The Power of Public Compute: Shaping the Future of AI

The Compute Conundrum: AI's Growing Hunger for Processing Power

In the rapidly evolving world of Artificial Intelligence, compute – the raw processing power needed to train and run complex algorithms – has become the lifeblood of innovation. As AI models grow increasingly sophisticated, access to vast compute resources has emerged as a key differentiator, determining who can participate in, and ultimately shape, the future of this transformative technology.

The rise of large language models (LLMs) like GPT-3 and GPT-4 has dramatically amplified this demand, creating a "compute divide" between well-resourced tech giants and smaller players, including researchers, startups, and non-profits. This divide threatens to stifle innovation and concentrate control over AI development in the hands of a few.

Public Compute Initiatives: Leveling the Playing Field

Recognizing the strategic importance of compute and the risks posed by the compute divide, governments worldwide are exploring "public compute" initiatives. These initiatives aim to leverage public funds to broaden access to computing resources, fostering a more diverse and inclusive AI ecosystem.

From direct provision of hardware and software to voucher schemes and public-private partnerships, these initiatives take a variety of forms. Some focus on supporting general scientific research, while others target specifically AI-related workloads like LLM development. The goal is to empower a wider range of actors to contribute to AI advancement, ensuring that the benefits of this technology are shared broadly.

Four Families of Public Compute Provision

Our research identifies four distinct approaches to public compute provision:

• Direct Provision (Generalist): Large-scale, multi-purpose facilities supporting a wide range of research activities. Examples include the US Department of Energy's supercomputers and the European High-Performance Computing (EuroHPC) Joint Undertaking. These initiatives offer economies of scale and strategic flexibility, but can be challenging to coordinate and may not always be optimized for AI-specific workloads.
• Direct Provision (AI-Focused): Infrastructure tailored specifically for computationally intensive AI tasks like LLM training. The UK's Isambard-AI project and the US's National AI Research Resource pilot are prime examples. While offering greater specialization, these initiatives face significant uncertainty around future demand and risk entrenching existing market monopolies.
• Decentralized Provision: Distributed networks of smaller facilities across regions, offering greater accessibility and promoting local capacity building. China's network of municipally-owned data centers and India's Open Cloud Compute initiative demonstrate this model. However, challenges remain around technical integration and achieving strategic scale.
• Market-Based Provision: Government subsidies or vouchers to help organizations access compute resources from commercial providers. India's compute vouchers under the IndiaAI mission are a case in point. This approach can leverage existing infrastructure, but risks further empowering dominant market players.

Navigating the Challenges: Recommendations for Policymakers

Across these diverse approaches, several key challenges emerge, including value capture by private interests, strategic coordination, achieving both flexibility and longevity, and mitigating environmental impact. To address these challenges, we offer a set of recommendations for policymakers:

Ensuring Public Benefit and Avoiding Value Capture

Public compute initiatives should prioritize public benefit. This can be achieved by insulating strategies from industry lobbying, using procurement strategically to support a diverse supplier ecosystem, and exploring access conditions that ensure public value creation.

Achieving Strategic Coherence

Effective coordination is essential. National-level mechanisms should align initiatives across different levels of government, while preserving local autonomy. Integrated strategies for infrastructure development and skills building are also crucial.

Balancing Flexibility and Longevity

Long-term planning is important, but flexibility is equally crucial. Strategies should set long-term targets and funding envelopes while maintaining adaptable delivery mechanisms. Modular software infrastructure can further enhance this flexibility.

Minimizing Environmental Impact

The environmental footprint of compute infrastructure cannot be ignored. Public compute initiatives should prioritize sustainability, requiring environmental commitments from both suppliers and users, and actively coordinating with clean energy development projects.

"Public compute has emerged as a critical frontier in AI policy...The challenge for policymakers is not just to provide compute resources, but to do so in ways that actively shape markets, support diverse innovation and advance broader societal goals."