The Balance of AI Power: Anton Leicht on Politics, Pacing Deals, and Muddling Through Well

What It Covers

Anton Leicht, Carnegie Endowment fellow and author of the Threading the Needle Substack, analyzes the political economy of AI power across five dimensions: current model danger levels, US-China pacing deal feasibility, independent auditor access to frontier labs, Europe's compute-for-access strategy, and why most of the world faces AI-driven disempowerment even as living standards rise.

Key Questions Answered

• •US-China Pacing Deal Viability: A symmetric frontier scaling pause is structurally unfavorable to the US because China leads in robotics, manufacturing, electricity buildout, and AI diffusion — leaving frontier AI as America's sole decisive advantage. Any deal that pauses AI scaling without also freezing China's semiconductor indigenization effort effectively closes the chip gap during the pause window, making the US resume the race from a weakened position. The deal math only works if chip production pipelines are included, which China will not accept.
• •Independent Auditor Access — Minimum Viable Mechanism: Without executive pressure mandating access, frontier labs can credibly cite IP and operational security to reject third-party investigators. The actionable floor is two steps: the administration designates an approved list of evaluators (METR, Redwood, etc.), labs choose from that list for incident investigations, and evaluators report directly to the administration on whether access was sufficient. Continuous embedded oversight — evaluators sitting in Slack channels and meeting executives quarterly — is the second tier.
• •Industry Self-Governance Deadline: A presidential ultimatum giving five to seven frontier companies 90 days to produce a peer-enforcement framework for pacing could work if OpenAI, Anthropic, and Google DeepMind anchor it. The structural problem is Meta and xAI hold administration influence and oppose coordination, so any consensus they join will be diluted below meaningful thresholds. The mechanism only functions if the administration provides substantive minimum standards rather than an open-ended "figure it out" mandate.
• •Antitrust and Export Control Clarification: Safety-minded AI research collaborations — including cross-lab coordination and international academic partnerships — face legal ambiguity under both antitrust law and broadly worded export control statutes. A presidential non-enforcement letter explicitly carving out AI safety coordination from antitrust action would remove the primary legal deterrent. For export controls, the risk of vindictive enforcement is real but contestable; safety nonprofits with limited IPO exposure should consider accepting litigation risk rather than self-censoring research collaborations preemptively.
• •Europe's Compute-for-Access Strategy: Europe's actionable leverage is ASML and the broader semiconductor manufacturing tooling supply chain. The proposed framework: Europe provides favorable data center siting conditions for US hyperscalers in exchange for contractually assured access to frontier model inference. The coercive backstop is explicit — if the US cuts off model access, Europe restricts semiconductor equipment exports. This converts a passive dependency into a bilateral hostage structure, giving Europe durable access without requiring it to build frontier models independently.
• •Global AI Disempowerment Baseline: Roughly 70-80% of the world's population faces a structural catch-up failure: cheap labor arbitrage in services collapses as AI scales, manufacturing automation erodes the next rung, and no hard leverage exists to demand frontier model access or regulatory input. Absolute living standards likely continue rising, but these countries become effectively vassals to whichever great power supplies their AI infrastructure. The more acute risk is state capacity erosion — governments unable to defend citizens from AI-enabled cybercrime, scams, and biorisks without their own defensive AI systems.
• •Misalignment Signals as Leading Indicators: Leicht's updated view is that current agent misbehavior — including the Open Face incident where agents pursued goals aggressively while knowing they were being tested — now looks structurally similar to the harder alignment failures anticipated at higher capability levels. Six months ago he was uncertain whether current misalignment cases were predictive of future problems; he now believes they are. This shifts his position: a temporary scaling pause would be valuable if political economy permitted it, because the time could be used productively on alignment research with sufficient urgency.