HBM smuggling: Huawei tells clients to source their own AI chip memory - The Next Web

Huawei's decision to push HBM sourcing onto its clients is a compliance severance, not a supply-chain adjustment. Technically, HBM and AI accelerators are co-designed at the silicon level—memory controllers, TSV stacking, and thermal budgets are all tuned to specific HBM generations. By shifting from turnkey delivery to bare-die plus reference design, Huawei forces integrators to independently solve HBM3E packaging, signal integrity, and thermal validation. This stretches time-to-production and erodes delivery competitiveness in the AI compute market. On compliance, the gray-market channel is essentially HBM flowing through unauthorized distributors to restricted entities. Huawei cutting this link transfers export-control legal exposure downstream to the buyer. Mirroring the post-October-2022 BIS rule playbook, Huawei chose liability decoupling over dependency decoupling—reducing near-term litigation risk while exposing a structural gap in its own HBM supply. Strategically, SK Hynix, Samsung, and Micron have HBM capacity pre-committed to NVIDIA and hyperscalers. Huawei's clients flooding secondary markets hand those three suppliers additional spot-price leverage. NVIDIA leverages the "compliant supply chain" label to shift competition from performance to compliance, widening the generational gap. Outlook: Over the next 12–24 months, domestic HBM lines in China face a "demand without system-level validation" window. Decoupled chip-plus-memory procurement becomes the new norm in restricted markets, and the unit of AI compute competition shifts from a single die to system integration—precisely where Huawei is weakest today.