Shenzhen just lit up China’s most powerful AI data center to date. The new cluster runs at 11,000 petaflops and is built entirely on Huawei’s Ascend 910C chips, no NVIDIA, no AMD, nothing imported. It’s the country’s first 10,000-card level full-stack domestically controlled intelligent computing cluster built entirely with advanced Chinese chips.
Add that to an existing 3,000-petaflop cluster already humming at the same facility, and the site now totals 14,000 petaflops of computing capacity, with nearly 50 organizations already signed up and a combined booking rate of 92 percent across both phases.
That’s a big number. But the real story isn’t the number, it’s what the number reveals.
Impressive achievement, uncomfortable context
Credit where it’s due: pulling off a 10,000-card AI cluster with homegrown chips, under heavy U.S. export restrictions, is no small thing. China has been cut off from NVIDIA’s most advanced silicon for years. The fact that Huawei and Chinese engineers built something this large, this fast, entirely on domestic technology is a genuine industrial achievement.
According to Xinhua News Agency, 100 petaflops equals the computing power of 50,000 personal computers, making Shenzhen’s new cluster the equivalent of 5.5 million PCs working in unison. That framing sounds massive. And it is, until you zoom out.
According to analysis by Xataka, the Shenzhen cluster represents a tiny fraction of the capacity of the largest American data centers currently in operation. What China is celebrating today as its biggest AI infrastructure achievement is roughly what OpenAI had on hand when it trained GPT-4 back in 2022. Four years ago.
The gap isn’t about ambition. It isn’t about money or engineers. It’s about chips, what China can build, and how many.
The Ascend 910C is good. just not good enough yet.
The chip behind all 10,000 cards is Huawei’s Ascend 910C, and Huawei has worked hard to frame it as a serious NVIDIA alternative. In some ways, it earns that framing. Huawei’s CloudMatrix 384, a system linking 384 Ascend 910C chips together, has been noted by analysts as capable of outperforming NVIDIA’s GB200 NVL72 on some specific metrics.
But there are trade-offs that matter. The 910C’s combined silicon footprint is around 60% larger than NVIDIA’s H100, with lower performance per square millimeter and per watt. While the CloudMatrix 384 offers 300 petaflops of BF16 compute versus 180 from NVIDIA’s GB200 NVL72, it burns through almost four times the power to get there.
On inference tasks, the 910C delivers roughly 60% of H100-class performance, and training remains an even bigger challenge. In a world where training frontier AI models is the whole ballgame, that gap is significant.
Then there’s the manufacturing side. TSMC processed orders from Sophgo, a design partner of crypto-mining firm Bitmain, and those chiplets later turned up inside Huawei’s Ascend 910 AI accelerator, a supply chain breach that has since put TSMC under scrutiny for a potential fine exceeding $1 billion from the U.S. Commerce Department. That backdoor is closing, and fast.
Can China catch up before the gap gets worse?
That’s the question everything else hangs on.
SMIC, China’s largest chip foundry, is pushing 7nm-class production using DUV techniques, but yields still lag behind TSMC and Samsung. Domestic production is growing, but the quality and volume gap is real and not disappearing overnight.
Huawei isn’t standing still either. The company has a multi-year chip roadmap lined up: the Ascend 950PR for Q1 2026, the 950DT for Q4 2026, the Ascend 960 in 2027, and the Ascend 970 in 2028. If they execute on that plan, the picture changes. But execution means manufacturing at scale, with the right yields, using equipment that Western sanctions are increasingly cutting off.
Meanwhile, the U.S. side isn’t waiting around. Oracle, Microsoft, Google, and Amazon are deploying clusters with hundreds of thousands of NVIDIA Blackwell GPUs. The ceiling keeps rising, and every month China spends catching up, the target moves further away.
The Shenzhen cluster is real, operational, and reflects a genuine appetite for computing power from AI startups, robotics firms, and research universities across the region. The 92% booking rate proves the demand is there. The engineers are there. The money is there.
What isn’t there yet, at least not at the scale needed, are the chips.
Shenzhen is celebrating its biggest AI milestone ever. Silicon Valley is looking at 2022 and calling it Tuesday.
What do you think, can China actually close this chip gap in the next few years, or is the U.S. lead just too far ahead at this point? Tell us in the comments, we want to hear your take!