http://x.com/i/article/2086758982493085696

Weekly Dose of China Tech [08.10.2026]

Optical Transceiver Bans, Unitree's IPO, DeepSeek's API Price Hike - and CXMT's Love Affair With Apple

Hi friends,

CT here. Han is tied up with another project, so I'm taking over the weekly tech roundup for now.

Hangzhou is being hit by Typhoon "White Dolphin," and I've spent the evening listening to the wind hammer against the windows. Chinese tech news felt a bit like that.

The biggest gust came from Washington, where the FCC is preparing restrictions on Chinese optical transceivers. Shares of Zhongji Innolight and Eoptolink briefly fell nearly 10% in a single day. But one detail got lost in the panic: Coherent, an American optical-communications company, has a factory in Wuxi. The supply chain is messier than the headline suggests.

Then there is CXMT and Apple. Apple is reportedly helping the Chinese memory maker seek approval to enter its supply chain, even though CXMT's orders are already booked well into 2027 and customers like Tencent and ByteDance could absorb much of its available capacity. Apple and CXMT increasingly look like a celebrity couple caught in a dating rumor: everyone has seen the photos, but they have not even had dinner together yet.

That is what storms do. They create enough noise that people react before they think. Breaking news matters, but the most useful information is often buried in the details: who benefits, who gets hurt, and what each side actually wants. So let's look past the headlines and see what really happened.

This Week Features...

The Apple Supplier Lens Technology Outgrows Its Patron

On May 14, during Donald Trump's visit to China, a woman sitting next to then-Apple CEO Tim Cook caught some attention. She was Zhou Qunfei, founder of Lens Technology, one of Apple's key component suppliers and a longtime maker of iPhone cover glass. At first glance, Zhou's story looks like the classic rise of a Chinese contract manufacturer: make something for someone else, master the process, scale it up, and get paid.

But Zhou does not seem interested in remaining someone else's craftsman-or servant. Lens has invested in Chinese smart-glasses maker Rokid and helped solve the mass-production problem for optical waveguide lenses. It also backed Qingtian Rent, a robot-leasing platform linked to AgiBot, while taking on component manufacturing for several Chinese robotics companies.

For years, Lens was criticized for depending too heavily on Apple. Zhou's answer now is not simply to find more customers, but to invest in companies that could become its future customers. That is also the focus of this week's story: Chinese manufacturers made fortunes from "Made in China," but the direction of their industries was still largely set by foreign customers like Apple and Tesla.

Now some of those suppliers are taking a riskier path. They are putting their own money into startups and emerging technologies, hoping to secure a place in the industries that may define the next decade. They are no longer waiting for the future to send them a purchase order-they are investing in the future themselves. And the stock market is starting to price them accordingly.

[Read the full piece →]

The Missing Middle Layer of AI, and Why China Built It First

In Wuxi, China, a new kind of AI infrastructure is taking shape. It does not train frontier models or manufacture chips. It produces something else: tokens.

China's enterprise AI market is no longer just a race to build the smartest model. According to industry tracking from the 2025 China Large Model Procurement Monitoring Report, Chinese organizations signed more than 7,500 contracts related to large-model projects in 2025, worth nearly RMB30 billion. What is striking is that many of these buyers were not primarily asking which model scored highest on a benchmark. They cared more about deployment, security, reliability and control.

Nvidia CEO Jensen Huang has described the next generation of AI infrastructure as an "AI factory"-a system that turns electricity and computing power into tokens. In the U.S., much of that capability is absorbed inside vertically integrated platforms such as AWS, Microsoft Azure and Google Cloud, or within the major AI labs themselves.

China did not invent this idea. What makes the Chinese market interesting is that it is becoming one of the first places where the concept is emerging as a standalone industry. Local governments, infrastructure providers and startups are building what they increasingly call "token factories": dedicated systems whose business is not to create the next foundation model, but to produce AI inference at scale.

[Read the full piece →]

It's Nearly Impossible to Remove China From Tesla

Any American company that builds a huge business in China eventually has to confront an uncomfortable question: What happens if that business has to be separated from the rest of the company?

The Wall Street Journal put that question on the table and picked perhaps the most obvious target: Tesla. According to its July 30 report, Tesla had explored possible ways to separate its China operations as part of preparations for a potential merger with SpaceX. Musk quickly denied it.

But the report creates a much more interesting thought experiment: What would actually happen if Tesla tried to give up Gigafactory Shanghai? We went through the factory's production, financing, suppliers, data and employees, and the conclusion was fairly simple-leaving China would create one problem after another for Tesla. And those might not even be the biggest problems.

So this week, we're running the simulation ourselves. You can play along.

[Read the full piece →]

The News...

(i) ByteDance's 5 Trillion Parameters

ByteDance is discussing training a model with more than 5 trillion parameters, according to LatePost, potentially making it the largest known AI model under development in China. The project would surpass Alibaba's Qwen 3.8-Max (2.4T parameters) and Moonshot AI's Kimi K3 (2.8T parameters), though it remains in an early stage.

Founder Zhang Yiming reportedly encouraged the AI team to pursue original architectures and maintain a "no distillation" approach. He argued that distilling models can quickly improve performance but mainly replicates existing capabilities such as Anthropic's Claude, making it difficult to achieve true breakthroughs.

ByteDance had considered distillation amid pressure from advanced overseas models and limited AI compute resources, but Zhang ultimately pushed the company toward long-term AGI development.

Read Full Article