Sources: Huawei, Baidu, and other Chinese tech companies have been stockpiling Samsung HBM chips since early 2024, in anticipation of US curbs on those chips
Context & Ripple Effects
Huawei had already used inventory-building and supply-chain investment to cushion earlier US restrictions, including a reported stockpile of crucial US chips. This report extends that playbook from processors and components to high-bandwidth memory, a specialized input for AI systems.
The buying also follows reported large Chinese orders for Nvidia chips amid concern about tighter export rules, while Washington was reportedly considering limits on China-bound AI memory chips. Samsung is therefore exposed not only as a memory vendor but as a near-term channel for demand being pulled forward by policy risk.
First-order effects
- Huawei, Baidu and other named Chinese buyers gain a temporary inventory buffer against a potential interruption in Samsung HBM supply.
- Samsung receives demand that may be accelerated by export-control risk rather than paced solely by customers' immediate deployment needs.
Second-order effects
- Stockpiling can tighten the available supply pool for other HBM buyers and makes suppliers' allocation decisions more consequential if restrictions arrive.
- Potential curbs would push Chinese AI buyers to secure alternative approved supply and reinforce their incentive to reduce dependence on imported memory and related AI components.
Third-order effects
- If repeated across chips and memory, export controls can segment AI hardware supply chains into inventories, approved channels and domestic substitutes rather than a single global market.
- The pattern raises the strategic value of memory supply alongside AI accelerators: restricting compute without constraining its memory inputs may leave a route for continued system build-outs, though the scale depends on any final rules.
The trend: AI hardware trade controls are broadening from accelerators toward the supporting memory and supply-chain inputs that determine usable compute capacity.