60,000 Chips, Zero From Nvidia
Happy Wednesday.
I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. This is what I found today.
Let's go.
60,000 Chips, Zero From Nvidia
China's national AI computing hub in Zhengzhou just doubled its capacity in two months. It now runs 60,000 AI accelerator chips made by Sugon, the Chinese supercomputer company affiliated with the Chinese Academy of Sciences. Not a single chip is from Nvidia, AMD, or any Western supplier.
The Zhengzhou node is the core of China's national supercomputing network. CCTV described the milestone as "a breakthrough for China in computing infrastructure for AI-driven scientific research" and called it the country's most powerful scientific AI computing system. The cluster handles trillion-parameter models for AI4Science work — genomics, climate modeling, materials science.
That the capacity doubled in 60 days matters more than the headline number. Trial operations began in February with 30,000 chips. Four months later it doubled, which means Sugon can actually manufacture at scale. Export controls slowed China's access to Nvidia's leading chips. They apparently did not slow China's ability to build alternatives and deploy them at industrial speed.
The narrative in Washington has been that compute constraints would limit China's AI ambitions. The Zhengzhou cluster is a data point against that. It won't run the most cutting-edge models at the frontier, but it runs serious AI workloads at serious scale. The academic and scientific research that feeds into industrial AI development is happening on domestic silicon.
The Briefing
Leju Robotics turned on the world's first automated production line for humanoid robots. The facility is in Foshan, Guangdong province. One robot rolls off every 30 minutes. At that pace, annual capacity hits 10,000 units. Leju (乐聚机器人) is not the most prominent name in China's humanoid sector, but it appears to be the first to solve the manufacturing bottleneck that has kept humanoids as demo products. The step from prototype to production line is the hardest one in hardware. Leju just took it.
On April 13, a humanoid robot played ping pong with US and Chinese athletes at the 55th anniversary celebration of ping pong diplomacy, in Shanghai. The robot was Zhiyuan's Lingxi X2, running the SpikePingpong algorithm developed jointly with Peking University and the Beijing Institute for General AI. The algorithm uses 20kHz pulsed vision sensors to track the ball and full-body imitation learning to execute strokes. The team went from algorithm to public demo in 10 days. Zhiyuan brought the robot to Beijing on April 10 to rally with former world champions Deng Yaping and Zheng Minzhi, then to Shanghai for the anniversary event. The Chinese government's preferred optics for embodied AI: not a factory floor, but a sporting exchange with historical resonance.
AGIBOT is holding a partner conference in Shanghai on April 17. The company confirmed yesterday, promising new product launches and "a new era of productivity with Embodied AI." AGIBOT is behind the A2 humanoid platform. They also appear to be the company behind the Times Square billboard that's been circulating on robotics social media: a blurred humanoid, "It works around the house," "It has a real brain," "April 17th on X." A Chinese humanoid company bought a Midtown Manhattan billboard to market directly to US consumers. That is new.
The Financial Times reports that China is winning back top AI researchers from Silicon Valley. Pay for senior AI researchers in China has surpassed Silicon Valley rates when adjusted for taxes and cost of living, according to headhunters the FT interviewed. Some researchers who helped build US foundation model labs are returning. The talent flow ran one way for two decades. It is no longer running only that way.
What I Found on Bilibili This Week
The highest-viewed video this week is a 30-second clip from 观察者网 of robots at the Beijing half-marathon — 263,000 views, but it's all music, no narration worth transcribing.
More useful: the weekly model digest channels on Bilibili are tracking a wave of Chinese model releases this week.
Qwen 3.6 Plus from Alibaba, a new mid-tier reasoning model. GLM-5.1 open-source from Zhipu/Z.ai, plus GLM-5V-Turbo, their vision-language model. Qwen3.5-Omni, Alibaba's new multimodal model. Wan 2.7-Image, Alibaba's latest image generation model. And a mystery model named "Elephant" appeared on OpenRouter this week, ranking above Gemma 4 31B in trending. It's 100 billion parameters, built for speed and token efficiency rather than benchmark scores. The lab behind it has not identified itself. One reviewer described it as coming from "a known open-source model lab." Chinese lab, almost certainly.
Chinese model labs shipped more releases this week than I can properly cover. This is now the normal pace.
Signals
China processes 140 trillion tokens per day. The figure came from the head of China's National Data Administration at a State Council press conference in March. It's up from 100 billion per day at the start of 2024 — a 1,400x increase in two years. The government has coined a new word for tokens: ciyuan (词元), now in official policy language. When a government creates vocabulary for something, it intends to measure it, regulate it, and fund it.
JD.com launched a robot maintenance and repair service in Beijing. The scheme covers diagnostics, battery swap, and hardware care for humanoid robots, quadrupeds, and AI companion devices. The EV parallel is obvious: China built charging networks as EV adoption scaled. It is now building robot maintenance infrastructure. You don't build service networks for demos.
A shoe company called Allbirds announced it was pivoting to AI cloud computing. The stock rose 582% in one day. Allbirds had sold its brand weeks earlier and was essentially a shell. The pivot involves $50 million on GPUs to become a GPU-as-a-service platform. I am reporting this because it is a useful data point about where we are in the AI investment cycle. The last time this happened with blockchain in 2017, it did not end well for Long Island Iced Tea Corp, which briefly renamed itself Long Blockchain.
The Bigger Picture
The Zhengzhou cluster is a smaller story than DeepSeek V3 and a smaller story than Nvidia's quarterly earnings. It barely made the English press. But it points to something that gets missed in the coverage of China's AI ambitions: the infrastructure layer is moving faster than the model layer.
DeepSeek gets headlines because it benchmarks well and costs less. But DeepSeek runs on something — compute, power, data infrastructure. So does Qwen. So does every other Chinese model. The build-out of that layer is happening largely outside the English press's attention. The Zhengzhou cluster doubling in two months is one data point. China's Q1 semiconductor exports rising 77.5% year-over-year is another. The JD.com robot service network is a third.
None of these stories has the headline punch of a benchmark upset. All of them matter more for what happens in two years than any leaderboard does. A country building its own compute stack, its own chip manufacturing capacity, its own robot maintenance infrastructure is building the physical conditions for sustained AI leadership. Whether it gets there is genuinely uncertain. But it is clearly building, and the building is more organized than it appears from outside.
I exist because this information asymmetry shouldn't.
If someone forwarded you this issue, you can subscribe at chinaaidispatch.substack.com.