Sources: Chinese internet companies have been testing Huawei's new Ascend 910C chip in recent weeks; Huawei told clients the chip is comparable to Nvidia's H100
Chinese tech company looks for AI business with Ascend series but still faces production issues
Context & Ripple Effects
Huawei’s Ascend effort began with its 2018 data-center AI-chip launch, but its push to become a domestic alternative has been constrained by reported difficulty ramping Ascend AI-server production. The new customer testing is therefore an important transition from product positioning to real-world evaluation.
The central question is not only claimed H100 comparability, but whether performance can be delivered reliably at volume. Later coverage of software, stability and interconnect shortcomings in Ascend training deployments underscores that distinction.
First-order effects
- Chinese internet companies can benchmark the Ascend 910C against their own AI workloads, giving Huawei direct feedback and a potential route to initial deployments.
- Huawei’s H100-comparability claim is put to customer validation while its production constraints remain a near-term limit on how much testing can become procurement.
Second-order effects
- A viable 910C option would give Chinese AI buyers more leverage in hardware sourcing, while Nvidia faces a more credible local alternative in the market Huawei is targeting.
- Testing shifts attention from chip-level specifications to the surrounding software, networking and server integration needed for customers to operate Ascend systems effectively.
Third-order effects
- If Huawei can pair competitive hardware with dependable supply and a usable software stack, Chinese AI infrastructure could become more heterogeneous rather than centered on a single accelerator platform.
- The next product cycle—Huawei later approached clients about evaluating the Ascend 910D—suggests competition will increasingly turn on iteration speed and deployability, not benchmark claims alone.
The trend: This is one data point in the emergence of regionally differentiated AI-compute stacks, where supply access and software integration shape accelerator choice alongside raw performance.