Smuggled HBM3E Hits $4,000 a Unit; Huawei Tells Clients to Buy Their Own

Smuggled HBM3E memory now costs roughly $4,000 per unit on China's grey market, up from a few hundred dollars a year ago. A Shenzhen-based broker says the price surge, driven by U.S. export controls, leaves Chinese AI chip builders facing memory bills of up to $32,000 per processor. Huawei, the largest affected buyer, has started telling some customers to secure their own supply.

The arithmetic is stark. An AI accelerator needs six to eight HBM modules. At current street prices, memory alone can reach $32,000 per chip. That figure comes from a single broker in Shenzhen, so treat it as an indicative quote rather than a market average. No exchange publishes prices for goods that move outside the law.

Where does the smuggled memory come from?

The parallel supply draws from three streams, according to the reporting. Organized smuggling is one. Dismantled older Nvidia GPUs contribute salvageable HBM. Surplus inventory leaked from chip designers also feeds the market.

The most inventive route exploits a loophole in U.S. rules. Networks pair HBM modules with underperforming server processors that fall below export thresholds. The combined sets ship legally on paper, routed through Hong Kong. Workshops in China then separate the memory from the processors.

The operation is fragile. The source describes a high error rate in the dismantling work. Yet the margins are wide enough that the practice keeps growing, and a sophisticated smuggling network has formed around it.

How large is the illegal trade?

South Korean customs offers the hardest numbers in this story. Authorities reported $560 million in trade violations by May. That total already exceeds the previous annual record, with a full year still to run. One-third of the violations involve strategic goods, including chips exported without authorization or under false declarations.

South Korea is now recording record levels of illegal chip exports. The country hosts the two most important HBM suppliers, Samsung and SK Hynix, which together control more than 80 percent of the global market. That concentration matters. Most of the world's HBM passes through a jurisdiction that is now a documented smuggling choke point.

What is Huawei telling customers?

Huawei cannot obtain enough HBM for its AI chips. Its domestic Chinese suppliers are not producing sufficient volumes. Direct purchases from Samsung or SK Hynix are blocked by U.S. export rules, and no workaround within legal channels has emerged.

The company's response, according to a source, is to shift the problem downstream. Huawei has informed some customers that they must acquire the necessary memory for its AI chips themselves. The claim rests on a single unnamed source. Huawei has not confirmed it publicly, and the scope of the instruction, whether it covers all customers or a subset, remains unclear.

For buyers, the practical effect is severe. A customer must now navigate the same constrained market, at the same inflated prices, without Huawei's negotiating weight. Some will inevitably turn to the parallel channels described above. The policy, if accurate, risks pushing Huawei's client base directly into the grey trade.

Can the Ascend 960DT change the picture?

Huawei says it will launch the Ascend 960DT early next year. The company claims the chip will deliver double the performance of its current models. That is a roadmap claim, not a shipping product, and no independent benchmark exists yet.

Note what the announcement does not address. Huawei has said nothing about the memory configuration of the 960DT. A performance doubling means little if each unit requires memory that customers cannot legally buy or can only source at $4,000 per module. The HBM constraint, not the logic design, now sets the ceiling on Chinese AI accelerator output.

Unless domestic HBM supply scales or export rules shift, the gap between announced chip performance and deliverable systems will widen, and the $32,000-per-chip memory bill will remain the real price of Chinese AI compute.