Reports describe an alleged way to move Samsung and SK hynix high-bandwidth memory (HBM) into China despite U.S. export restrictions: attach the memory to low-performance server processors, ship the combined packages, then separate the components after arrival. The reporting also points to high prices and limited domestic alternatives, but it does not independently establish how widespread the shipments are or confirm every detail of the route.
4
6
How the reported package route works
According to reports citing The Information, traders place HBM on a board or substrate alongside a low-end server processor. The combined package is then exported as a processor-based product. The alleged workaround relies on assessments that focus on the processor’s computing performance rather than the HBM included with it.
4
9
After the packages reach China, technicians reportedly heat them to soften the solder, remove the HBM stacks and discard the processors. The process is described as delicate, with a risk of damaging the memory.
4
6
Some accounts say packages pass through Hong Kong before entering mainland China. These are reports of suspected evasion—not evidence that the shipments are legally exempt from export controls.
10
15
Reported HBM prices and supply pressure
A Shenzhen broker reportedly quoted about $4,000 for one HBM3E unit. Other coverage says that is up from several hundred dollars, but the available reporting does not establish a market-wide price or independently verify the comparison.
8
10
11
The reported supply route comes amid constraints on domestic production. Reuters reported that China’s CXMT planned to begin HBM3E production even as estimated yields stood at 20%–30%, compared with about 70% at SK hynix. Those figures came from people familiar with the matter, not audited company disclosures.
46
What the reports say about Huawei and DeepSeek
The reported shipments are said to reach customers in Huawei’s AI ecosystem. Separate coverage says Huawei has asked some customers to source memory themselves, and reports that the Ascend 960DT is expected in the first quarter of 2027. Those details remain reported plans, not confirmation of actual customer arrangements or product availability.
4
8
10
A separate report says DeepSeek plans to train its V5 model primarily on Huawei Ascend 950 chips, with the project expected to materialize in 2027. The same report characterizes Ascend’s role with DeepSeek so far as primarily inference. The plan has not been confirmed by DeepSeek or Huawei in the evidence available here.
17
If the V5 plan goes ahead, it would be a significant reported expansion of Ascend’s role—but a proposed training project is not proof of a completed, large-scale run. The available reporting does not independently verify training results or demonstrate that China has secured a reliable, self-sufficient supply of advanced AI memory.
17
46