Published on Sep. 28, 2026

Alibaba used its annual Apsara Conference in Hangzhou to lay out the most ambitious AI infrastructure plan any Chinese company has announced so far. On September 22, 2026, its chip subsidiary T-Head unveiled the Zhenwu V900, an AI accelerator with three times the computing performance of its predecessor, the Zhenwu M890, reported the tech portal Xin Dongxi (in Chinese).

The chip has 216 gigabytes of memory and a chip-to-chip bandwidth of 1.2 terabytes per second. A single cluster can scale to as many as 500,000 chips. Mass production is scheduled for the first quarter of 2027. A successor named Zhenwu J900 is to follow in the third quarter of 2028. The new V900 is “the most powerful AI chip in China today in terms of computing performance,” CEO Eddie Wu said.

The chip is one of three pillars. Alibaba's ambitions extend beyond AI chips to AI models and the AI clould as the cornerstones of what it calls the era of machine intelligence. “If tokens are the electricity of the AI era, chips are the generators,” Wu was quoted as saying in the business newspaper Huaxia Shibao on September 23, 2026 (in Chinese).

Machines will take over 99.9 percent of thinking tasks in the future, Wu believes. “The total amount of thinking supplied by machines in the future will be more than 1,000 times the total amount of human thinking.”

On models, the Qwen team plans to train Qwen 5 with 5 to 10 trillion parameters, designed for complex tasks over extended periods. Alibaba’s current flagship, Qwen 3.8-Max, has 2.4 trillion.

With regard to the cloud, Alibaba aims for more than 20 gigawatts of global data center capacity by 2032, starting with a new site in the Netherlands in October, reported the South China Morning Post.

Alibaba has committed more than US$53 billion to AI over three years and raised about US$10.2 billion in a share offering in August, news reports say. Alibaba’s profit fell by more than 75 percent in the last quarter after a spending spree of nearly US$10 billion.

Wu did not say whether Alibaba can produce enough of the new chips in the short term. This is an open question, since U.S. export controls block Chinese firms from using TSMC.

Alibaba is assembling a domestic full stack of models, chips and cloud at the moment when Washington is trying to keep advanced silicon out of China.

State-owned enterprises (SOEs) in China are getting serious about robots. They commission them from private robot makers and also invest real money in their own robot manufacturing.

Most of the self-developed machines are specialist robots for dangerous jobs on power lines, in tunnels and in mines, but the procurement lists of the state enterprises increasingly include humanoids as well.

Tens of meters below ground, a tunnel segment assembly robot developed by the state-run China Railway Construction now grips, positions and fits the concrete segments behind a shield tunneling machine. The misalignment is no more than 5 millimeters, reported the industry newspaper Zhongguo Gongye Bao on September 24, 2026 (in Chinese).

State Grid has built an ice-removal robot that a drone lifts onto a high-voltage line. It moves at one meter per second and clears the roughly 500 meters between two towers in ten minutes, said Yuan Hong of State Grid Intelligent Technology.

Another SOE, CSSC Chongqing Equipment, built a humanoid that sprays insulation inside ship compartments, a job for which workers are more and more hard to find. It runs on a large AI model and currently still trails human workers in accuracy by around 10 percent, but is learning fast. The humanoid form is best for work in confined spaces, its maker says.

In August, Norinco Group formally established a robot innovation consortium of central state-owned enterprises with more than 100 organizations. Private robot makers have technology but lack application scenarios, the newspaper argues. The entry of the state sector is changing that.

“The core uniqueness of central SOEs taking the lead is not that they have the strongest individual technologies, but that they possess China’s largest pool of real-world application scenarios for robots,” said Bai Wenxi, deputy chairman of the China Enterprise Capital Alliance.

The private sector in China, deemed important for competition and innovation by central government planners, is integrated into the roadmaps. Deep Robotics for example, a private company in Hangzhou, has a five-year framework agreement worth 1 billion yuan (about US$148 million) with State Grid. The company currently holds 85 percent of the market for power inspection robots.

The entry of the state giants answers a question that hung like a shadow over China’s robot boom, namely who will buy the machines. Annual procurements of state-owned enterprises are likely to amount to hundreds of billions of yuan (tens of billions of US dollars). This will also be a boost for the private robotics sector in China, as the procurement lists of State Grid and PetroChina already include Unitree, AgiBot, UBTech and Fourier.

Envision, a Chinese green energy company and wind turbine maker, is connecting AI data centers directly with green wind power. The biggest project of this kind so far has now fired up its servers and promises to deliver exceptionally cheap tokens.

Envision’s Galaxy Campus in Ulanqab, Inner Mongolia, began operations in August, reported the business portal Guandian on September 19, 2026 (in Chinese).

The facility covers 120,000 square meters, about 20 football fields.
The volume of AI inference or training facilitated by this facility with cheap, green energy is noteworthy. Galaxy Campus can run one million chips in parallel and has computing power on the scale of one million petaflops, which makes it the single AI data center with the highest token output capacity in the world, according to Envision.

More than 80 percent of its electricity is directly connected green power. It is the first flagship project of Envision’s “Mission Gobi” initiative, which aims to develop 5 gigawatts of green AI data center capacity in desert and arid regions worldwide by 2030.

Ulanqab is known as the “capital of wind power.” Electricity for data centers there costs about 50 percent less than in the Beijing-Tianjin-Hebei region, and computing power more than 60 percent less than in first-tier cities, according to official statistics.

Envision’s founder Zhang Lei wants to make tokens “as cheap as potatoes,” Guandian reported, adding that whether this goal can be reached will depend on actual demand and long-term costs.

Cheap tokens are the basis of China’s strategy to make capable AI ubiquitous. In Beijing’s view, this is at least as important as winning the global race for the single best AI model.

This week’s Deep Dive looks at the European suppliers, Siemens and ABB among them, that sell the power technology such campuses need.

Alibaba’s DAMO Academy has released DAMO EAGLE, an AI model that screens for esophageal cancer using nothing more than a routine chest CT scan, reported the financial news portal Cailianshe on September 24, 2026 (in Chinese).

The model was developed with Sichuan Cancer Hospital and Sun Yat-sen University Cancer Center and needs neither intubation nor contrast agents.

The esophagus is a hollow tube prone to collapse and motion artifacts on CT scans, which made detecting small early lesions “a task historically considered impossible,” the researchers write in Nature Medicine.

EAGLE was trained on 6,813 patients and validated across 12 centers in three countries with 80,612 patients. In external test cohorts it reached 98.5 percent specificity, with 90 percent sensitivity for cancer and 52.5 percent for precancerous lesions. It is Alibaba’s fourth cancer-screening model, after pancreas, stomach and colorectal cancer.

Publication is not deployment, however. False-positive rates in real screening populations and performance across CT vendors still have to be proven.

China accounts for roughly half of the world’s esophageal cancer cases and deaths, and most patients are diagnosed late. A model that turns screening into a byproduct of scans hospitals already perform pushes the marginal cost of screening toward zero.