Chinese AI Models Are Cheaper and Gaining Fast. What That Means for U.S. AI

Chinese AI startups including DeepSeek and Alibaba-backed Moonshot AI are releasing open-weight models priced far below U.S. alternatives, and the gap in usage is widening fast. In June 2026, monthly token usage of Chinese models ran 70% higher than for U.S. models. U.S. policymakers are now debating whether to restrict access to these models, but analysts warn that doing so could raise costs for American businesses and slow AI adoption broadly, while Nvidia and U.S. cloud providers also stand to lose revenue from serving Chinese model traffic.

What happened

Data point	Detail
Chinese vs U.S. token usage gap (June 2026)	Chinese models 70% higher, per Apollo Chief Economist Torsten Sløk
U.S. compute lead over Huawei	14,600 MW of energizable compute expected from U.S. chip companies in 2026, roughly 20x Huawei’s output
China chip independence forecast	2028, per Ryan Fedasiuk, American Enterprise Institute
U.S. frontier model lead	A couple of months ahead of Chinese rivals, per Sløk
Trump-Xi meeting	Expected September 24

The competitive pressure is not coming from hardware. It is coming from pricing and licensing. Open-weight models (models whose trained parameters are publicly released) allow any company to deploy them on their own servers or via a cloud provider without paying royalties. That strips a core revenue lever from closed-source U.S. labs like OpenAI and Anthropic.

Laila Khawaja, director of Gavekal Technologies, says Chinese models are already threatening the pricing power of Anthropic and OpenAI for non-frontier tasks. Those tasks, think document summarisation, customer support, data extraction, make up the bulk of real enterprise AI workloads.

The distillation dispute

A separate flashpoint has emerged around AI distillation, a training technique where a smaller model learns by querying a larger one. Michael Krastios, the White House’s Science and Technology advisor, posted on social media alleging that Moonshot’s Kimi K3 model used “industrial” large-scale distillation of Anthropic’s model to build its own capabilities. Moonshot plans to release Kimi K3 as a free download.

Krastios drew a line between “legitimate” distillation aimed at creating smaller, efficient models and what he called “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology.” China’s Ministry of Commerce pushed back, calling the allegations without factual or legal basis and labelling the framing “AI hegemonism.” Beijing urged both countries to work together on AI governance.

Why it matters for U.S. businesses

If Washington restricts access to Chinese models, American companies currently using those cheaper options would be forced back toward more expensive U.S. alternatives. That benefits OpenAI and Anthropic in the short term but puts U.S. firms at a cost disadvantage versus European and other global competitors who keep access to the lower-priced Chinese options.

There is also a downstream effect on chip demand. Nvidia and U.S. cloud providers earn revenue from running Chinese model workloads. Fewer deployments means less compute spend. As covered in our look at how OpenAI and Anthropic are competing for the AI lead, the battle between open-weight and closed-source models is shaping up as one of the defining commercial questions in AI right now.

Fedasiuk’s warning about Huawei is worth taking seriously. The U.S. tried to push Huawei out of 5G infrastructure during the first Trump administration and largely failed. He argues the U.S. needs to get American chips into more sockets globally before Chinese chip production scales in 2028, specifically to avoid repeating that outcome in AI infrastructure.

Our take

The 70% usage gap is the number that should focus minds. Export controls have slowed Chinese chip development, but they have not stopped Chinese labs from building competitive models. The lead the U.S. holds, a couple of months on frontier capability, is not a moat. It is a head start.

For businesses, the practical question is simpler than the geopolitics: if you are using a Chinese open-weight model for internal tasks today, watch Washington closely this autumn. A restriction before or after the September 24 Trump-Xi meeting could change your cost structure overnight. If you are building on a closed U.S. model and pricing feels high, that pressure from below is real and not going away regardless of policy outcomes.

If you are thinking about which AI tools and models to build into your own workflows, the open-weight versus closed-source question is directly relevant to cost and vendor lock-in. Our AI integration work regularly involves exactly this trade-off for clients, and right now the answer genuinely depends on your risk tolerance for geopolitical disruption as much as model quality.

What to do about it

1. Audit which AI models your business currently uses and note whether any are Chinese open-weight models (DeepSeek, Kimi, etc.).
2. Check the licensing terms for each model you deploy: open-weight licenses vary and some have commercial restrictions already.
3. Build a short list of U.S.-based alternatives for any critical workflows, so a restriction does not catch you without a fallback.
4. Watch the September 24 Trump-Xi summit and any executive orders or export control updates that follow it.
5. If you are evaluating new AI tooling, factor in the possibility of access restrictions when choosing between open-weight Chinese models and U.S. closed-source APIs. Talk to your technology team or agency about which architecture gives you the most flexibility.

The safest position right now is knowing exactly what you are running and having a tested alternative ready, not waiting for a policy announcement to start that audit.

Source: Bing News · Anthropic