The IndiaAI compute mission follows a bidding process where empanelled vendors must match the lowest bid price for the AI compute and services. Photo: indiaai.gov.in
The Ministry of Electronics and Information Technology announced the launch of a continuous empanelment process for AI compute providers, which allows firms to apply on an ongoing basis to supply AI compute and related services. While this may seem like a good move in the short term, the process impedes market dynamics and creates bureaucratic hurdles for both providers and users of compute infrastructure. Allowing markets to function freely and offer services that meet consumer needs is necessary for long-term sustainability.
Current approach: challenges
The IndiaAI compute mission follows a bidding process where empanelled vendors must match the lowest bid price for the AI compute and services. Some vendors undercut the market price by as much as 89% by “doing things that were never done before to reduce operational costs”, as one vendor put it. In addition, the IndiaAI mission will subsidise up to 40% of the compute costs for eligible users addressing priority use cases such as healthcare and education.
The government’s move to empanel private compute providers and offer additional subsidies stimulates demand in the near term. This might support the development of India-based compute providers. In addition, if done well, the continuous empanelment process could prevent cartelisation among compute providers and allow newer providers to enter the market. However, it is worth asking if the bidding process that requires matching the lowest bid price is sustainable and how effectively it addresses the goals of developing sovereign computing infrastructure and encouraging innovation in priority areas.
The lowest bid tendering process creates incentives to compromise on the quality of service as operational costs need to be kept to a bare minimum. Slim profit margins will also leave little room for investments in research and development.
Furthermore, the fact that many of the providers agreed to the lowest price seems to be a sign of low private market demand for AI compute in India. Yotta, an empanelled vendor providing a majority of the compute under the IndiaAI mission, claims that just 25% of the demand for their Nvidia H100 chips is from India. While government intervention might stimulate demand in the near term, the below-market prices and the additional subsidy offered may cap private demand at what is available at the reduced prices.
The end-user policy for AI services under the IndiaAI compute mission has several requirements to apply for the compute that create hurdles for users. This involves meeting various qualification criteria and a project and subsidy evaluation process. For instance, a startup applying for compute needs to be registered with Startup India or recognised by the Department for Promotion of Industry and Internal Trade. It will also need to demonstrate experience in AI/Machine Learning and have an annual revenue of ₹50 lakh-₹200 lakh or funding above ₹100 lakh. The monitoring and evaluation team would then evaluate the project and, based on several criteria, approve the allocation and any subsidy requested.
Such a process is necessary to ensure accountability and due diligence, but it adds too much friction, which will hinder innovation. DeepSeek has been disruptive in building AI models comparable to frontier models developed by OpenAI at a fraction of the cost. Some reasons for its success were that it was incubated from a hedge fund, ran its own data centre, and had no business model. This allowed engineers to experiment without navigating through bureaucratic processes or internal coordination to get compute resources.
India’s choice to build its sovereign computing infrastructure may seem like a good move in the short term, as it provides ownership of computing resources and supports the development of a domestic infrastructure-as-a-service market.
Things to prioritise
However, some concerns remain. To create a sustainable market, providers must compete for market share by offering solutions that meet consumer needs instead of just undercutting prices. In addition, the ₹4,500 crore earmarked for the IndiaAI compute pillar over five years will fund only the subsidy component for eligible projects. It is interesting to explore whether there is enough demand for projects that qualify for the subsidy. If not, a significant part of the budget allocation might remain unutilised. Finally, India’s current compute capacity at ~19,000 GPUs pales compared to investments in compute in the U.S., EU, or China. Meta alone is building $10 billion in a data centre. This seems to imply that India’s focus is not on building the most capable AI models but rather on addressing Indian use cases, which is a practical choice.
Concerns that must be prioritised include scaling up energy infrastructure as demands are likely to grow as compute utilisation grows. Further, any hurdles to importing compute infrastructure can also be an area that can be streamlined through government intervention.
Market trends indicate that compute usage is moving from training to inference time, which requires different kinds of AI chips. As the hype settles down and competitors to Nvidia emerge, chip costs may also go down. Government interventions should leave the market as agile as possible to adapt to these shifts. Allowing private players to function freely will enable the market to keep pace with future developments and also provide a pathway for the transition after the IndiaAI mission sunsets.
Published – April 28, 2025 01:57 am IST