America’s AI Labs Are Under Threat From Cheap Chinese Rivals

Abstract

America’s leading AI developers pour vast sums into each frontier release, only to find competitors have already caught up — and the most dangerous of those competitors give their models away for free. Chinese open-weight models are now months, not years, behind the frontier, run far cheaper per task — DeepSeek’s best at $0.04 a job against Fable’s $2.75 — and are widely judged “good enough” for many tasks. Washington’s unpredictable approach to granting access to American models is itself pushing customers toward those alternatives.

Framing

• Published in the Business section under the rubric Weights of the world, 21 July 2026.
• Standfirst: demand for open-weight models is soaring.
• The core squeeze: it would be bad enough if only rivals that charge for their technology were nipping at their heels, but Anthropic and OpenAI must also sprint to stay ahead of labs that give their models away for free.

Two releases in one week

• 16 July — Moonshot AI, a Chinese startup, released Kimi K3: only slightly less capable than Fable or Sol, the latest offerings from Anthropic and OpenAI respectively.
  • Its parameters (weights) were to be made publicly available at no cost later that month, meaning it can be downloaded and run on any server capable of doing so.
• 19 July — Alibaba unveiled a preview version of a new addition to its Qwen family, which the company says is second only to Fable.

How far behind is the frontier?

• Per analysis by research firm Epoch AI, the most powerful open-weight models — which typically come from China — are a matter of months behind the technology’s frontier.
• Adoption is spreading rapidly. On OpenRouter, a popular model marketplace:
  • In May, customers used systems from the top American labs about as much as those from the top Chinese ones, measured by token consumption.
  • The following month, use of American models was up by two-thirds — but had tripled for Chinese ones.

The reaction in Washington

• Consternation not just among America’s leading model-makers but in its government, which fears the country losing its AI lead and creating a new opening for hacking.
• Some within the Trump administration are said to be pushing measures to slow the spread of Chinese models:
  • adding their makers to trade blacklists;
  • warning American companies against their use.
• Grumbles that Chinese labs use American models to train their own systems — “distillation”.
• The counterproductive effect: the American government’s unpredictable approach to granting access to advanced models such as Fable and Sol is itself spurring adoption of China’s open-weight alternatives.

The open-weight field before K3

• GLM-5.2, launched in June by Z.ai, is better at so-called agentic tasks — such as creating software — than anything currently available from Google.
• Models from DeepSeek and MiniMax, two other Chinese labs, have also proved popular.
• Per a report by Mozilla: open-weight models remain behind the frontier in reasoning and complex agentic work, but are now widely perceived as “good enough” for many tasks.

Cost

• Cost matters more as companies grapple with the rising expense of AI. Artificial Analysis tracks the cost for selected models to perform each task in its test battery:
  • DeepSeek’s most capable system: $0.04 per job.
  • Kimi K3, which is not yet open-weight: $0.95 per job.
  • Fable users can expect to pay $2.75.
• Why open weights are cheaper: more or less any cloud provider can sell access to an open-weight model, driving down profit margins.
  • Example: DeepSeek offers its v4 flash model at $0.28 per million tokens, while the cheapest third-party provider offers it for $0.18 per million.
• Efficiency by necessity: many open-weight models are trained with a specific focus on running efficiently, partly in response to the difficulty Chinese labs face accessing computing power under American restrictions on advanced chips.
  • “Mixture of experts” designs — activating only the part of an AI system relevant to a query — have been used extensively by open-weight developers, with DeepSeek and Alibaba making big strides.

The deeper appeal: dependence risk

• For three weeks Fable was unavailable globally, owing to a hurried attempt by America’s government to withhold the model from all but American citizens.
• That prompted many companies to examine their dependence on individual labs, and encouraged experimentation with open-weight alternatives that can be accessed through many providers and perhaps run on internal servers.
• Raffi Krikorian, Mozilla’s chief technologist, describes the episode as a “shot across the bow”.

The catch

• China’s open-weight models may not be entirely dependable for long.
• Its government is said to be looking at ways to control foreigners’ access to its technology, which could involve restricting the ability to download its open-weight models outside China.
• America’s labs would surely be delighted.