Anthropic's 'Structured Access' Proposal Defines US AI Policy

The race for artificial intelligence leadership is increasingly a battle over physical infrastructure. As the computational cost of training frontier models skyrockets—with Google’s Gemini Ultra demanding an estimated 50 times more compute than GPT-4—the policy governing access to this power becomes a critical determinant of national competitiveness. This debate has intensified following a public warning from AI research company Anthropic, which cautioned the U. S.against centralizing its AI infrastructure in a state-controlled model similar to China’s. This development brings the fundamental philosophies of AI governance into sharp relief, highlighting a global divergence in strategy between the US, China, and the EU. The infrastructure policies emerging from this US-China AI governance comparison will shape the landscape of innovation for the next decade.

Key Points

• Anthropic advocates for a “US structured access AI proposal,” where private industry builds infrastructure under robust government-set safety standards developed by bodies like the U. S. AI Safety Institute (USAISI).

• Training state-of-the-art models requires substantial investment, with the Stanford AI Index documenting estimated costs of over $78 million for GPT-4 and nearly $191 million for Gemini Ultra.

• China is implementing a state-directed “national computing power network,” treating compute as a strategic asset to be centrally allocated, a stark contrast to the market-led approach favored by US industry leaders.

• Global AI governance frameworks demonstrate diverging paths: the US debates infrastructure access models, China implements centralized state control, and the EU’s AI Act establishes risk-based rules for AI systems entering the market.

The Trillion-Parameter Price Tag

The debate over AI infrastructure policy is grounded in staggering economic and technical realities. The 2024 Stanford HAI “Artificial Intelligence Index Report” documents an exponential growth in the compute required for frontier models. This isn’t a minor increase; it’s a compounding surge that directly translates into massive capital expenditure. The estimated training costs—$191 million for Gemini Ultra—represent only a fraction of the total investment.

This demand fuels a global data center construction boom. According to data from Synergy Research Group, spending on data center hardware and software reached nearly $74 billion in the fourth quarter of 2023 alone, with full-year 2023 spending surpassing $260 billion, a significant portion allocated specifically for generative AI. These facilities are not just warehouses for servers; they are highly specialized clusters packed with tens of thousands of advanced GPUs. Control over who can build, access, and operate these engines of innovation is now a central point of national strategic contention.

Silicon Superpowers: Competing Infrastructure Visions

At the heart of the policy discussion are two fundamentally different models for managing national AI compute resources. Anthropic’s recent AI infrastructure recommendation news has clearly defined the choice facing US policymakers.

The company champions a framework it calls “structured access.” In this model, the government’s role is not to build data centers but to ensure they are used safely. It relies on a competitive private market to drive innovation in hardware and efficiency, while government bodies like the USAISI at NIST would establish mandatory safety testing and auditing standards for the most powerful models. Public resources, such as the National AI Research Resource (NAIRR) pilot program, are then designed to supplement—not supplant—the private market by providing academics and startups with crucial access to computational resources.

This stands in direct opposition to the “China approach.” China is actively building an integrated “national computing power network,” a strategy that treats compute as a public utility akin to the power grid. As documented by think tanks like the Carnegie Endowment for International Peace, this state-led investment is part of a broader push for technological self-sufficiency and aims to create a unified resource that can be centrally managed and allocated to advance national priorities, prioritizing strategic alignment over open market competition.

Digital Infrastructure’s Democracy Dilemma

The choice between these models carries significant consequences for innovation, safety, and competition. Expert analysis from organizations like the Center for Security and Emerging Technology (CSET) supports the market-driven view, arguing that competition fosters a diversity of technical approaches. CSET research demonstrates that overly restrictive licensing “inadvertently lock[s] in today’s leading technologies and business models,” stifling future innovation. A centralized, government-permitted system, as Anthropic’s Jack Clark testified before a Senate subcommittee, creates a single point of failure and reduces the pace of research.

However, other experts from institutions like Brookings identify specific market failures. The immense cost of frontier AI creates high barriers to entry, concentrating power within a few tech giants. This has led to arguments for a stronger government role, perhaps through public-private partnerships or an expanded NAIRR, to ensure academia is not locked out of cutting-edge research. Furthermore, national security proponents argue that government visibility into the largest compute clusters is necessary to monitor for dangerous capabilities and prevent misuse.

This positions the US debate between two poles, with the EU offering a third way. The EU’s AI Act largely sidesteps direct infrastructure control, instead focusing on placing risk-based obligations on AI systems being sold within its market. This creates a complex regulatory patchwork that global AI companies must now navigate.

Computing Power, Computing Politics

The policy decisions made today will define the trajectory of AI development for years to come. The US is charting a course between a fully private market and complete state control, leveraging institutions like the USAISI to act as a referee. This “structured access” model bets on the ability of a competitive ecosystem to drive progress, balanced by robust, government-mandated safety evaluations. It presents a distinct alternative to China’s top-down, state-directed strategy and the EU’s market-focused regulatory framework.

The central question remains: which approach will most effectively balance the immense drive for innovation with the critical need for safety and democratic oversight? The answer will determine not only the next generation of technology but also the global distribution of power that comes with it.