China has overtaken the US in AI performance and cost efficiency. Moonshot AI’s Kimi K3 now ranks first in six of seven technical domains, including coding and data analytics, while costing as little as $0.50 per million tokens - up to 50 times cheaper than OpenAI or Anthropic.

This shift wasn’t accidental. US chip bans forced Chinese engineers to optimize for efficiency. The result: leaner models that outperform heavier, guardrail-laden American counterparts. As Saagar Enjeti noted, the sanctions backfired - Beijing turned constraint into advantage.

The economic pressure is immediate. Frontier labs like OpenAI and Anthropic face IPOs while drowning in hardware debt. If Chinese models offer 95% of the capability at 1% of the cost, the US business model collapses. Alex Finn warns this creates a "gasoline price" disparity - forcing American firms to pay $800 per gallon while the rest of the world pays $8.

The irony deepens in cybersecurity. When an AI agent attacked Hugging Face, US models like Fable 5 refused to assist in analyzing the exploit, citing safety protocols. Developers turned to GLM 5.2, a Chinese model, to patch 15 critical bugs. As Sriram Krishnan put it: US defenses are being outsourced to foreign intelligence.

Policy responses are split. The White House debates banning Chinese models, with OpenAI and Anthropic lobbying for protection. Meanwhile, Treasury Secretary Scott Bessent floats sanctions on Moonshot AI, accusing it of distilling Anthropic’s Fable to build K3. But Commerce Department officials argue for competing with open US models instead of isolation.

The commoditization of intelligence is accelerating. Krishnan observes that leading open-weight models are now Chinese - Kimi K3, Qwen - and developers are adopting them for tasks from code generation to exploit research. The raw model is no longer the moat. Value is shifting to the "harness": proprietary interfaces, workflows, and routing layers that manage cost and context.

Coinbase’s internal AI gateway, which routes tasks to the cheapest capable model, exemplifies the new logic. Ory Goan notes only 0.1% of companies can do this at scale. For the rest, the choice is stark: adapt or rely on subsidized foreign intelligence.

The race is no longer just about who builds the smartest model. It’s about who controls the stack, the access, and the rules. China’s open release isn’t altruism - it’s a calculated move to dominate the next layer of digital infrastructure. As Anand Kappen warns, once they solve their compute bottlenecks, the open door will close.