Sources: DeepSeek plans to use Huawei's Ascend AI chips to train smaller versions of its upcoming R2 models but will still use Nvidia chips for largest models
The Information :
Context & Ripple Effects
DeepSeek's R2 hardware plan follows an earlier R2 delay tied to Ascend training issues, making the division of workloads a practical response to differing chip capabilities rather than a clean platform switch.
The broader coverage shows DeepSeek continuing to cultivate alternatives: Huawei-linked researchers later reported Ascend 910C use for V4 post-training, while DeepSeek also began exploring an in-house inference chip.
First-order effects
- DeepSeek can put smaller R2 training runs on Huawei Ascend hardware while retaining Nvidia for the largest training jobs, reducing the operational risk of relying on either platform alone.
- Huawei gains a meaningful training workload and associated software-validation opportunity; Nvidia retains the highest-end portion of DeepSeek's model-development compute.
Second-order effects
- DeepSeek must maintain model-training workflows across two hardware and software environments, raising the value of tooling, optimization work, and portability between stacks.
- Huawei has an incentive to close the gaps exposed by large-model training, while Nvidia's position is reinforced where scale and training reliability are decisive.
Third-order effects
- If this workload segmentation persists, AI developers may increasingly allocate compute by task size and stage rather than standardize on one accelerator vendor.
- The pattern points to a more heterogeneous AI-compute market: alternative chips can win bounded workloads first, while leadership in frontier-scale training remains harder to dislodge.
The trend: AI labs are adopting heterogeneous compute strategies, using alternative accelerators for workloads they can support while reserving top-tier hardware for the most demanding training.