The AI Race America Isn't Running

In May, DeepSeek permanently cut the price of its V4-Pro model by 75 percent. The company attributed the move to efficiency gains from Huawei Ascend 950 processors, and the explanation it gave distinguished the cut from a promotional loss leader. ^[1]

It is an efficiency gain being passed through. — Sanchit Vir Gogia

That same spring, Anthropic was moving in the opposite direction. In March, it quietly reduced Claude session limits during peak hours, a throttling analysts read as a push to move power users from fixed subscriptions to revenue-guaranteed API consumption. ^[2] In April, it blocked third-party tools like OpenClaw from Claude Pro and Max subscriptions entirely. The company's head of Claude Code offered a rationale. ^[3]

We've been working hard to meet the increase in demand for Claude, and our subscriptions weren’t built for the usage patterns of these third-party tools. — Boris Chernyshov

Two companies facing the same cost pressure. One passed the savings to customers. The other raised the walls and raised the price. That divergence is not a tale of corporate character. It is a structural fact: the proprietary, high-margin business model that funds American AI cannot match the economics of Chinese open-weight models, which treat efficiency as something to distribute rather than capture. The competition has shifted. For years, the AI race was measured by a single question: whose model is the smartest? Stanford's 2026 AI Index found the US-China performance gap has "effectively closed," with Anthropic holding just a 2.7 percent advantage as of March. ^[4] Artificial Analysis, using a different methodology, finds a stable three-to-nine-month intelligence lag for Chinese models. ^[5] The two assessments disagree on the margin. They agree on what the margin means: cost, not capability, is now the decisive competitive variable. And on cost, the numbers are not close. DeepSeek's latest V4 Flash trails OpenAI's GPT-5.6 Luna by a single Intelligence Index point while costing 60 percent less per task — even after OpenAI cut its own prices by 80 percent. ^[6] The V4-Pro runs 12 to 19 times cheaper than GPT-5.5 and Claude Opus 4.7 for equivalent work. ^[1] The market has noticed. Chinese models have outpaced American models in global weekly token consumption for five consecutive weeks, handling 12.96 trillion tokens to America's 3.03 trillion. All six of the world's most-used AI models were Chinese. ^[7] Chinese open-weight models surpassed American ones in Hugging Face library downloads in January, meaning the adoption runs through the developer ecosystem, not just end-user chat. ^[8] American companies are adopting them. Pinterest uses DeepSeek R-1 for its recommendation engine; Airbnb runs Alibaba's Qwen on customer service. Pinterest's CTO reported that open-source techniques yield in-house models 30 percent more accurate than off-the-shelf proprietary options. ^[9] Microsoft is exploring a self-hosted, fine-tuned version of DeepSeek-V4 to power a lower-cost tier of Copilot Cowork, after OpenAI and Anthropic shifted to usage-based pricing that made high-volume enterprise users prohibitively expensive. A Microsoft VP described the bind. ^[10]

We have users who do hundreds of tasks a week, which is great, they’re way productive, but the consequence is the costs can go very high. — Charles Lamanna

Tencent's new Hy3 model hit number one on the OpenRouter global usage leaderboard in its first week, with API call volume up 68-fold over the previous generation. The company is giving it away free on WorkBuddy through August to drive enterprise adoption. ^[11] Tencent's total AI investment pledge for 2026 is just over $5 billion. ^[12] American tech giants, meanwhile, are projected to spend nearly $900 billion on AI infrastructure this year — up from $450 billion in 2025, with $1.4 trillion expected in 2027. They have borrowed over $400 billion to fund the expansion, and the ultimate financial returns remain, in the careful language of analysts, "deeply uncertain." ^[13] The capital is deployed on the assumption that compute abundance is the competitive moat. But the competition has moved to efficiency, and on that field $900 billion buys remarkably little. The American response to this shift has been a series of moves that all target the old race — the intelligence contest — while the new one runs on cost. OpenAI launched GPT-5.4 Mini and Nano in March, describing them as "our most capable small models yet," a direct response to the efficiency pressure. But the models stay inside OpenAI's proprietary pricing layer. ^[14] The efficiency is real; the economics are not. Anthropic's 2026 enterprise strategy has been to build lock-in features — Claude Desktop for Enterprise, private marketplaces, enterprise plugins — while simultaneously throttling session limits and blocking third-party tools, extracting more revenue from existing customers rather than competing on price. ^[2][3] A tech-backed nonprofit called Build American AI is paying social media influencers thousands of dollars per video to promote American AI and frame Chinese models as a national security threat. ^[15] US officials describe the contest in moral and existential terms — "superhero versus supervillain," in the words of Representative Brian Mast — even as the performance gap closes and Chinese models handle two-thirds of global AI usage by token volume. ^[16] Treasury Secretary Bessent calls America "an AI superpower" and says China is "trailing substantially." Trump claims "we are leading China by a lot." ^[16] The rhetoric is designed for an intelligence race the data says is over. The spending is designed for a compute race the market has already routed around. The mechanism is straightforward. Chinese open-weight economics pass efficiency gains through to customers as price cuts. The proprietary American model requires capturing those same gains as margin — because the labs have investors, because the infrastructure debt must be serviced, because the business model was built for a world where superior intelligence commanded a premium. That world is gone. The business model remains. On Monday, the Trump administration removed a significant regulatory barrier, exempting open-weight AI models — explicitly including Chinese rivals like Moonshot's Kimi K3 — from government safety testing. National Cyber Director Sean Cairncross argued a strict regime would "strangle growth" and be "obsolete 48 hours" after implementation. ^[17] The exemption does not clear every obstacle. House committees are investigating American firms for using Chinese models. ^[18] Trump is weighing further sanctions on Alibaba, Moonshot, DeepSeek, and ByteDance, though he has also warned against restricting in ways that would leave the US "second to China," saying "we have to be careful in both ways." ^[19] Apple has geoblocked ByteDance apps for US iOS users. ^[20] But the safety-testing exemption removes the one barrier that could have slowed adoption at the model level, and the investigations target individual firms, not the API-access channel through which enterprises actually use these models. There is one development that could alter the trajectory, and it comes from Beijing. China is considering national security laws to restrict overseas access to its most advanced models — limiting API access, blocking model-weight downloads, and designating AI technology leaks as national security offenses. ^[18] If enacted, such restrictions would cut off the very channel driving global adoption of Chinese models. It would be the one thing that saves the United States from a competition its business model cannot enter. The irony is that the only effective defense against Chinese AI dominance would have to be imposed by China itself.