DeepSeek-Ascend Release Shifts China's AI Limit to Chip Supply

By
CTOL Staff Reporter
1 min read

DeepSeek has open-sourced a full Ascend port of its Nvidia software stack, and Huawei's Ascend 950 cluster cloud service went on commercial sale in China the same day. Each Ascend 950 chip runs at about 1 petaflop of FP8 compute, half an Nvidia H100's rating, so matching Nvidia from here depends on chip volume more than clever code.

Liang Wenfeng, DeepSeek's founder, did the sum himself in May: training a top model would take about 50,000 of Nvidia's GB300 chips, or about 200,000 of Huawei's Ascend 950s. Four Huawei chips for every Nvidia one: that ratio is now China's AI problem, because on Wednesday DeepSeek removed the other one. It published a complete port of its own training software, the compilers and communication libraries that make a chip usable at scale, onto Huawei's Ascend hardware, and said every operator it uses to train its models now has a matching Ascend version. DeepSeek credited Huawei with unreserved support. The release landed on the day Huawei had set for its Ascend 950 cluster cloud service to open to paying customers in China, and a day after China Mobile announced a Beijing supernode built from 1,024 of the chips. None of this needed coordinating in public. DeepSeek removes the reason to stay tied to Nvidia's software, and Huawei supplies the metal underneath it.

Nvidia's real moat was never the chip. It was CUDA, which Nvidia opened to general-purpose programming back in 2006, two decades of kernels, libraries and institutional habit built on top of it long before anyone imagined a training market worth hundreds of billions of dollars. Moving a production training pipeline to a new chip family has usually meant rewriting that whole layer by hand, which is why software, more than any export licence, kept Chinese labs loyal to Nvidia. TileLang, the compiler DeepSeek uses, changes that arithmetic. An engineer writes an operator once and compiles it for either an Nvidia GPU or an Ascend chip, with the Ascend version wrapping Huawei's own low-level instruction set underneath. What used to require a rewrite is now closer to a recompile. Huawei's own figures point the same direction: outside contributors to its CANN developer community now outnumber Huawei's own employees, and more than 40 large models have already been pretrained on Ascend chips.

The hardware still lags. A single Ascend 950 chip runs at roughly 1 petaflop in the FP8 format used for large-scale training, about half of what an Nvidia H100 already delivered years earlier, and a small fraction of what Nvidia's current Blackwell chips do. Huawei's answer is brute scale: more chips, more power, more floor space. Wire 1,024 of those chips together, as China Mobile did this week in a Beijing data center, and the resulting machine is rated at 1 exaflop of FP8 compute with 256 terabytes of shared memory, simply the single-chip number multiplied by the chip count. DeepSeek's own new supernode, co-built with Huawei, is a separate and much smaller 128-chip design. It proves the software runs; the 1,024-chip machine is where scale gets tested.

Beijing is running Nvidia access through a filter, approving some chips for import while keeping others out by category. China's industry ministry asked ByteDance and Alibaba to submit purchase plans for Nvidia's RTX PRO 5500, an inference-class chip outside the current US export limits, and signaled it would approve them; ByteDance alone could seek around a million units, with shipments starting in late December. Nvidia's H200, cleared for export by Washington months ago, has barely shipped to China because Beijing restricted it on its own initiative. Put the two policies together and the shape is a valve that opens for chips that run other people's already-trained models and stays shut for chips built to train new ones, a policy DeepSeek's release just made cheaper to enforce.

Skeptics will say open-sourcing a kernel library is a long way from training a frontier model on Ascend, and they are right that DeepSeek still trains on Nvidia today. But Liang told investors this month, according to The Information as cited by Reuters, that moving training to Huawei chips is a priority that must succeed, and DeepSeek expects a new batch of Huawei training chips in the fourth quarter or early next year. The excuse that Ascend's software was too immature to trust with a full training run is gone, because the lab best placed to judge it, DeepSeek, has just put its own training code on the line. What decides the pace of migration from here is production: how many Ascend 950 chips Huawei's fabs can turn out, and how fast the 960 generation, pulled forward to the first quarter of 2027, actually ships in volume.

For Nvidia shareholders the read is unsentimental: China training revenue is gone for good, whatever comes out of the trade talks that followed last week's Trump-Xi meeting in Washington. What remains addressable is inference volume of the RTX PRO 5500 kind, a real business, but one priced and margined nothing like selling into frontier training runs. For Cambricon and Hygon, the best-known domestic rivals, DeepSeek's decision to build its software bridge specifically to Ascend, leaving them outside the handoff, is close to a verdict. The Chinese AI stack is consolidating around Huawei's silicon regardless of whether that silicon is the best available on a spec sheet next to Nvidia's.

Huawei has already dated the next milestone. Its Lingqu Ascend 950 cloud service goes on sale outside China on November 30, and the 960DT chip, originally due in the fourth quarter of 2027 and rated at twice the 950's FP8 throughput, now arrives in the first quarter. We expect DeepSeek's next major model to disclose that at least part of its training ran on Ascend. If it does, the migration Beijing has been engineering since the earliest US export curbs will have moved from a software demonstration to a production fact, and the chips doing the training will have come from a Chinese fab.

You May Also Like

This article is submitted by our user under the News Submission Rules and Guidelines. The cover photo is computer generated art for illustrative purposes only; not indicative of factual content. If you believe this article infringes upon copyright rights, please do not hesitate to report it by sending an email to us. Your vigilance and cooperation are invaluable in helping us maintain a respectful and legally compliant community.

Subscribe to our Newsletter

Get the latest in enterprise business and tech with exclusive peeks at our new offerings

We use cookies on our website to enable certain functions, to provide more relevant information to you and to optimize your experience on our website. Further information can be found in our Privacy Policy and our Terms of Service . Mandatory information can be found in the legal notice