https://x.com/i/article/2104737816509100032

【X 长文】Weekly Dose of China Tech [9.28.2026]

Hi friends,

One of the most interesting things happening in China tech right now is what happens after a company hits a wall.

Huawei hit a very big one. US restrictions cut off key smartphone and chip supply, and its overseas business fell sharply. But five years later, the company is selling watches and earbuds, supplying solar projects, winning telecom contracts, and finding new opportunities in AI infrastructure abroad.

The same pattern is appearing elsewhere.

DeepSeek is running into compute shortages and looking at Huawei chips. Alibaba is building its own cloud and accelerator stack. CXMT is pushing domestic memory into flagship phones. CATL is moving production into Europe. Even in robotics, companies are building the infrastructure around the machines before the technology is fully mature.

China’s tech industry is not simply trying to build substitutes for everything it has lost access to.

It is increasingly trying to redesign the system around the constraint.

Anyway, let’s take a look.

Weekly News...

(I) DeepSeek May Need Huawei Chips for Its Next Giant Model

DeepSeek appears to be preparing for a much larger compute bill. According to The Information, the company is considering using up to 950 Huawei 950DT chips for training, while reportedly working on a 2-trillion-parameter model and planning an even larger 8-trillion-parameter system.

But Huawei is unlikely to replace Nvidia overnight. Limited domestic chip and HBM supply means DeepSeek may still need Nvidia capacity alongside Huawei hardware.

The bigger story is that China’s AI race is moving toward a multi-chip strategy. The question is no longer just whether Huawei can replace Nvidia, but whether Chinese AI companies can assemble enough compute from several sources to keep scaling.

(II) Kimi K3 Just Got a New Route to Enterprise Customers

Moonshot AI has brought Kimi K3 onto Amazon Bedrock, putting a Chinese open-weight model directly inside AWS’s enterprise cloud ecosystem.

The arrangement is a useful test of whether open models can generate revenue through distribution rather than direct model hosting. AWS says customer data stays within its cloud and is not shared with Moonshot or used for training.

For Moonshot, the appeal is simple: enterprises may want the flexibility of an open model without having to operate one themselves. Open source lowers the barrier to access; cloud platforms may provide the monetization layer.

(III) Alibaba Wants 20GW of Cloud Capacity

Alibaba says it wants more than 20GW of global cloud data-center capacity by 2032, as strong AI demand drives another wave of infrastructure spending.

At the same time, T-Head has unveiled the Zhenwu V900, which Alibaba says delivers three times the compute performance of the M890 and can support clusters of up to 500,000 accelerators.

Alibaba is increasingly building the whole stack around AI: chips, cloud, and models. The 20GW target shows the company is betting that the infrastructure underneath AI may be a larger and more durable business than the models themselves.

(IV) Unitree Just Put a Price on the Robot Hand

Unitree’s new Dex5-S starts at RMB 39,900 and packs 22 motors and 22 degrees of freedom into a 620-gram robotic hand.

The product addresses one of humanoid robotics’ least...

Read Full Article