The AI Balance of Power

Council on Foreign Relations | Adam Segal

The United States and China currently dominate a bipolar artificial intelligence landscape, with the U.S. holding 75 percent of global GPU cluster performance as of May 2025. This technological duopoly faces disruption from rapid private-sector scaling and the potential emergence of superintelligence within three to five years, threatening to permanently widen the gap over middle powers.

This concentration of capability stems from massive commercial R&D spending, which exceeded $109 billion in 2024 and now surpasses most state military budgets. Private firms now wield sovereign-like digital power. A late 2025 dispute between Anthropic and the Pentagon over military AI deployment in Iran underscores growing friction between state regulators and frontier labs. While Beijing enforces military-civil fusion to direct its tech sector, Washington must foster collaborative relations with private innovators to maintain its global lead and prevent international partners from rejecting the American tech stack, which would slow global adoption.

Comment

The Pentagon's designation of Anthropic as a supply chain risk in late 2025 exposes a fundamental command and control friction in Western military technology adoption. Unlike traditional defence acquisition, where the state dictates terms to contractors, the proprietary nature of the Claude model leaves the state dependent on commercial goodwill. This power asymmetry limits the Pentagon's ability to enforce compliance without jeopardising active U.S. Central Command operations in Iran. Sovereign authority no longer guarantees operational control. Consequently, Washington's leverage is severely constrained when San Francisco-based labs restrict model access based on corporate safety charters.

The mechanism of this dependency lies in the closed-source architecture of the Claude model, which prevents the U.S. military from hosting and maintaining the software on sovereign servers. Because Claude requires continuous API access and cloud compute infrastructure managed by the developer, the Pentagon cannot unilaterally sustain the system if the firm terminates service. This operational vulnerability is compounded by the lack of in-house technical expertise within the Defense Information Systems Agency to audit or replicate Anthropic's proprietary neural networks.

Strategic Question for Discussion

What happens to the concept of sovereign operational command if the Pentagon remains dependent on continuous API access to proprietary systems like Anthropic's Claude during active combat operations?

The trajectory indicates that traditional command structures will be forced to accept a model of shared sovereignty, where private tech firms hold an effective veto over specific mission parameters. To mitigate this, the military is likely to pivot toward smaller, open-source models that can be fine-tuned and hosted locally on secure tactical networks. However, this shift may require sacrificing the cutting-edge capabilities of frontier models, presenting a stark trade-off between operational autonomy and technological superiority.

Share your assessment in the comments below.

To read in details click here