Overview

• According to multiple reports, DeepSeek’s efforts to train its R2 model on Huawei Ascend chips were derailed by persistent hardware and software failures, prompting a shift back to Nvidia GPUs for the training phase.
• Engineers cited unstable Ascend accelerators, inadequate interconnect speeds, and immature CANN software as insurmountable obstacles that prevented completion of a single training run.
• Financial Times sources say Chinese authorities had encouraged adoption of homegrown Ascend hardware to reduce reliance on US technology under tightening export controls.
• Extended data-labeling requirements also contributed to pushing back R2’s expected May debut, and although some Chinese outlets suggest a launch in the coming weeks, DeepSeek and Huawei have not confirmed any timeline.
• The setback underscores the challenges of building an indigenous AI training stack and reaffirms Nvidia’s entrenched role in powering large-scale model development amid geopolitical supply-chain tensions.