Huawei and DeepSeek Close the Gap With Nvidia, Report Says
DeepSeek's deepening partnership with Huawei is helping the Chinese hardware champion close the distance to Nvidia's chips and the software that runs on them, according to a report from SemiAnalysis which concludes that Jensen Huang's nightmare is starting to come true. The South China Morning Post reports that the research firm points to DeepSeek's recent software releases and its increased testing on Huawei's Ascend 950PR processor.
The arrangement suits both sides. DeepSeek gets silicon tuned for its models, and Huawei gets a demanding customer that pushes its toolchain forward, which is why the report describes the relationship as a symbiosis rather than a supply deal.
And the software is where the shift shows. DeepSeek released six modules built for Ascend that mirror tools it had previously written for Nvidia chips, among them an Ascend version of TileLang, a language for producing the high performance kernels that decide how fast a chip actually runs.
DeepSeek said the Ascend version of TileLang can help build a domestic software ecosystem for graphics processors that is "independent and controllable". That framing matters because the code layer, not the chip alone, is what keeps developers tied to a vendor's toolchain, and moving that layer onto Ascend gives Chinese teams a path that does not depend on the American one.
SemiAnalysis expects DeepSeek to rely "primarily" on Huawei's "already mature and well developed" stack by 2027 or 2028, "especially with the external and internal pressure to build a fully domestic Chinese ecosystem". Beijing has made technology independence a national priority while US export controls limit access to advanced chips and the equipment that makes them, which is the pressure the report describes.
But the hardware still trails. The Ascend 950PR lags Nvidia's H100 in absolute performance, according to the report. Nvidia has been shipping the H100 since 2022, which is a long lead in a market that moves this quickly.
Huawei's engineers are nonetheless earning praise. Their best Ascend kernels reach 99.7% of their theoretical ceiling, a measure of how close the code comes to what the silicon can deliver, and the report calls that good work.
The analysis also sets out what the collaboration has not fixed. Optimisation has concentrated on large scale supernode deployments, which leaves unresolved problems for developers who want to run Ascend cards on standard x86 host processors rather than Huawei's own designs, and configurations outside DeepSeek's specific setups may run into compatibility trouble.
So the benefits could reach past the two companies. SemiAnalysis argues that other operators serving DeepSeek can reuse the kernels and the deployment practices, and that other laboratories can build new models on the shared compiler work. "If these users contribute fixes and optimisations back, the gains can compound," its analysts wrote.
Huang has warned about this direction in public before. Ahead of DeepSeek's V4 model he said on a podcast that it would be "a horrible outcome" for the United States if DeepSeek tuned its models to run on Huawei hardware, and that China "will become superior to" the US as a result. Huawei has announced "Day-0" inference support for the V4 and V4.1 releases, meaning its chips could serve the models as they arrived.
SemiAnalysis says Huawei is closing the chip and software gap with Nvidia, with DeepSeek's help.
DeepSeek shipped six Ascend modules mirroring its Nvidia tools, including an Ascend version of TileLang.
The firm expects DeepSeek to rely primarily on Huawei's stack by 2027 or 2028.
The Ascend 950PR still trails Nvidia's H100, though Huawei kernels reach 99.7% of their theoretical ceiling.
Optimisation has focused on large supernodes, leaving gaps for standard x86 hosts.
Source: SCMP
