Chinese open-source models capture over half of enterprise AI traffic

A shift in global AI traffic

Chinese AI models have rapidly captured a significant portion of the global developer market. Between January and August 2026, these models accounted for 55% of all enterprise token traffic on Vercel's developer gateway. This represents a massive jump from just 11% recorded only seven months earlier.

The surge is primarily driven by two specific models: DeepSeek V4 and Kimi K3. These models have allowed Chinese labs to challenge U.S. dominance by providing high-performance options that are both affordable and accessible. This shift has unsettled U.S. policymakers and fundamentally redrawn the competitive map for global AI infrastructure.

This growth is not attributed to state subsidies or market coercion. Instead, it is the result of openly released code and weights being priced lower than any American frontier lab. These models are spreading into production stacks faster than regulators have been able to respond to the trend.

DeepSeek's efficiency and pricing

DeepSeek launched its V4 and V4-Pro models in April 2026 to reposition the open-weight tier. The V4-Pro model uses a mixture-of-experts system, which stores 1.6 trillion total parameters but only activates 49 billion of them during inference. This design allows the model to deliver near-frontier output with much lower compute overhead than traditional dense models.

The lab released these models under the permissive MIT License, allowing developers worldwide to inspect, modify, and deploy the weights for commercial use. The API for V4-Pro settled at $0.44 per million input tokens in May 2026. This pricing puts the cost of output tokens at roughly 17 times lower than equivalent models from OpenAI.

DeepSeek also made a notable hardware choice by engineering V4 to run natively on Huawei's Ascend SuperPoD products. The lab withheld standard pre-release access from Nvidia and AMD, breaking a long-standing industry convention. Huawei confirmed full support for the model on its Ascend 950-based supernode clusters on the same day of the launch.

Kimi K3 sets a new scale

While DeepSeek focused on pricing, Moonshot AI's Kimi K3 reset expectations regarding model scale. Released as an open-weight system in July 2026, Kimi K3 features 2.8 trillion total parameters. It is currently the largest open-weight system in existence and is the first model Moonshot describes as entering the open 3T-class.

Benchmarks place Kimi K3 as a high-performing contender, ranking fourth on the Artificial Analysis Intelligence Index v4.1. It trails only Claude Fable 5, GPT-5.6 Sol, and a second Sol configuration. In evaluations for long-horizon task completion, Kimi K3 matched Claude Fable 5 and outperformed both Claude Opus 4.8 and OpenAI's GPT-5.5.

The availability of Kimi K3 means sovereign governments and enterprises can now download and self-host a frontier-tier coding model. This allows them to avoid paying recurring fees to U.S. cloud providers. Mohammed Soliman of the Middle East Institute noted that an open, high-quality model like Kimi K3 does change the calculation for those building domestic infrastructure.

The logic of open-source growth

China's strategy of releasing model weights is a structural solution to an infrastructure deficit. Because Chinese labs lack the global cloud footprint of companies like Google or Microsoft, they cannot serve worldwide demand through hosted compute alone. By releasing weights, they allow their technology to propagate on hardware that users already own.

This strategy avoids the need for Chinese labs to build out massive data centers for global deployment. The results are measurable, with Chinese models now accounting for approximately 30 percent of global open-source AI usage. This is a significant increase from less than 2 percent at the end of 2024.

Alibaba's Qwen family has seen massive success, crossing 700 million downloads on Hugging Face in early 2026. By August, the family surpassed Meta's Llama as the most-downloaded model family on the platform. These models are becoming the primary choice for developers from Bangalore to San Francisco.

New export controls and exemptions

The rapid expansion of Chinese AI has prompted a new regulatory response from Beijing. In mid-October 2026, Chinese authorities for commerce and science jointly published a draft catalog of 23 AI technologies for export control. This includes large language models, autonomous agents, and specific chip designs or computing clusters.

The proposed restrictions apply to systems meeting certain criteria, such as having 2 trillion or more parameters or reaching a global top-10 benchmark ranking. However, the proposal includes a specific exemption for technology already in the public domain. This means models like Kimi K3, which were released on Hugging Face in July, would not fall under these new restrictions.

This move shifts the focus from what China can receive to what it permits to leave the country. The outcome of the public comment period, ending November 15, will determine if the open-source offensive continues. It will decide whether Beijing will preserve the distribution model that made Chinese AI competitiveness visible to the entire world.