Zhipu and DeepSeek are putting Chinese chips through the hardest tests
Two leading Chinese model labs are using home-grown chips for demanding work, which a commentary argues is what those chips have always lacked.
The short version
- The commentary argues Chinese chips lacked demanding customers, not government support or enthusiasm.
- Zhipu runs most of its newest model compute on domestic cards and bought a compiler software company.
- DeepSeek rewrote its inference code from Nvidia's CUDA to Huawei-style CANN tooling and open-sourced it.
- Rumors say both labs are training multi-trillion-parameter models on domestic cards. This is unconfirmed.
- The piece is opinion, and its efficiency claims rest on unnamed estimates.
Why it matters
A chip improves when someone demanding runs real work on it and reports what breaks. If this account is right, Nvidia's grip on China weakens through usage, not just export rules. Treat it as an argument, since the piece gives no independent benchmarks we can check.
The argument: chips need tough customers
As Cnyes (Anue) carried it, a commentary originally from Gongfu Caijing starts from a simple idea. A chip is judged less by its launch event than by how demanding its toughest customer is.
The author says Chinese chipmakers never lacked policy support or goodwill. What they lacked was a customer willing to hand them the heaviest workloads. While Nvidia's China-specific cards stayed easy to buy, the hardest jobs went to Nvidia, and domestic cards got lighter tasks. Feedback therefore stalled at "it works."
What Zhipu and DeepSeek are doing
The piece says Zhipu treats domestic cards as its main hardware, not a side option. Most compute for its latest model family reportedly runs on them, and one lighter model runs entirely on clusters built with chips from Hygon, a Chinese chipmaker. Zhipu also bought a company that makes a compiler, the software layer that translates a model's math into instructions a particular chip can run efficiently.
DeepSeek, the commentary says, rewrote its inference code, the code that runs a trained model, away from Nvidia's CUDA toolkit to a Chinese alternative called CANN, then released it openly. It also reportedly finished post-training of a recent model on a domestic chip. The author cites the analysis firm SemiAnalysis as saying part of DeepSeek's newest design was shaped with that hardware in mind.
The rumored biggest test
The most dramatic claim is a rumor: both labs are training models with several trillion parameters on domestic cards. Parameters are the adjustable numbers inside a model, and more of them means a bigger, costlier job.
Training at that scale means thousands of chips running for months without a failure, so any weakness in networking, memory or software gets exposed. If the rumor is true, it would be a harsher test than any benchmark. It is still a rumor, and the piece offers no proof.
Reading it with care
The author also cites a Bloomberg survey saying Chinese companies plan to put a large share of their AI chip budgets into domestic products over the next year. A big domestic maker has reportedly raised its production target sharply as well.
Note the limits. This is commentary written with a clear point of view. The claim that Zhipu's cost per unit of output is close to Nvidia's rests on unnamed institutions, and no numbers are given. Independent tests would settle it.
What to watch
- Will independent benchmarks confirm that domestic clusters match Nvidia on cost and efficiency?
- Will either lab confirm the reported very large training runs on domestic cards?
Try it
Before planning around any new model release, check which hardware and software stacks it officially supports, and whether the maker published results for your setup.
Sources and links
- ReportingCnyes (Anue): the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Cnyes (Anue) · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
