Huawei’s plans for new AI chips are running into a shortage of a crucial component: high-bandwidth memory. The company is reportedly asking some customers to obtain that memory themselves, a move that can push buyers toward unauthorized supply channels.
Memory supply is constraining chip plans
Huawei plans to launch the Ascend 960DT early next year. The company says the chip will double the performance of its current chips, but that ambition depends on access to enough High-Bandwidth Memory (HBM).
Huawei’s Chinese suppliers cannot produce sufficient quantities, while U.S. export rules block direct purchases from Samsung and SK Hynix. Together, those companies control over 80 percent of the global market. That combination leaves Huawei and its customers facing a gap between the memory needed for the chips and what can be obtained through authorized channels.
Asking customers to source memory shifts part of the procurement problem onto them. It does not make the component easier to obtain; it gives buyers responsibility for finding supply amid restrictions and shortages. The reported result is greater reliance on a black market that has developed to meet demand.
Smugglers use a workaround
The reported smuggling method pairs HBM with low-end server processors whose performance falls below the export threshold. The packages are routed through Hong Kong into China, where workshops separate the memory chips from the processors.
The arrangement takes advantage of the way the goods are bundled and shipped. Once the shipment reaches China, the memory can be recovered for use in AI hardware. The process is fragile, and the error rate is high, but the profits are large enough to make the risk worthwhile for those involved.
This route illustrates how a supply restriction can create incentives for intermediaries to find indirect paths. It also makes procurement less predictable: memory may be available, but obtaining it can depend on a complicated chain in which some of the work is technically difficult and prone to failure.
Scarcity pushes costs higher
A broker in Shenzhen said prices for smuggled HBM3E have risen from a few hundred dollars to around $4,000 per unit in just one year. Each AI chip needs six to eight units, putting the memory cost alone as high as $32,000.
That cost changes the economics of building AI hardware. Memory is not an incidental add-on when a chip needs multiple units and each one has become far more expensive. Buyers must account for both the price and the difficulty of securing enough supply.
The steep price increase also shows how scarcity can reward unauthorized channels. When customers cannot rely on ordinary procurement, the cost of the scarce component can rise sharply, even as the recovery process remains unreliable.
The informal market draws on several sources
Smuggling is only one part of the supply network described in the report. Older Nvidia GPUs are being stripped for parts, and excess inventory from chip designers who scrapped projects or cut orders also enters the market.
South Korea’s customs authority had uncovered trade violations worth $560 million by May, already exceeding the previous full-year record. A third of that total involved strategic goods such as chips exported without proper authorization or with false declarations.
These figures point to a broader enforcement challenge. Supply can come from newly routed shipments, recovered components, or inventory left over from changed business plans. Tracking each source and verifying that exports are properly declared becomes harder when components move through multiple hands.
For Huawei’s customers, the immediate issue is whether they can secure enough HBM for planned AI chips, and at what cost. For the wider market, the reported activity shows how restricted access and limited production can redirect demand into channels where prices, quality, and delivery are less dependable.