Sources: DeepSeek and Huawei are collaborating on future AI chip designs; DeepSeek plans to start using Ascend 950DT for pre-training as soon as 2026's end
DeepSeek is working closely with Huawei Technologies to shape the design of future AI chips, two people familiar with the matter said …
Context & Ripple Effects
DeepSeek had already mapped a mixed compute strategy: it planned to use Huawei Ascend chips for smaller R2-model training while retaining Nvidia hardware for its largest models in its earlier Ascend training plan. By September 2026, Liang Wenfeng had reportedly identified training on Huawei chips as one of DeepSeek's biggest bets, alongside expected training-chip deliveries in late 2026 or early 2027 as DeepSeek raised its Huawei training commitment.
The reported joint design work would move the relationship beyond a buyer-supplier arrangement. It ties DeepSeek's model-training requirements more directly to Huawei's forthcoming accelerator roadmap, though both the collaboration and the proposed deployment timeline rest on source reporting.
First-order effects
- DeepSeek would gain direct input into future Huawei AI-chip designs, allowing its pre-training workloads to influence the hardware it expects to use.
- Huawei would secure a demanding model developer as a design partner and prospective large-scale customer for Ascend training hardware, if the reported arrangement proceeds.
Second-order effects
- Huawei's chip roadmap would become more closely coupled to DeepSeek's software and training requirements, making compatibility and optimization part of the commercial relationship rather than a post-release task.
- DeepSeek's earlier split between Huawei for smaller-model training and Nvidia for its largest workloads faces pressure toward a more deliberate multi-vendor compute strategy as Huawei's training chips become available.
Third-order effects
- The partnership points toward an integrated AI stack in which model developers help shape accelerators, potentially shifting competition from stand-alone chip performance toward co-optimized hardware, software, and training workflows.
- If other Chinese AI developers pursue similar arrangements, accelerator vendors may compete for anchor customers' workloads and engineering feedback as intensely as for chip purchase orders.
The trend: AI compute is moving toward co-designed stacks, with model builders seeking greater influence over the chips and systems that determine training capacity.