⚡ Uncle Cat AI Radar
IndustryPolicy

Huawei Customers Told to Source HBM for AI Chips

Huawei is reportedly asking some buyers to procure high-bandwidth memory themselves, exposing a severe constraint in China’s AI-chip supply chain.

Huawei has reportedly told some customers ordering its AI chips to source their own high-bandwidth memory, or HBM, according to reporting from The Decoder and South Korea’s Herald Business. The unusual instruction reflects a supply-chain bottleneck: Chinese companies face difficulty obtaining advanced HBM directly from Korean memory suppliers, while domestic alternatives remain constrained in availability and performance.

HBM is not a peripheral component. It sits close to the accelerator and feeds data to the compute cores at very high bandwidth, making it central to training and inference performance. A chip design can be competitive on paper yet remain difficult to deploy at scale if the surrounding memory supply is insufficient.

The reported arrangement effectively shifts procurement risk from Huawei to buyers. Customers may need to locate compatible memory through their own channels before Huawei can complete a system. That raises questions about qualification, packaging, warranty responsibility, and the ability to build repeatable server configurations. It also creates incentives for intermediaries and gray-market procurement, precisely the kind of opaque supply chain that export controls seek to disrupt.

The reports cite memory costs of roughly $24,000 to $32,000 per package in the Chinese market before adding the accelerator, server, or electricity. Those figures should be treated cautiously because the underlying supply contracts and configurations are not public. Huawei has not, in the cited coverage, publicly confirmed a universal requirement applying to all customers or products.

Why it matters: China’s AI hardware challenge is becoming less about designing a substitute accelerator and more about assembling an entire reliable system around it. If buyers must independently secure HBM, scale-up becomes slower, more expensive, and harder to standardize. The development also shows why memory suppliers, packaging capacity, and procurement controls can matter as much as the headline chip itself.

Sources