THE UNITED STATES-CHINA Economic and Security Review Commission does not typically traffic in alarmism. So when the congressional advisory body released a report this week warning that Chinese open-source AI models now run inside roughly four out of five American startups building on open-source stacks, the figure landed with the dull thud of a problem that had been visible for months but politely ignored. The estimate originates from Andreessen Horowitz general partner Martin Casado, who observed that among startups pitching his firm with open-source architectures, there is about an 80% chance they are running on Chinese models — Alibaba's Qwen family, DeepSeek, or MiniMax. The number is a venture capitalist's field observation, not a rigorous census, but it reflects a structural reality that Washington's AI strategy has not yet reckoned with: America dominates the penthouse of artificial intelligence while China is quietly occupying every other floor.
The economics are not subtle. On OpenRouter, the aggregation platform widely used as a proxy for developer demand, Chinese models have overtaken American rivals in total token consumption since February. MiniMax and Moonshot charge $2 to $3 per million output tokens; Anthropic's Claude Sonnet 4.5 runs about $15 for the same volume — a near sixfold gap. MiniMax's M2.5 model alone saw usage surge 476% in a single month. For the agentic workloads that now define the industry's frontier — AI coding assistants, autonomous task-runners, multi-step research agents — the cost differential is existential, not incremental. A chatbot summarizing a document might consume 30,000 tokens; an AI agent can burn through 20 million on a minor coding task. One Hong Kong-based developer told the Financial Times he now spends $50 a day using Moonshot's Kimi for 80% of his work, reserving Claude for complex tasks. Running Claude alone would cost him $900 daily. Startups, which optimize for survival before sovereignty, will make the same calculation every time.
The comfortable lie
The reassuring narrative in Washington goes something like this: American labs hold the high ground. Claude Opus 4.6 and GPT-5.4 sit atop capability benchmarks; Chinese models cluster in the cheaper-but-weaker quadrant of the cost-performance scatter plot. As long as the United States leads on frontier intelligence, the argument runs, the token price war is someone else's problem. But this framing confuses the AI race's scoring system. Frontier models serve a narrow elite — researchers, complex enterprise deployments, the most demanding reasoning tasks. The vast majority of the world's AI inference, the billions of tokens consumed daily by coding assistants, customer service bots, document processors, and the swelling army of autonomous agents, does not require the best model. It requires the cheapest model that clears the quality bar. And Chinese labs are clearing it at a fifth of the price.
The USCC report frames the dynamic as a self-reinforcing loop. Cheap open-source models drive global adoption; adoption generates usage data and community fine-tuning contributions; that signal feeds back into the next model generation. Chinese-origin models already account for 41% of all Hugging Face downloads over the past year, according to the platform's own data. Alibaba's Qwen family has surpassed Meta's Llama in cumulative global downloads. The distribution advantage compounds over time in ways that chip export controls — Washington's primary lever — were never designed to address. As the USCC's vice chair Michael Kuiken put it, the assumption that restricting advanced semiconductors would be sufficient to contain Chinese AI competitiveness is looking increasingly fragile.
The security implications are not hypothetical. In March 2026, Germany's Federal Office for Information Security confirmed that a pilot project using DeepSeek-V3 for automated parliamentary briefing summaries had transmitted classified metadata — document classification levels, internal committee codes, timestamped access logs — to a DeepSeek server cluster in Shanghai. Not the documents themselves; the metadata. The German government did not realize the exposure until after the fact. Researchers at the Centre for International Governance Innovation have described the pattern as "infrastructure colonization": the broad adoption of Chinese large language models embeds foreign political assumptions and data pathways into the architectures of software, workflows, and institutional knowledge systems. Foreign Affairs went further, noting that researchers have demonstrated "sleeper agent" behaviors — potentially dangerous functions embedded in a model that surface only in specific contexts — could be inserted by developers into the LLMs now powering countless startup products.
Yet warning developers not to use Chinese models is roughly as effective as warning them not to use AWS in 2012. The incentive structure points one way. Siemens CEO Roland Busch said publicly this week that he saw "no disadvantages" to using Chinese open-source AI to train the company's industrial automation models, citing their cost advantage and ease of customization. When the CEO of a $150 billion German industrial conglomerate shrugs off the risk, telling a seed-stage startup in San Francisco to pay five times more for patriotic reasons is not a viable national security strategy.
The deeper problem is that America's AI policy treats the technology like a weapons system — something to be controlled through export restrictions and classification — when it is increasingly behaving like a commodity. China understood this first. Alibaba reorganized its entire AI operation in mid-March around a concept it calls "Token Hub," consolidating five units under CEO Eddie Wu with a 380 billion RMB ($53 billion) three-year investment commitment and a mission statement that reads like an energy company's: "create tokens, deliver tokens, apply tokens." Beijing designated "computing-electricity synergy" a national priority in its 2026 work report, explicitly linking cheap energy policy to AI competitiveness. The Chinese government is treating tokens the way it once treated solar panels and EV batteries — as a strategic commodity where scale and cost leadership reshape global dependency.
If the pattern holds, the United States risks reprising a familiar dynamic: inventing the technology, leading on performance, and watching the volume market migrate overseas along with the leverage that comes with it. The fact that Claude and GPT-5 remain the most capable models in the world is cold comfort if the global default inference layer runs on Qwen and DeepSeek. The best chipmaker in the world still needs customers. The question is whether Washington will treat cheap Chinese tokens as a trade problem, a security problem, or simply the price of an open internet — and whether it will decide before the dependency becomes too deep to unwind. ■
For more, join 75,000 subscribers getting tech's favorite brief here
