Huawei says its Ascend supernode based on the Ascend 950 AI chips will fully support DeepSeek V4, as DeepSeek launches a preview of its V4 model
Reuters
Context & Ripple Effects
DeepSeek and Huawei had already been reported to be optimizing V4 for Huawei hardware, while bulk orders tied to V4 reportedly lifted expected Ascend 950PR pricing. The announced support turns that preparation into a deployable hardware-model pairing as V4 enters preview.
The coverage also shows a mixed-compute transition rather than a clean break: DeepSeek previously planned to use Ascend chips for smaller R2 training workloads while retaining Nvidia for its largest models. Later work on Ascend 910C for V4 Pro post-training extends Huawei’s role beyond inference compatibility.
First-order effects
Huawei can position Ascend 950-based supernodes as ready infrastructure for organizations deploying DeepSeek V4, reducing integration uncertainty for the two named platforms.
DeepSeek gains a validated domestic hardware route for V4 deployments, while Huawei gains a prominent workload to support demand for its Ascend systems.
Second-order effects
Chinese cloud and enterprise buyers evaluating V4 can compare an integrated DeepSeek–Ascend stack with alternatives rather than treating model availability and accelerator procurement as separate decisions.
Demand linked to V4 could further tighten the commercial connection between model releases and Ascend capacity or pricing, following reports of bulk orders for the 950PR.
Third-order effects
If compatibility work expands from deployment to more training and post-training stages, Chinese AI stacks could become less dependent on a single external accelerator ecosystem, though the prior reliance on Nvidia for DeepSeek’s largest models shows that transition remains incomplete.
The strategic unit of competition shifts toward co-optimized model, chip and supernode packages—not standalone models or accelerators—making software tooling and workload support as important as chip availability.
The trend: This is one data point in the build-out of vertically coordinated, domestically sourced AI compute stacks around model-specific hardware optimization.
DeepSeek is trolling They don't say *what* they trained on, only “eh MXFP4” Hawks will now go on a hunt for Blackwells in Inner Mongolia In a year they'll publish “insights into V4” with details of an experimental Huawei cluster man I love this lab so much [image]
> The library supports SM90 (Hopper) and SM100 (Blackwell), with no Huawei Ascend support I wanted to say “TileLang doesn't even support Ascend” but apparently there is a variant still, I repeat that all this “DeepSeek spent 9000 years adapting for Huawei” is most likely bullshit…
Buried in the fine print: DeepSeek says V4-Pro throughput is currently limited by high-end compute supply. Prices will drop significantly once Huawei Ascend 950 super nodes ship at scale in H2. DeepSeek is publicly tying its API economics to domestic chip infrastructure. That's […
DeepSeek 4 upon Huawei's Ascend chipsis apparently being released in a live event via Bilibili, China's equivalent to Youtube, today at 7pm Beijing Time on April 24. [image]