GoI’s Rs 10,000 crore plan for a ‘sovereign AI’ computing infrastructure needs a rethink

The government's plan to establish its sovereign computing infrastructure has several problems.

First, there is a massive demand for the current generation of NVIDIA GPUs like the H100, which power AI-related tasks, and the lead times stretch up to 52 weeks. Even if the government orders these GPUs now, they won't be able to use them until a year later. By then, newer and more advanced GPU models may be available, and the number of cloud service providers that offer AI computing infrastructure is also likely to increase. Over time, there might also be advancements in alternative chips for AI processing, such as FPGAs or ASICs. These developments could make the sovereign computing infrastructure outdated and redundant.

Second, managing computing infrastructure demands specialised expertise and continuous investment to stay updated with the latest hardware and software. It also lacks the seamless integration with other services typically offered by commercial cloud providers. Third, the government's allocation of computing resources to important projects disrupts the efficient allocation of resources usually achieved by market pricing. The flexibility and scalability offered by commercial cloud providers would also be missing. Lastly, the social cost of the government's spending is high; every Rs.1 spent is estimated to cost society Rs.3 effectively. Such spending is only justifiable if the societal benefits exceed this high threshold.

A key argument for establishing sovereign computing infrastructure hinges on the narrative of Aatmanirbhar Bharat, or self-reliance. Technology is becoming more entangled with geopolitical considerations in the recent past. Many nations, including the US and China, have identified AI as a critical technology and are attempting to keep the chokepoints of these technologies under their control. Even so, promoting competition in this sector can help mitigate these geopolitical risks. Encouraging domestic and international cloud computing services, rather than direct government involvement, is a viable strategy.

Another justification cited for sovereign computing infrastructure is that the government can facilitate the creation and use of Indian datasets that comply with data protection guidelines. However, this argument does not hold water. Data will flow to locations where storage and processing are most efficient. The security or utility of the data is not compromised by storing it internationally rather than domestically. Even if strict data localisation requirements exist for certain sectors, private markets are more efficient at addressing that requirement. Although the use of sovereign computing resources might be indispensable for specific military and national security applications, such scenarios should remain exceptions and should not become the standard practice

Where should the government focus its efforts?

While there are no market failures requiring a sovereign computation infrastructure, markets might not always work effectively at different stages of the AI supply chain, such as data, computation, models, and application. Government intervention could address these shortcomings. Some potential areas are listed below.

For example, NVIDIA, a chip manufacturer, holds more than 90% of the market share for GPUs, essential for AI-related tasks. This market dominance stems from NVIDIA's early entry into the market and the widespread adoption of its proprietary computing platform, CUDA. As a long-term risk mitigation strategy, the government could fund research on the next generation of open-source architectures that could replace GPUs for AI-related tasks.

Major cloud service providers like Amazon, Microsoft, and Google may bundle their proprietary AI models with their cloud services. This type of vertical integration, spanning various segments of the AI supply chain, could be a deterrent to competition. It's crucial for competition regulators to vigilantly monitor and address any practices that may stifle competition in this sector.

Lastly, AI applications developed globally might not work as well in Indian contexts or might not focus on India-specific use cases. For instance, India’s unique conditions or extreme diversity might need data representative of the population for AI to work effectively. The government can direct its efforts towards creating such India-specific datasets. These datasets should be made open access following the principle of “public money, public data”. They would not only benefit research but also enhance the performance and relevance of AI applications. Bhashini is one such initiative under MeiTY to capture the diversity of Indians to simplify the diversity of languages in the country and provide data and tools for real-time translation.

The Indian government's ambitious plan to build sovereign computing infrastructure, while well-intentioned, may not be the most prudent use of resources, given the high social cost and the complexities involved. Public resources should instead be channelled towards areas which the markets would not address.

Bharath Reddy is a researcher in the High-Tech Geopolitics programme at the Takshashila Institution. Views are personal and do not represent the stand of this publication.