Can China train frontier models at scale on domestic hardware?

China can already run competitive AI models on domestic hardware, but it still has not demonstrated a frontier-scale training run using a fully domestic stack.

Frontier training requires thousands of accelerators to work together reliably at scale. That coordination depends on fast connections between chips, sufficient memory and stable software. A weakness anywhere in the system can stop the run.

China still relies on Nvidia stockpiles and approved imports for frontier training while Huawei develops domestic infrastructure around Ascend. Bringing that infrastructure to scale will also require more chip output from SMIC and continued progress in domestic memory and networking.

The 2027–28 model generation is the first that could plausibly be trained on Huawei’s domestic infrastructure. A fully disclosed training run would confirm that China can train frontier models without foreign hardware.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。