From immersive gaming experiences to life-altering medical diagnostics, these advancements are powered by intricate mathematical models trained on large computers to mimic human cognitive functions – the essence of artificial intelligence (AI). 

Many countries have developed national AI plans as they continue to recognise AI as a pivotal technology for economic growth. While these plans are largely aspirational, outlining strategies to harness AI benefits and establish protective measures, they commonly share a significant limitation. 

They were formulated without a comprehensive assessment of a country's access to the necessary compute capacity for creating, training, and utilising AI models. In essence, nations are aspiring to leverage economic gains from AI without ensuring adequate compute capacity to realise these ambitions. 

The Problem AI Poses For Policy-Makers

The relative neglect of this issue can partly be attributed to the inherently technical nature of AI compute, which requires a level of expertise beyond that of most policymakers. Additionally, some government officials may mistakenly perceive AI compute as commoditised and readily available, similar to traditional information technology infrastructure. 

This misconception may stem from the widespread availability of commercial cloud providers offering scalable and robust services, as noted in the US National AI R&D Strategic Plan. While a universally accepted definition of AI compute capacity is lacking, it comprises a specialised combination of software and hardware components, including processors, memory, and networking. 

This integrated system is designed to support AI-specific workloads and applications, ranging from large data centres, supercomputers, and cloud providers to more compact data science laptops and workstations. In the realm of AI advancement, access to computing resources stands as a crucial facilitator, aligning with data and algorithms. 

Consequently, the ability to quantify and comprehend a country's access to AI compute capacity is foundational for establishing a robust AI ecosystem. This framework is essential for ensuring that countries and their citizens can fully enjoy the benefits of AI technologies. The absence of a clear framework for measuring and benchmarking relative access to AI compute capacity poses a risk. 

Without such guidance, countries may struggle to make well-informed decisions regarding the necessary investments to realise their AI plans. This could potentially lead to a division in the nation’s ability to compute complex AI models essential for a competitive edge in the global digital economy. 

India is one of them. In the first week of March this year, the government cleared an IndiaAI Mission with a budgetary outlay of Rs 10,371 crore. The government’s support comes as a bid to build an artificial intelligence hub through a series of initiatives including one to develop computing infrastructure involving over 10,000 GPUs or graphics processing units. However, the way forward may not be that easy to define. 

The Indian Context

In this context, for India to cultivate a robust domestic industry capable of competing with global giants, effective AI governance must ensure unimpeded access to various stages of the supply chain. Several crucial considerations come into play at each stage of this supply chain. 
The performance of models trained on large, diverse, and high-quality datasets is notably superior. 

Studies project that, for training extensive language models, top-tier data from sources like books, academic papers, news articles, and Wikipedia may be depleted by 2027.

Consequently, access to proprietary data emerges as a pivotal factor in differentiating AI model training. This encompasses datasets such as the extensive code repositories, found on platforms like GitHub, meticulously organised web indices compiled by search engines through web crawling, and audio-video recordings from platforms like Zoom. 

Predominant tech platforms often function as monopolies or duopolies due to network effects, consolidating markets across various domains, from search engines and social media to ride-sharing and food-delivery services. Their privileged access to proprietary data grants them a substantial advantage in innovation over competitors.

This access not only enhances the development of AI applications but also facilitates seamless integration into existing offerings, further solidifying their market dominance. Research underscores the existence of fairness issues in AI applications, particularly in medical diagnoses, gender classification, recidivism prediction, and other areas where certain groups are underrepresented in training data. 

This imbalance can result in elevated error rates and biased outcomes for minorities, women, the elderly, and vulnerable populations. Given India's remarkable diversity, the government can play a pivotal role in fostering inclusivity by creating publicly accessible open datasets representative of the population. 

Initiatives like Bhashini, an AI-driven language translation system, backed by the Indian government, strive to capture the diversity of Indian languages, providing opensource databases and tools for real-time translation. Such practices contribute positively to both research and commercial applications. 

As the global community actively grapples with the complexities of AI comprehension and regulation, it becomes imperative for India to strategically adapt to its unique conditions in the ever-evolving geopolitical landscape. 

Implementing decisive measures is not merely a recommendation but a necessity to ensure unimpeded access to AI technology that can help harness its profound benefits. In navigating this dynamic frontier, the Indian government holds the key to fostering an environment where inclusivity works hand-in-hand with innovation to grow. This can perhaps propel the nation to the forefront of the AI revolution.

AI Compute Divide

Examining supercomputers provides a tangible illustration of the problems that India and many other developing countries have to face; many suppliers operate in concentrated markets, often located outside the country of use. This scenario raises concerns about interdependencies between countries, resulting in what can be termed an "AI compute divide" for advanced and intricate AI models.

 An analysis of the 2023 TOP 500 supercomputers highlights that fewer than 35 countries possess a system qualifying as a top supercomputer, with a majority situated in the United States, Japan and China, leaving the global south notably underrepresented. 

Moreover, evidence suggests the existence of an AI compute not only divides between countries but also within national boundaries of the nation. In the United States, for instance, only a limited number of private firms and universities can afford to acquire and maintain leadership-class AI supercomputers, potentially creating disparities among researchers. 

Similarly, a recent report from the United Kingdom Government Office for Science reveals challenges faced by smaller research centres and businesses in accessing large-scale computing platforms, constraining the scope of their AI development efforts. 

Assessing Investment In Domestic AI Capacity

Nations must carefully assess how investments in domestic AI compute capacity align with diverse policy objectives. A prime example is the strategy of "scaling up" AI compute, where investments are directed towards a limited number of larger AI systems tailored for training the most intricate models. 

This scaling-up approach significantly contributes to advancements in critical domains such as Natural Language Processing (NLP), precision medicine, and autonomous vehicle development. 

On the flip side, the "scaling out" approach involves investing in a larger number of smaller AI systems to facilitate AI research and development (R&D) initiatives, including workforce training and student education.

In this scenario, the emphasis is on accessibility rather than groundbreaking achievements. Notably, this scaling-out strategy is more prevalent in Southeast Asia, specifically in countries like Thailand and Indonesia. Here, multiple smaller AI clusters are strategically deployed across universities with government support, aiming to broaden access. Governments should tailor AI compute investments according to their specific policy objectives. 

In essence, there are varied methods to enhance domestic AI compute capacity, and the most resilient approach depends on a country's unique context and domestic needs. Potential strategies encompass investments in nationally owned or sponsored AI supercomputers and strategic collaborations with global and regional commercial cloud providers.
 
It's essential to acknowledge that valuable AI compute resources can also be on a smaller scale, particularly beneficial for students and junior researchers. Even the provision of a data science laptop or workstation, which doesn't entail the infrastructure overhead of a data centre, can be a potent avenue for fostering AI innovation, expanding access, and addressing the compute divide. 

The development of AI systems involves several components – data, computation, models, and applications – envisioned as layers. Data and computation form the foundational layers contributing to the model, which subsequently supports various applications. Companies engaged in AI model or application development encounter entry barriers at each stage.
 
Vertical integration, where a single company controls multiple stages of the supply chain, is not uncommon. Google serves as an exemplar in this regard, showcasing a high level of integration across stages. Its operations encompass developing and training proprietary AI models using extensive data, offering cloud services, and seamlessly integrating AI systems into diverse applications for web and Android users. 

(The author is a US-based management consultant, with an interest in Artificial Intelligence, financial services and driving humanity forward. Views expressed are personal)