This essay is part of a series that highlights the main takeaways from discussions that took place at Carnegie India’s eighth Global Technology Summit, co-hosted with the Ministry of External Affairs, Government of India.

---

As the conversation on artificial intelligence (AI) continues to evolve around the risks and opportunities involved, an issue coming into sharp focus is “compute.”

What Is Compute?

In technical terms, compute is a measure of the calculations that can be performed by a processor, usually represented in floating-point operations per second, or FLOPS. For reference, the world’s most powerful supercomputer clocks nearly 1.2 exaflops—more than a billion-billion calculations per second.

A more holistic understanding of compute is that it is a technology stack that combines a hardware layer of graphic processing units (GPUs), an infrastructure layer of data centers and server optimization algorithms, and a software layer of development frameworks. 

Another way to think about compute is that it provides the sophisticated technical capabilities required to generate meaningful insights from large volumes of data for specialized operations such as natural language processing and object recognition. In fact, studies have shown that, on average, the performance of a large language model (LLM) improves with an increase in the compute power made available to it.

The Scarcity Problem

Access to compute creates the promise of technological progress, which explains why it is considered a strategic geopolitical asset. However, even as demand for compute continues to climb around the world, governments are facing an acute supply shortage.

This problem of scarcity stems from a combination of various factors: the high cost of the raw materials and specialized equipment required to manufacture silicon chips, the shortage of skilled professionals capable of developing and maintaining advanced compute systems, and the concentration of these resources in the hands of a few private corporations.

Moreover, with the factors of production being concentrated in the developed world, the concern is that the digital divide between the Global North and South will continue to widen. 

To Maximize or Optimize?

While the Indian government recognized compute as an element of its AI strategy in 2018, only recently has it outlined the specific steps required to enhance its compute capacity.  

However, given the problem of scarcity, there is a need for deeper analysis on how India’s compute capacity should be enhanced and whether it is the right goal to pursue.

An emerging debate based on discussions at the Global Technology Summit (GTS) is whether India’s national AI objectives can be achieved through small, custom, and open-source models designed for specific use cases, which will require less compute, instead of large, compute-intensive models that are designed for general use.

Should India choose to adopt a use-case-led strategy, it may be more efficient to optimize the use of existing compute resources and step away from the global arms race for compute. This would be in contrast to China’s approach, which has resolved to scale up its compute capacities for strategic reasons. 

India’s choice on this matter will have implications for its technological capabilities, economic competitiveness, and national security. We call this the “compute conundrum.”

Democratizing Access

Whether India chooses to maximize or optimize its compute capacity (or both), it must take efforts to democratize access to its compute resources, especially for academics and startups.

Two ideas presented at the GTS merit attention in this regard:

1. A global repository similar to a “CERN for AI” or the AI-on-demand platform (AIoD) in the EU, when combined with suitable cross-border data sharing frameworks, can help make compute more readily available to researchers.
2. In the Indian context, some have proposed developing a “digital public infrastructure (DPI) for compute,” through which micro-data centers running on interoperable standards would enable small businesses to shift their workloads from hyperscale cloud service providers and avoid vendor lock-in.

India’s Compute Strategy

We outline three factors that should inform India’s compute strategy.

1. Scalability: India’s AI strategy is largely use-case-driven. To that end, the value of large, general-purpose models is limited. Instead, compute resources may be optimized for specific applications using smaller models. At the same time, the DPI experience suggests that positive network effects emerge at population scale, so India’s compute infrastructure should be scalable enough to meet these demands as and when the need arises. This could be achieved by tapping into compute capacity in the private sector—Meta, for example, intends to acquire 350,000 H100 GPUs in 2024, more than ten times the Indian government’s target of 25,000 GPUs.  
2. Sovereignty: India’s desire to develop sovereign AI infrastructure aligns with its larger goal of national security and technological self-sufficiency, with a sharp eye on neighboring China’s efforts to scale up its own compute capacity. However, India should be circumspect in placing geographically linked restrictions, as it has done in the past—doing so would prevent it from fully participating in the global AI market. Instead, India should partner with like-minded countries on compute, cybersecurity, data sharing, and research, in the spirit of “collaborative AI.”
3. Sustainability: India has set ambitious targets to achieve net-zero emissions. Given the large carbon footprint of running AI models, the desire for more compute may run counter to India’s environmental goals. Therefore, at least a portion of India’s planned R&D investments in AI should be directed toward developing energy-efficient compute infrastructure such as nuclear fusion, which is likely more sustainable. 

A Reliable Measure

To inform its compute strategy, the government should conduct a comprehensive survey to measure existing compute capacity and projected needs over the next few decades.

While some data from the National Supercomputing Mission has been made publicly available, it is not a reliable measure of national compute capacity. As a recent OECD paper notes, “Without a clear framework to help countries measure and benchmark their relative access to AI compute capacity, countries may be unable to make fully informed decisions on which investments are needed to fulfill their AI plans.”

By investing in a reliable measure of existing capacity and projected needs, India will be better positioned to address the “compute conundrum” and realize its digital aspirations.