DeepSeek’s priority is strategic resilience rather than near-term chip efficiency: it wants a viable domestic training stack before U.S. restrictions make Nvidia supply even less accessible. The trade-off is substantial—Huawei hardware reportedly remains less efficient for frontier training—but committing real workloads helps improve the chips, software, and supply chain that China needs for long-term autonomy.

• Why persist after the 910C setback: An earlier effort to train DeepSeek’s R2 on Ascend 910C reportedly failed because of hardware/software limitations. The reported estimate that roughly four Huawei processors are needed to match one Nvidia GB300 illustrates the current performance and efficiency gap. But DeepSeek gains supply security, reduces exposure to U.S. policy, and helps create the large-scale deployment feedback needed to close that gap.

• Training-chip timetable: DeepSeek reportedly expects Huawei to begin supplying purpose-built training chips in late 2026, with broader availability around early 2027. Huawei has separately said its Ascend 960DT is scheduled for the first quarter of 2027, while it expects Ascend 950DT systems to see wide use for model training next year. 1

• The 160,000-chip project is principally inference, not yet frontier pre-training: DeepSeek plans to install at least 160,000 Ascend 950DT accelerators at a roughly one-gigawatt site in Ulanqab, Inner Mongolia. That would be among the largest known domestic AI-chip clusters, but reporting says the deployment is intended to run models—inference—rather than replace Nvidia for training immediately; timing depends on Huawei’s production capacity. 2

• The key shift is from proving inference compatibility to making training viable: V4 was released in a version adapted for Huawei chips, demonstrating model-and-software co-design rather than simple porting. This establishes the ecosystem—compiler, frameworks, kernels, networking and model architecture—that Huawei needs before its chips can be dependable for large training runs. 3

• DeepSeek is not betting on Huawei alone: Reuters reported that DeepSeek is also developing its own AI chip, which could reduce dependence on both Nvidia and Huawei. That makes the Huawei relationship a bridge and an ecosystem-building effort, not necessarily DeepSeek’s permanent end-state. 4

• Export controls are reshaping market incentives: U.S. controls limiting access to Nvidia’s most capable AI accelerators have converted domestic substitution from an industrial-policy preference into an operational necessity. Huawei says demand for its AI computing products exceeds supply and that it believes Ascend’s China market share now exceeds Nvidia’s—an important claim, though it is Huawei’s own assessment rather than an independently published market-share measure. 1

• Broader implication: Restrictions may preserve a performance edge for Nvidia at the high end in the short term, but they also accelerate Chinese customers’ investment in domestic chips, software and data-center infrastructure. DeepSeek’s Huawei-optimized V4, planned 950DT-scale inference deployment, prospective training-chip deliveries, and in-house chip effort together support China’s broader objective of a self-sufficient semiconductor and AI-compute stack. 34

DeepSeek’s priority is strategic resilience rather than near-term chip efficiency: it wants a viable domestic training stack before U.S. restrictions make Nvidia supply even less accessible. The trade-off is substantial—Huawei hardware reportedly remains less efficient for frontier training—but committing real workloads helps improve the chips, software, and supply chain that China needs for long-term autonomy.

• Why persist after the 910C setback: An earlier effort to train DeepSeek’s R2 on Ascend 910C reportedly failed because of hardware/software limitations. The reported estimate that roughly four Huawei processors are needed to match one Nvidia GB300 illustrates the current performance and efficiency gap. But DeepSeek gains supply security, reduces exposure to U.S. policy, and helps create the large-scale deployment feedback needed to close that gap.

• Training-chip timetable: DeepSeek reportedly expects Huawei to begin supplying purpose-built training chips in late 2026, with broader availability around early 2027. Huawei has separately said its Ascend 960DT is scheduled for the first quarter of 2027, while it expects Ascend 950DT systems to see wide use for model training next year. 1

• The 160,000-chip project is principally inference, not yet frontier pre-training: DeepSeek plans to install at least 160,000 Ascend 950DT accelerators at a roughly one-gigawatt site in Ulanqab, Inner Mongolia. That would be among the largest known domestic AI-chip clusters, but reporting says the deployment is intended to run models—inference—rather than replace Nvidia for training immediately; timing depends on Huawei’s production capacity. 2

• The key shift is from proving inference compatibility to making training viable: V4 was released in a version adapted for Huawei chips, demonstrating model-and-software co-design rather than simple porting. This establishes the ecosystem—compiler, frameworks, kernels, networking and model architecture—that Huawei needs before its chips can be dependable for large training runs. 3

• DeepSeek is not betting on Huawei alone: Reuters reported that DeepSeek is also developing its own AI chip, which could reduce dependence on both Nvidia and Huawei. That makes the Huawei relationship a bridge and an ecosystem-building effort, not necessarily DeepSeek’s permanent end-state. 4

• Export controls are reshaping market incentives: U.S. controls limiting access to Nvidia’s most capable AI accelerators have converted domestic substitution from an industrial-policy preference into an operational necessity. Huawei says demand for its AI computing products exceeds supply and that it believes Ascend’s China market share now exceeds Nvidia’s—an important claim, though it is Huawei’s own assessment rather than an independently published market-share measure. 1

• Broader implication: Restrictions may preserve a performance edge for Nvidia at the high end in the short term, but they also accelerate Chinese customers’ investment in domestic chips, software and data-center infrastructure. DeepSeek’s Huawei-optimized V4, planned 950DT-scale inference deployment, prospective training-chip deliveries, and in-house chip effort together support China’s broader objective of a self-sufficient semiconductor and AI-compute stack. 34