Chinese models reportedly carry 55% of Vercel's enterprise AI traffic

What the report says

As Latin Times reported, Chinese AI models handled 11% of token traffic on Vercel's enterprise developer platform in January 2026. By August, that share had grown to 55%. A token is a small chunk of text that an AI model reads or writes, and it is the usual unit for billing and for measuring usage.

The article credits two models for most of the jump: DeepSeek V4 and Moonshot AI's Kimi K3. Vercel is a service developers use to build and host web apps, so this number describes what developers on that one platform chose to run. It is not a measure of the whole market.

A caution on the number. We have not seen Vercel's own data. The figure comes from Latin Times, and the word "controls" in its headline is stronger than what the traffic share alone shows. Traffic does not tell us revenue, quality or how many customers are paying.

DeepSeek V4: cheap and open

DeepSeek released V4 and V4-Pro on April 24, 2026. V4-Pro is a "mixture-of-experts" model, meaning it holds a very large set of specialized components but only switches on a few for each piece of text. It stores 1.6 trillion parameters (the adjustable numbers that make up a model) and uses 49 billion at a time, which keeps the computing cost down.

Both V4-Pro and a smaller V4-Flash came out under the MIT License, one of the most permissive licenses around. Anyone can inspect, change and use the weights commercially. DeepSeek's hosted service for V4-Pro costs about $0.44 per million input tokens, a price it made permanent on May 22 after a promotional period.

The report says this puts output costs roughly 17 times below comparable OpenAI models. That is the outlet's comparison on output tokens, so check current price lists before relying on it.

The report also notes a hardware choice. DeepSeek built V4 to run on Huawei's Ascend chips and skipped the early access to the model that Nvidia and AMD usually get for tuning. Huawei said it supported V4 on launch day.

Kimi K3 and the size record

Moonshot AI opened Kimi K3 to paying API users on July 16 and published the weights on July 27 under its own license. The model has 2.8 trillion total parameters, with 104 billion active per token, which Latin Times calls the largest open-weight model. "Open-weight" means the trained model files can be downloaded and run on your own hardware. It also handles images as input and has a context window of one million tokens, which is roughly how much text it can consider at once.

On the Artificial Analysis Intelligence Index v4.1, K3 ranks fourth. Only Claude Fable 5 and two Sol entries from OpenAI rank above it. In GPU-kernel tests, which measure how well a model writes low-level code for graphics chips, it matched Claude Fable 5 on long tasks. It beat Claude Opus 4.8 and GPT-5.5, and scored competitively but behind on some tests against GPT-5.6 Sol.

The practical point is that a government or company can now host a top-tier coding model itself, without recurring fees to a U.S. cloud provider. Mohammed Soliman of the Middle East Institute told Rest of World the model "does change the calculation" for governments building their own AI infrastructure.

Why open release spreads so fast

The report argues that open release is a workaround, not generosity. Chinese labs lack the worldwide cloud networks of Google or Microsoft, so serving global demand from their own servers is expensive. Publishing the weights lets users run the models on hardware they already have.

The numbers cited back that up. Chinese open models, led by Qwen, DeepSeek and Kimi, make up about 30 percent of global open-source AI usage, up from under 2 percent at the end of 2024. Alibaba's Qwen family passed 3 billion total downloads by August and, per the report, overtook Meta's Llama as the most-downloaded family on Hugging Face.

Beijing's export draft

On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology published a draft listing 23 AI technologies that could be added to the national export-control list. It covers language models, autonomous AI agents, advanced chip designs and large computing clusters. A temporary pre-export registration rule began the same day.

For models, the draft applies if any one of three tests is met: 2 trillion or more parameters, training compute above 10²⁶ operations, or a top-10 global ranking on performance or safety benchmarks. Public comment runs through November 15.

Importantly, the draft exempts technology already public, which includes Kimi K3's weights. Serving users abroad through apps, websites or APIs is also excluded. What the final rules say, and whether they slow future releases, is still open.