The US Is Winning the AI War It Declared — and Losing the One That Matters

Microsoft built a billion-dollar business selling OpenAI's models to Chinese companies through Azure, with ByteDance as its largest customer ^[1]. Now OpenAI's most important commercial partner is exploring something that would have been unthinkable a year ago: a self-hosted, fine-tuned version of DeepSeek-V4 to power a cheaper tier of Copilot ^[2].

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

There are two AI wars, and the United States is winning the wrong one. The first is the war Washington declared — a restrictions war fought with export controls, executive orders, and investigations. The House Foreign Affairs Committee has advanced more than 20 bills tightening chip-equipment exports, including the MATCH Act, which would give Japan and the Netherlands 150 days to align or face unilateral US enforcement ^[3]. Trump signed an executive order requiring companies to provide the government access to frontier models up to 30 days before release ^[4]. The Commerce Department ordered Anthropic to disable its Fable 5 and Mythos 5 models for foreign nationals ^[5]. Anthropic banned China-headquartered companies and their subsidiaries from its platform ^[6]. Apple geoblocked ByteDance apps for US iOS users ^[7]. House committees launched investigations into American firms Cursor and Airbnb for using low-cost Chinese AI models ^[6]. On the surface, the US is doing everything it said it would do. But the war that determines AI leadership is adoption — who builds with whose models, whose infrastructure processes the world's tokens, whose ecosystem becomes the default. And on that front, the numbers have turned. In a single week at the end of March, Chinese AI models processed 12.96 trillion tokens globally. US models processed 3.03 trillion. All six of the world's most-used AI models are now Chinese ^[8]. On OpenRouter, a platform that routes API calls for American firms, Chinese models now account for 58% of tokens processed — up from under 10% at the start of 2025. DeepSeek is the most widely used Chinese model among US companies ^[9]. The US-China Economic and Security Review Commission reported that roughly 80% of US AI startups now use Chinese open-source base models, with Alibaba's Qwen surpassing Meta's Llama in cumulative downloads ^[10]. The mechanism is not mysterious. In May, DeepSeek permanently cut V4-Pro API prices by 75%, making the model 12 to 19 times cheaper than GPT-5.5 or Claude Opus 4.7 for equivalent tasks ^[11].

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

The price gap is structural. DeepSeek's efficiency gains come from Huawei Ascend 950 processors, which pass through cost advantages that US labs, paying Nvidia margins and burning scarce American electricity, cannot match. And electricity is the second mechanism. US data center power demand is projected to reach 9.5% to 15% of total US electricity consumption by 2030 ^[12]. Over 100 local moratoriums on data center construction have been proposed across the country, with communities citing grid instability and water contamination — including an 18-month pause in Wisconsin after a $1 billion Meta facility contaminated drinking water ^[13]. In March, Anthropic throttled Claude usage during peak hours, reducing session limits for Pro users ^[14].

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before. — Thariq Shihipar

Capacity constraints push users toward alternatives at the moment competition is fiercest. And the Stanford-led 2026 AI Index Report removed the last excuse: the performance gap between US and Chinese frontier models has effectively closed, with Anthropic's leading model holding just a 2.7% edge across major benchmarks ^[15]. When the quality difference is 2.7% and the price difference is 12 to 19 times, the market does not need a lecture on patriotism. The restrictions war is also producing a backfire Washington's own architects warned about. China's AI chip self-sufficiency has risen from 10% in 2021 to 41% in 2026, with Morgan Stanley projecting 86% by 2030 ^[16]. Huawei's rotating chairman Xu Zhijun made the causal chain explicit.

If the US hadn't forced our country, our companies, and our industry, we wouldn't have done something like this. — Xu Zhijun

Nvidia CEO Jensen Huang had warned of exactly this outcome.

if they can’t buy from us, they’ll build their own — and then we’ll have a competitor we wouldn’t have otherwise created. — Jensen Huang

DeepSeek excluded Nvidia and AMD from the optimization phase of its V4 model entirely, giving Huawei and HiSilicon a multi-week head start on a parallel Chinese tech supply chain ^[17]. Zhipu AI trained its GLM-5.2 model on 100,000 Huawei Ascend 910B processors, bypassing Nvidia completely, and produced a model that matches frontier US systems in coding at less than one-tenth the cost of Anthropic's Fable 5 ^[5]. Alibaba unveiled its Zhenwu M890 chip as a domestic alternative to Nvidia's H100, delivering 560,000 units to more than 400 customers across 20 industries ^[18]. Both sides are now walling off their ecosystems. On July 1, China imposed its own sweeping AI export controls: blocking Meta's $2 billion acquisition of Manus, barring AI professionals from overseas travel, and requiring government sign-off before firms accept American capital ^[19]. Alibaba banned employees from using Anthropic's Claude Code after detecting hidden tracking code that identified users in China or affiliated with Chinese AI labs ^[20]. The US, meanwhile, is investigating its own companies — Cursor and Airbnb — for using Chinese models, while a nonprofit called Build American AI pays social media influencers to frame Chinese AI as a national security threat ^[6][21]. But the bifurcation is not symmetric. When the performance gap is 2.7% and the price gap is 12 to 19 times, the walled garden that costs more is the one that shrinks. US officials continue to describe the contest in moral terms. Treasury Secretary Bessent calls America "an AI superpower" with China "trailing substantially" ^[22].

We are leading China by a lot. — Donald Trump

The Stanford number — 2.7% — is the hinge of the whole contest. The cheaper side now offers near-parity. The US is winning the war it can restrict and losing the one it cannot.