.png)
The most direct evidence came in February 2026. Anthropic reported, in findings covered by The Decoder, that DeepSeek, Moonshot and MiniMax had generated more than 16 million exchanges with Claude through around 24,000 fraudulent accounts. Computing’s breakdown shows that MiniMax accounted for over 13 million of these exchanges, Moonshot for more than 3.4 million and DeepSeek for about 150,000.. The requests focused on coding, reasoning and the ability to carry out complex tasks autonomously, precisely the skills on which US labs have spent most and for which businesses pay the highest prices. In a memo to the US congressional comitte, OpenAI made similar accusations against DeepSeek the same month, and Google disclosed an attempt to copy its Gemini model using more than 100,000 prompts. These remain company allegations, however, and nobody has measured how much the Chinese labs actually gained.
Source: Anthropic
The spending gap adds weight to the argument. US hyperscalers are expected to pass $1tn of annual investment in 2027, based on estimates from Evercore and Bank of America. By comparison, ByteDance, China’s largest spender, plans to invest about $28 billion in 2026, EqualOcean reports. Labs such as DeepSeek or Moonshot have no comparable infrastructure at all. Supporters of the subsidy argument point out that most of the cost of building an AI model comes before the final version. Labs spend heavily on experiments that fail, on collecting data and on testing. A lab that learns from a rival’s answers skips these costly steps and goes straight to the result, while the rival has paid for all of them.
Source: Company guidance
Yet the results are close. Research institute, Epoch AI, notes that Chinese models have trailed the best American ones by an average of seven months since 2023. However, open-weight models in general, including those developed in the United States, trail the leading closed models by a comparable margin, suggesting that the delay is not specific to China. Their adoption is also growing rapidly. On OpenRouter, a platform that lets developers switch between AI models, Chinese models went from about 1% of usage in late 2024 to around 46% by mid-2026, according to Forkast, citing CNBC. The reason is price. OpenRouter’s own analysis finds these models 60% to 90% cheaper than leading US alternatives. Even American companies have adopted them. In March 2026, the coding tool Cursor launched a model built on Moonshot’s Kimi, without disclosing it at launch.
Source: OpenRouter
The idea that Chinese companies do not invest in AI does not survive a closer look. Alibaba and Tencent together spent more than RMB 120bn, around $17bn, in the second quarter of 2026 alone, according to TMTPost. The famous claim that DeepSeek trained a leading model for just $5.6mn covered only the final training run, as DeepSeek itself stated. The research firm SemiAnalysis estimated that the company’s servers alone were worth around $1.6bn, nearly 300 times that figure. Chinese labs spend far less than American ones, but they are far from working on a shoestring budget.
Ideas also travel in both directions. Several of the most important techniques for making AI cheaper to train and run were developed by DeepSeek and published openly, and US labs have since adopted them. CSET data cited by Al Jazeera show that China also produces more AI research than the United States, with around 27% of global publications in 2024 against 12% for the US. Limited access to advanced chips played a role too. Because US export controls restrict China’s access to the most powerful hardware, Chinese engineers were pushed to get more out of less, making cheap models partly a genuine engineering achievement.
Source: Epoch AI
In addition, most of this American spending is safe from copying. The majority of hyperscaler investment does not go into creating new AI models. It funds data centres used for cloud services rented to other companies, for advertising systems and for running AI tools for paying customers. Google Cloud alone has more than $460bn of contracted future revenue, and Meta attributed part of its 33% revenue growth in the first quarter of 2026 to AI-driven ad targetingas reported by Yahoo Finance. A Chinese lab copying a model’s answers takes nothing away from these businesses. Only the money spent on developing new models is truly exposed, and it represents a fraction of the headline figures.
It is also worth asking who captures the benefit of cheap Chinese models. Of 178 large open-weight models released by Chinese labs in 2026, 59% carry the Apache 2.0 licence and 22% the MIT licence, both of which allow free commercial use. The main winners are therefore developers and businesses around the world, including American start-ups. In 2025, Andreessen Horowitz, one of Silicon Valley’s largest venture capital firms, reported that up to 80% of the start-ups pitching it with open models were building on Chinese ones. Seen this way, Chinese labs subsidise the global software industry as much as US companies subsidise Chinese labs.
Copying is also becoming harder. US labs are deploying tools that detect suspicious usage patterns and can deliberately degrade the answers sent to suspected copiers. The scale of the alleged campaigns also varies widely. The 150,000 exchanges Anthropic attributes to DeepSeek are small compared with the data needed to train a modern model, which suggests that most of its progress came from its own work..
Finally, American labs continue to capture the vast majority of AI revenue. Anthropic’s annualised revenue passed $65bn at the end of July 2026 and OpenAI’s reached about $40bn. By comparison, Epoch AI estimates that no other lab, Chinese ones included, earns more than $1bn.
Source: Bloomberg, Reuters, Epoch AI