It is necessary to reevaluate the Indian government's Rs 10,000 crore proposal for "sovereign AI" computer infrastructure

It is necessary to reevaluate the Indian government's Rs 10,000 crore proposal for "sovereign AI" computer infrastructure

The goal of using AI's advantages will be more effectively achieved in fields that the commercial sector often ignores. When it comes to buying GPUs and overseeing computer infrastructure, the government may not be able to keep up with the market's agility. As an alternative, the government may support research, encourage competition, control anti-competitive behavior, and assist in the creation of datasets unique to India.

Globally created AI apps may not concentrate on use cases unique to India or might not perform as effectively in Indian settings.

With a budget of Rs 10,000 crore, the federal government has unveiled plans for an ambitious artificial intelligence (AI) computing effort. The program aims to establish a "Sovereign AI" computer infrastructure that can provide Indian entrepreneurs, particularly those in the fields of education, healthcare, and agriculture, computing resources as a service.

AI is without a doubt a technology with a wide range of applications and tremendous transformational potential. The government's plan to construct the required computing infrastructure, however, is not the most efficient use of tax dollars. By concentrating its efforts and resources on areas that the private sector does not usually handle, the government will better achieve its goal of helping Indians reap the advantages of artificial intelligence.

Exists a failure of the market?

In general, it makes more sense to train and implement AI systems using cloud computing infrastructure rather than preserving specialized, on-premises computer resources. It has advantages including scalability, simplicity of administration, and cost efficiency. Major players like Google Cloud, Microsoft Azure, and Amazon Web Services dominate the industry with a combined 65% worldwide share, mostly due to their economies of scale. This kind of government intervention is unnecessary since the market can still operate efficiently in spite of this concentration.

The government's ambition to establish its own national computer infrastructure has several flaws.

First, lead times for the current generation of NVIDIA GPUs, such as the H100, which power AI-related workloads, expand to 52 weeks due to the high demand. The government won't be able to utilize these GPUs for another year, even if they purchase them today. By then, more sophisticated and recent GPU models could be accessible, and there will probably be a greater number of cloud service providers providing AI computing infrastructure. Alternative processors for AI processing, such FPGAs or ASICs, could potentially progress over time. Sovereign computer infrastructure may become superfluous and redundant because to these advancements.

Second, in order to maintain current hardware and software, managing computer infrastructure need specific knowledge and ongoing investment. Moreover, it doesn't integrate well with other services that are often provided by for-profit cloud providers. Third, the government's distribution of computer resources to priority programs usually impedes the effective distribution of resources realized via market pricing. There will also be a lack of the scalability and flexibility provided by commercial cloud providers. Lastly, there is a substantial social cost associated with government expenditure; for every rupee invested, society is expected to lose Rs 3. Only when the societal advantages exceed this maximum can such investment be justified.

One of the main justifications for creating a sovereign computer infrastructure is the narrative of Aatmanirbhar Bharat, or self-sufficiency. Technology and geopolitical concerns are becoming increasingly intertwined in recent years. Numerous nations, such as the US and China, have recognized artificial intelligence (AI) as a crucial technology and are working to manage the chokepoints of this field. However, encouraging competition in the area could assist in reducing these geopolitical dangers. In lieu of direct government intervention, encouraging conventional cloud computing services is a workable approach.

The government may assist in the production and use of Indian datasets that adhere to data protection laws, according to another juicy detail included in the announcement for sovereign computing infrastructure. This argument, however, is not persuasive. Data will travel to the most efficient sites for processing and storage. Data security and usefulness are not compromised when stored abroad as opposed to locally. Certain jurisdictions have severe standards for data localization, although private markets are more effective at achieving such needs. While some military and national security applications would need the use of sovereign computer resources, such situations ought to be the exception rather than the rule.

Where should the state direct its attention?

It is not always the case that markets can function efficiently at various phases of the AI supply chain, including data, computation, models, and application. This is despite the fact that there is no market failure that calls for sovereign computer infrastructure. Government involvement can eliminate these inadequacies. The regions below are a few potential locations.

For instance, NVIDIA, a chip manufacturer, has a market share of over 90% for GPUs required for operations involving AI. The reason for this market domination is NVIDIA's early foray into the industry and the extensive use of their in-house developed computing platform, CUDA. The government might provide funding for studies on open-source, next-generation architectures that could eventually take the place of GPUs for AI-related jobs as a long-term risk mitigation measure.

Prominent cloud service companies such as Google, Microsoft, and Amazon are able to combine their cloud services with their own proprietary AI models. This kind of vertical integration, which crosses many AI supply chain tiers, can discourage rivalry. Competition authorities should keep a close eye on this industry's activities and take appropriate action against those that limit competition.

Lastly, AI apps created outside of India may not concentrate on use cases unique to that country or might not function as effectively there. For AI to function well, for instance, data that is representative of the population may be needed due to India's particular circumstances or great variety. The government may focus on producing these India-specific datasets. The idea of "public money, public data" should guide the release of these datasets for public use. These will improve AI applications' performance and relevance in addition to helping research. One such MeiTY project is Bhashini, which aims to simplify the variety of Indian languages and provide information and resources for real-time translation, thereby capturing the diversity of the people living in India.

With the significant societal costs and complexity required, the Indian government's ambitious ambition to develop sovereign computer infrastructure may not be the most practical use of resources, despite its good intentions. Public funds should be directed on endeavors that the market will not support instead.