‘Jensen’s nightmare’: Huawei, DeepSeek narrow AI chip ecosystem gap with Nvidia: report

South China Morning Post | Ann Cao

Huawei Technologies is narrowing the gap with Nvidia’s chip hardware and software ecosystem through closer work with AI start-up DeepSeek, a SemiAnalysis report published on Thursday said. The firm wrote that software results and testing on the Ascend 950PR show the ecosystem gap is closer, and that Jensen Huang’s “nightmare is starting to come true”.

DeepSeek released six software modules for the Ascend ecosystem that mirror ones it built for Nvidia chips, including an Ascend version of TileLang, a programming language for kernels, the code that operates chip hardware. SemiAnalysis expects DeepSeek to rely “primarily” on Huawei’s stack by 2027 or 2028. Gaps remain. The Ascend 950PR, unveiled in March, lags Nvidia’s H100, which began shipping in 2022, though Huawei’s best kernels reach 99.7 per cent of their theoretical ceiling. Optimisations have centred on large supernodes, leaving problems for developers using x86 host processors. Beijing has made tech self-reliance a national priority. Shared fixes could compound gains.

Comment

The Ascend 950PR’s usefulness depends on its kernels as much as its silicon. Kernels are the low-level routines that turn a processor’s raw capacity into work an AI model can use. The chip still lags Nvidia’s H100 in absolute performance, yet kernels reaching 99.7 per cent of their theoretical ceiling suggest tuning could recover part of that deficit. DeepSeek supplies what a chip designer rarely has: one demanding workload to tune against.

Google’s Tensor Processing Units offer a documented precedent for this arrangement. Google built those accelerators alongside its own machine-learning software and used them in its own services before offering them to outside customers. An anchor customer gives a chip a guaranteed workload and a feedback loop that open-market sales rarely provide. Huawei’s Day-0 inference support for DeepSeek V4 and V4.1 resembles that loop, though DeepSeek is an independent start-up and not a Huawei division.

The precedent has limits in this case. Optimisations centre on DeepSeek’s large supernode set-ups, and developers running Ascend hardware with x86 host processors face unresolved issues. Wider use of the Ascend 950PR would depend on other operators reusing DeepSeek’s kernels and contributing fixes back.

Strategic Question for Discussion

What happens to the Ascend 950PR’s prospects beyond DeepSeek if the kernels tuned for its supernode set-ups do not carry over to x86 host processors?

My assessment is that the chip would stay a strong domestic option for one dominant customer while struggling to become a general platform. The pattern suggests ecosystem gaps close fastest where a single workload dominates and slowest where hardware configurations vary. Reliance on DeepSeek would then be both the Ascend 950PR’s main strength and its ceiling.

Share your assessment in the comments below.

To read in details click here