate of Rs 65 per GPU-hour translates to approximately 72 US cents at current exchange rates. A 100-GPU training run over 24 hours costs approximately Rs 156,000 — roughly $1,800 — compared to approximately Rs 360,000 to Rs 480,000 at unsubsidised commercial rates for equivalent hardware. For teams training models in the 7B to 70B parameter range, the IndiaAI rate makes fine-tuning on domain-specific Indian data economically viable in a way that was previously out of reach for early-stage startups.

What This Means for Indian AI Teams

For Indian startups and research teams building AI products for the domestic market, the IndiaAI Compute programme changes the economics of model development in two ways.

First, the direct cost of training and fine-tuning models is substantially lower than commercial alternatives for qualifying organisations. A startup building a multilingual customer service model, a healthcare AI, or an agricultural advisory tool for Indian farmers can now run training experiments and production fine-tuning at a fraction of the cost of AWS, Azure, or Google Cloud at list prices.

Second, running training workloads on Indian infrastructure — rather than routing through foreign cloud providers — addresses the data localisation and sovereignty considerations increasingly relevant under India's Digital Personal Data Protection framework. For teams processing sensitive health data, financial records, or personal information of Indian users in model training pipelines, running those workloads on IndiaAI compute rather than a foreign cloud simplifies the compliance picture.

The Government's target of 100,000 GPUs by December 2026 — including the diversified hardware mix introduced by the Trillium tender — suggests that India's publicly accessible AI compute will continue to expand meaningfully through the rest of the year.

The Bottom Line

India's IndiaAI Mission has scaled from its initial 10,000-GPU target to more than 38,000 deployed GPUs by mid-2026, accessible at Rs 65 per GPU-hour to registered Indian startups, researchers, and government agencies — roughly 60 per cent below commercial cloud rates. The Yotta Shakti Cloud, powered by over 20,000 NVIDIA Blackwell Ultra GPUs, is the programme's flagship infrastructure deployment. A third tender adding approximately 3,850 units — including 1,050 Google Trillium TPUs in the first instance of non-NVIDIA accelerators in the IndiaAI stack — is currently under execution. For Indian startups building AI for India-specific use cases, IndiaAI compute is now the most cost-effective, legally compliant infrastructure option available for training and fine-tuning frontier-scale models.