The Power of Public Compute: Shaping the Future of AI

Compute: The New Currency of AI

In the rapidly evolving world of artificial intelligence, compute, or computational power, has emerged as a fundamental currency. It not only dictates the technical prowess of AI systems but also influences who gets to build them, shaping the competitive dynamics of the entire AI sector. As AI models become increasingly sophisticated, access to large-scale compute resources is paramount for research advancements, commercial success, and broader participation in AI development.

The Rise of Public Compute Initiatives

Recognizing the importance of compute access, governments worldwide are launching "public compute" initiatives. These initiatives utilize public funds to provide access to crucial compute resources, ranging from providing hardware like GPUs, offering vouchers for private cloud services, to granting access to publicly operated supercomputing projects.

From the multi-billion dollar Stargate Project in the US to the UK's AI Opportunities Action Plan, there's a growing urgency to grasp the profound impact of compute availability. The debate centers on how to ensure these initiatives yield tangible public benefits, a key concern discussed at the AI Action Summit in Paris. This report, supported by the Mozilla Foundation, explores these public compute initiatives, offering insights and recommendations for policymakers.

Four Approaches to Public Compute Provision

Our research has identified four distinct models of public compute provision:

Direct Provision (Generalist)

This model involves large-scale, general-purpose infrastructure designed to support a wide array of applications, from academic research to public sector projects. Examples include the US Department of Energy supercomputers and the European High-Performance Computing (EuroHPC) Joint Undertaking. While offering economies of scale and strategic flexibility, this model faces challenges like coordinating large projects and potential hardware specification issues. The risk of spreading investments too thinly across diverse compute needs also needs careful consideration.

Direct Provision (AI-Focused)

Specifically designed for AI workloads like frontier LLM development, this model utilizes GPU-centric architectures with flexible access and AI-specific support. The UK's AI Research Resource and the US's National AI Research Resource pilot are prime examples. This approach, while promising for supporting AI research and the domestic AI industry, carries risks of entrenching monopolies and over-reliance on the assumption that advanced foundation models automatically lead to scientific breakthroughs.

Decentralized Provision

This model utilizes public resources to facilitate networks of smaller facilities distributed across regions. India's Open Cloud Compute and China's network of municipally-owned data centers showcase this approach. While potentially beneficial for market shaping and supporting domestic compute industries, the scalability and effectiveness of large-scale decentralized investments remain uncertain.

Market-Based Provision

This model involves using public funds to subsidize access to compute resources from existing commercial providers, such as through voucher schemes. India's IndiaAI mission's compute vouchers are an example. This approach, while leveraging existing infrastructure and expanding access, risks simply subsidizing incumbent companies without addressing underlying dependencies in the compute supply chain.

Key Recommendations for Policymakers

To maximize the public benefit of compute investments, policymakers should focus on:

• Avoiding Value Capture: Insulating public compute strategies from industry lobbying and exploring conditional access based on public benefit commitments.
• Achieving Strategic Coherence: Developing integrated strategies for compute infrastructure and skills, coordinating energy and water requirements, and establishing national-level coordination mechanisms while preserving local autonomy.
• Balancing Flexibility and Longevity: Setting long-term targets with flexible delivery models and building modular software infrastructure.
• Squaring Compute Investments with Environmental Goals: Reviewing the environmental impact of AI on data center demand and requiring environmental commitments from suppliers and users.

The future of AI hinges on equitable access to compute resources. By implementing these recommendations, policymakers can harness the power of public compute to drive innovation, support research, and ensure a more inclusive and beneficial AI landscape for all.