Inside Chinese AI: ByteDance, Z.ai, humanoid robots requiring passport ID and the server GPU card market

The co-founder of GPTunneL — a platform for accessing Chinese AI models — spent eight days in China in early July 2026, meeting sequentially with teams from ByteDance, Z.ai, SiliconFlow, and MiniMax. This is an inside report: negotiations every day, the server card market, and a shop where humanoid robots are sold by passport.

What ByteDance Is Preparing for July 10, 2026

ByteDance — the developer of TikTok and the video generative model Seedance — announced the release of a new product on July 10, 2026. The authors intentionally withhold the exact details until the official release, but point to the context: the launch fits into a broader trend of making flagship models more affordable. SiliconFlow and MiniMax are simultaneously cutting prices — competition among Chinese AI providers has become noticeably sharper than a year ago.

One of the most revealing moments of the trip happened even at the airport: one of the three people who met the authors turned out to be the owner of a company that sells data to OpenAI, Meta, and Anthropic. This illustrates the real state of the market — Chinese and Western AI ecosystems remain deeply intertwined, despite the official technological confrontation between the two countries.

• Companies on the trip program: ByteDance, Z.ai (GLM model), SiliconFlow, MiniMax
• ByteDance announced the release of a new product on July 10, 2026
• Z.ai is switching from NVIDIA to Huawei Ascend chips
• Server GPUs are sold in open Shanghai markets alongside consumer electronics
• Humanoid robots — available for retail sale with passport presentation

How Z.ai Is Leaving NVIDIA for Huawei

Z.ai — the team developing the GLM language model, one of the most powerful Chinese LLMs — is switching from NVIDIA chips to Huawei Ascend accelerators. The reason is straightforward: American export restrictions have blocked supplies of flagship H100 and H200 GPU series to China, and major laboratories have to rebuild their training infrastructure on alternative hardware. The transition is painful: the Huawei development ecosystem is significantly inferior to the mature NVIDIA CUDA environment in terms of tools, documentation, and support — but there are almost no alternatives.

At the same time, server graphics cards continue to appear on open markets. The authors describe a Shanghai market where data-center-level cards lie on shelves next to iPhones and ordinary consumer electronics — a vivid illustration of how sanctions have restructured the gray logistics of AI components.

"One of those who met us at the airport turned out to be the owner of a company from which

OpenAI, Meta, and Anthropic buy data" — from the GPTunneL report.

Robots by Passport — and Beijing Cleaners Are Still Ahead

One of the striking episodes of the trip was a visit to a shop where humanoid robots are sold at retail, but with a mandatory passport presentation. For China, this is already a reality: you can come in, choose, and buy a humanoid — like a large household appliance. Nevertheless, the authors honestly record the gap between storefront appeal and operational maturity: in real tasks — cleaning offices and city streets — Beijing human cleaners still consistently outperform their robotic counterparts. Industrial maturity of humanoids is still ahead.

What This Means

Chinese AI in July 2026 is simultaneously competitive models, an emerging chip base, and parallel supply channels. ByteDance, Z.ai, and SiliconFlow are moving at the same pace as American laboratories, but by different rules — their own infrastructure, their own data, their own market. Observing this from inside is a fundamentally different experience than reading analytics from the outside.

*Meta is recognized as an extremist organization and is banned in Russia.