The India AI Mission is currently embroiled in a “computing power crisis.” As a core initiative to drive the development of domestic AI technology and compete for global AI influence, this plan has been severely hindered by significant delays in GPU delivery, leaving the participating startups in a development bind. These startups, which were expected to accelerate foundational model development with ample computing resources, now face the embarrassing situation of being unable to proceed due to a lack of essential resources.
This GPU delivery disruption has not only hindered the progress of individual projects but also reflects multiple obstacles India must overcome in its pursuit of AI technological sovereignty. From international supply chain constraints to domestic resource allocation issues, each bottleneck is testing the Indian government and enterprises’ adaptability and execution capabilities.
Development Divide Amid the Computing Power Shortage
Among the first batch of four startups selected by the Indian government to build sovereign foundational models, Sarvam stands out as the only company that has made some progress amidst the crisis. According to the original plan, the company was to receive 4096 NVIDIA H100 GPUs within six months, with a total computing cost of 2.4671 billion Indian Rupees. Despite only receiving an initial batch of 1500 GPUs, with the remaining 2596 still pending, Sarvam has managed to launch models like Sarvam-M and Sarvam Translate. They also plan to open-source their IndiaAI model once permitted, making them particularly notable in an otherwise stagnant environment.
However, other companies are facing far greater difficulties. Soket AI Labs, Gnani.ai, and Gan.ai have yet to receive any GPUs or financial support, and their projects have nearly come to a halt. Soket is working on the EKA project, which aims to develop a 120 billion parameter language model. Due to a lack of sufficient computing resources, they can only rely on existing GPUs to build small models slowly. CEO Abhishek Upperwal admits that before achieving large-scale computation, the team can only “do their best” to avoid wasting time. Similarly, Gnani.ai’s 14 billion parameter voice AI model and Gan.ai’s video creation tool have also stalled due to insufficient computing power. These models were originally planned to be developed within 6 to 10 months, but now the time frame has been significantly shortened, and if GPUs are not acquired soon, the development cycle may be delayed, severely affecting the open-source strategy.
Multiple Factors Contributing to Delivery Disruptions
The delay in GPU delivery has various causes, with U.S. export controls being a primary factor. The AI chip ban, effective in January 2025, places India on a “second-tier country” list, requiring special permits for importing H100 GPUs. The ban limits India to importing a total of 50,000 H100 equivalents between 2025 and 2027. Although India can apply for “National Verification End-User” (NVEU) status to receive up to 320,000 GPUs, the complex approval process means no company has successfully been granted approval, directly limiting access to high-end GPUs, on which Indian companies are heavily reliant.
In addition, supply chain and logistics bottlenecks have worsened the delivery issues. Even with approval, the actual delivery of GPUs remains difficult. For example, Yotta, a company planning to expand its GPU inventory to 32,768 units by the end of 2025, has ordered 16,000 H100s, but production delays and international logistics issues have caused its delivery schedule to fall behind. Furthermore, stringent customs inspections for imported chips have further slowed down the delivery process.
Moreover, there are significant issues with the government’s execution efficiency and resource allocation. While the IndiaAI plan promised to provide computing resources to startups, there are flaws in the funding and GPU allocation mechanisms. Sarvam’s subsidy of 986.8 million Rupees only covers 40% of its total computing costs, and other companies have not received any funding. The government’s creation of a 34,000 GPU backbone network has added 15,000 GPUs, but the allocation priorities are unclear, making it difficult for startups to quickly access the resources.
Solutions and Future Challenges
In response to the GPU delivery crisis, both the Indian government and businesses are actively exploring alternatives. Some companies are attempting to reduce their reliance on GPUs by optimizing algorithms. For instance, Ziroh Labs has developed the KompactAI system, which allows large models like Llama2 to run on regular CPUs, achieving a 3x improvement in inference speed and an 80% reduction in power consumption. Soket AI Labs is also exploring a phased training strategy, starting with smaller models using existing GPUs and gradually expanding to models with hundreds of billions of parameters.
In terms of international collaboration and localized procurement, the Indian government is negotiating with NVIDIA to procure GPUs under a “leasing and sub-leasing” model, and is working with domestic companies like Cyfuture and Netmagic to expand GPU inventories. Yotta’s collaboration with NVIDIA to build the GIFT-City AI data center, set to open in March 2025, will deploy 16,000 H100 and GH200 GPUs, which is expected to alleviate some of the computing power pressure.
However, the future of India’s AI mission still faces numerous long-term challenges. Talent drain is a significant issue, with top researchers leaving for tech companies in the U.S. and Europe. Domestic universities lack research funding, and basic research is weak. Private companies are more inclined to invest in short-term outsourcing businesses. Geopolitical risks related to the supply chain continue to persist, and procuring GPUs via third-party countries will increase costs by 30% to 50%. Global chip shortages and logistics fluctuations also contribute to uncertainty. Additionally, Chinese open-source models are performing strongly in the Indian market, and the local AI application ecosystem remains weak. If India cannot break through the computing power bottleneck within the next 1 to 2 years, it risks missing the key window in global AI competition.
The GPU delivery crisis within India’s AI mission is the result of multiple factors, including international supply chains, geopolitical issues, and domestic policy execution challenges. Although various responses have been initiated, solving the problem in the short term remains difficult. To achieve the goal of “technological sovereignty,” India still has a long road ahead.
