This two-part series explores strategic recommendations for the government on compute (Part I) and data (Part II).

Compute: The Backbone of AI Innovation

Compute stack comprising hardware, software, and infrastructure layers underpin modern digital communication. Advanced compute stacks are quintessential for training AI models. While the hardware layer includes chips such as GPUs, the infrastructure layer generally includes data centres, which pool thousands of chips together. The software layers include apps that enable the use of chips, domain-specific optimised languages, and data management services. 

India has sought to democratise access to compute and promote indigenous production, but progress has been slow. The National Semiconductor Mission, launched in 2022, aims to domestically produce GPUs, but currently, the production is primarily limited to less sophisticated legacy chips like DRAM chips.

In 2020, the Indian government proposed AIRWAT, an AI-specific cloud computing infrastructure, to offer access to compute for startups, academics, and researchers. It was created without relying on commercial cloud solutions. However, AIRAWAT’s capacity (656 GPUs) pales in comparison to the capacity of over 10,000 GPUs used by Meta and Microsoft to train their models.  

Recognising the compute deficit, the government allocated approximately INR 4,568 crores under IndiaAI Mission to build a national AI compute capacity of at least 10,000 GPUs. Building upon the lessons of AIRAWAT, the government would probably use a PPP model to build capacity.

India’s IT ministry is considering either a rent-and-sublet model or a marketplace approach. The former involves the government purchasing GPUs and then furnishing them to businesses. Under the latter, the government would directly negotiate or help companies negotiate with hyperscalers to acquire compute capacity. 

A Marketplace Approach: The Way Forward

We believe that a marketplace approach is more suitable for the deployment of the compute infrastructure under the IndiaAI Mission. Firstly, under the rent-and-sublet model, the government would need to acquire thousands of GPUs, incurring significant financial costs. Secondly, the marketplace approach offers more flexibility to actors in the ecosystem by enabling them to negotiate based on their business needs.

Hyperscalers can offer different compute allocations and services based on the needs of an AI business—training, fine-tuning a model, or inference. 

The government can further consider innovative decentralised or federated cloud/compute solutions. Federal learning aims to reduce centralisation by training data at the edge. The IndiaAI Report 2023 advocated for edge computing to meet India's compute needs, and a proof-of-concept of this is People+AI's “Open Cloud Compute” (OCC) project. The OCC is a network of micro-data centers based on interoperable standards.

Reassessing India’s Approach to Compute

The IT Ministry should also critically evaluate whether maximising compute capacity is strategically the best decision. So far, it appears that the government has focused on the hardware layer of the stack, with an emphasis on maximising compute capacity.

This approach mirrors that of the US and China, which is geared toward developing advanced AI models—a race that India may not be able to compete in or even should. A more optimal approach would be to focus on sectors that cater to India’s strategic interests and needs, avoiding a race to the bottom - continuous expansion of compute.

This pragmatic strategy would allow India to leverage its less advanced chips to train models optimised for specific applications with relatively limited computational requirements.

Strategic use cases can be identified based on either areas where India is well-suited for success, or priority areas and sectors where AI can be deployed for social good at scale.

In the former, India could concentrate on areas with abundant indigenous data, such as content in numerous Indian languages or data from the business outsourcing industry, useful for creating automation tools. For the latter, sectors can be chosen based on the transformative impact of AI. For instance, the NITI Aayog AI Strategy 2018 highlights agriculture, healthcare, education, and e-governance as key areas for intervention.

The focus on the hardware layer also neglects other layers of the stack, diminishing autonomy in designing technical architecture, despite India’s potential to play a major role in the software layer.

India has a world-class, market-leading software industry, and its open-source approach may resonate with businesses that are looking to tackle Nvidia's closed ecosystem approach. 

Within the first 100 days, the IT ministry should conduct a comprehensive assessment of the computing requirements to ensure the limited compute budget under the IndiaAI Mission is allocated in a Pareto-efficient manner.  

To be fair, the IT ministry did carry out a similar assessment under its IndiaAI Report 2023, but the report had one major flaw: it focused on how India can expand its compute capacity. It, however, missed a more preliminary inquiry—what approach would be best suited to leverage India’s strengths and capabilities?

The IT ministry’s assessment should aim to fill this gap, and this can be informed by comparative best practices.

Concluding thoughts 

It is imperative to accelerate the momentum built in promoting AI innovation and development over the last few years, given the transformative potential of AI. Lack of prioritisation may leave India trailing in this critical technological race.

We believe a pragmatic view that relies on a use-case-specific strategy would be best suited for India. Instead of focusing solely on hardware, emphasis should be placed across the stack to enable flexibility and opportunity for actors in the AI ecosystem. These recommendations allow India to chart its unique path, tailored to capitalise on its strategic interests, strengths, and aspirations.

(Rudraksh is an associate and Rutuja leads government affairs practice at Ikigai Law. Views expressed are personal.)