Two frontier-AI superpowers, two mirror-image governance systems.
The United States and China are the world's two frontier-AI superpowers, and they have built almost mirror-image governance systems to compete with each other. The US, especially since January 2025, has stripped back precautionary federal rules in favour of a deregulatory, export-and-infrastructure-led strategy; the only consistently active federal AI body is NIST's CAISI, whose entire toolkit is voluntary testing agreements with the same labs it's meant to be checking.
China has done the opposite: no single comprehensive law, but a fast-moving stack of binding, enforcement-backed rules that keeps expanding into new categories — agentic AI in July 2026, AI companion apps in April 2026 — under the Cyberspace Administration of China. Each side treats the other's chip and compute posture as the real battlefield: export controls on one side, chip self-sufficiency and open-weighting on the other.
Sectoral and executive-led — no binding federal frontier-model law. The prior administration's safety-testing executive order was repealed; the current administration's America's AI Action Plan prioritises infrastructure and export over precautionary rules. States (Colorado, California) remain the most active source of binding AI rules.
Targeted, stacked binding rules with no single comprehensive statute, but enforcement teeth throughout — the CAC-led approach just added its first dedicated framework for AI agents (July 2026) and interim rules for AI companion apps (April 2026).
Both avoid a single comprehensive AI law, but for opposite reasons and with opposite practical effect. The US avoids one because the current administration treats precautionary regulation as a competitive liability; China avoids one because a single static law would be too slow to extend to new product categories — its 'no comprehensive law' looks more like 'many fast, binding laws' than genuine light-touch governance.
Low-to-moderate direct government involvement in AI development itself; industrial strategy runs through infrastructure and export policy (the 'Stargate' initiative, America's AI Action Plan) rather than state-directed model-building. The frontier labs — OpenAI, Anthropic, Google DeepMind, Meta, xAI — are all private and US-headquartered.
Very high state involvement — the AI+ Initiative directs AI integration across manufacturing and public services, and DeepSeek's open-weighting of frontier-competitive models is widely read as a deliberate soft-power strategy, not just a technical choice.
The US leads on private-sector frontier capability and lets companies compete largely on their own terms; China leads on rapid state-directed diffusion into the economy and treats open-weighting as strategic statecraft. DeepSeek's releases are a direct challenge to the US assumption that export controls would meaningfully slow Chinese frontier capability — that a Chinese lab could release something frontier-competitive despite chip restrictions has reshaped the export-control debate inside the US itself.
NVIDIA's near-monopoly on frontier AI chips is a central US strategic asset; export controls restrict advanced NVIDIA and AMD chip sales to China specifically to slow Chinese frontier capability.
US export controls have accelerated a state-backed push for domestic semiconductor self-sufficiency, led by SMIC's foundry capacity and Huawei's Ascend AI chips — real progress, but still lagging the leading edge by multiple process-node generations.
This is the one dimension where the two countries are in direct, explicit conflict rather than parallel divergence: US chip policy is designed specifically to constrain China's compute, and China's chip policy is a direct response to that constraint. Whether the controls have 'worked' is genuinely contested — DeepSeek's frontier-competitive open-weight models are the strongest evidence for the sceptics' case that controls only accelerated Chinese self-sufficiency without preventing capability gains.
No dedicated AI ministry; the White House AI & Crypto Czar role has sat vacant since March 2026, leaving CAISI (a NIST body) and OSTP as the most consistently active federal AI institutions — CAISI's power is entirely voluntary testing agreements, not statutory authority.
The Cyberspace Administration of China is a genuine lead regulator with real enforcement power, co-issuing binding rules jointly with NDRC, MIIT, the Ministry of Public Security and SAMR depending on the AI category.
China's institutional structure is built to issue and enforce binding rules quickly across agencies; the US's is built around voluntary cooperation with the very labs it monitors, and currently lacks even a single accountable political owner for AI policy inside the White House.
Frontier model competitiveness and compute infrastructure over precautionary regulation.
Content and platform-safety control extending into agentic and companion AI.
The US's stated focus is external — winning the capability and infrastructure race; China's is substantially internal — controlling what AI-mediated content and interaction looks like for its own population, even as it also competes on frontier capability via DeepSeek and others.
Shared
US
China
The US is betting that private-sector dynamism and a light federal touch produce the most capable frontier models, and that export controls alone can preserve a compute advantage over China — a bet complicated by CAISI's growing reliance on voluntary agreements that could be unwound without any legislative action, and by DeepSeek's demonstration that chip constraints don't fully block capability progress.
China is betting that fast, binding, stacked regulation lets it control domestic AI-mediated risk — companion apps, agentic systems — without slowing frontier progress, using open-weighting as a soft-power tool to offset compute disadvantages. That bet depends on continued success stories like DeepSeek and on its domestic chip industry eventually closing the process-node gap.