Huawei Asks Customers to Source HBM for Ascend 960DT Themselves: Black-Market Memory Costs $4,000

Imagine buying a server from a vendor only to find the most expensive component missing from the box, with the vendor suggesting you track it down yourself. That is roughly the situation with Huawei's new AI chips. According to The Information, the company is telling some customers to source memory for their accelerators on their own. In effect, purchasing is being pushed onto the black market.

What happened

Huawei plans to release the Ascend 960DT early next year and promises twice the performance of its current chips. But there is a bottleneck: HBM (High-Bandwidth Memory), without which a modern AI accelerator is useless. Chinese manufacturers cannot produce it in the required volumes. Buying directly from Samsung and SK Hynix, which together hold more than 80% of the global market, is not possible because of US export restrictions.

As a result, a market has emerged to fill the shortage. Smugglers exploit a loophole in the rules: HBM is packed together with cheap server processors whose performance falls below the export-control threshold. The shipment is sent to China via Hong Kong. There, workshops extract the memory from these products. The process is fragile and the defect rate is high, but the margins cover everything.

Besides smuggling, the market is fed by old Nvidia graphics cards that are stripped for parts, and by surplus stock from chip developers that have shut down projects or cut orders.

What about prices

According to a broker in Shenzhen, HBM3E has risen in price over the past year from a few hundred dollars to roughly $4,000 per module. One accelerator needs six to eight modules. Do the math: memory alone costs up to $32,000 per chip. That is no longer a minor markup but the main line item in the budget.

South Korean statistics confirm the scale. By May, customs had identified trade-rule violations worth $560 million, already more than the record for the whole previous year. About a third of that amount involved strategic goods: chips exported without permission or with false declarations.

What this means in practice

For product managers and founders building AI products around Chinese infrastructure, the conclusion is simple: a chip's advertised performance says nothing about its real availability. If memory comes as a separate line item through the gray market, delivery times, price and warranty become a lottery.

For developers optimizing inference for Ascend, there are two risks. First, accelerator fleets may be assembled from memory of unclear origin with a high defect rate, which hurts stability. Second, Ascend 960DT volumes at launch will most likely be limited, which means the tooling ecosystem around it will develop more slowly than the presentation promises.

Limitations

All the data comes via The Information and an anonymous broker, and there is no independent verification of prices. It is unknown how many customers received such an offer from Huawei or in which countries. Huawei itself, judging by the report, has not commented publicly. The $4,000 figure should be treated as a benchmark, not a market quote.

Nevertheless, the signal is strong. US export controls have shifted the bottleneck from compute to memory, and now it is HBM that determines how many AI accelerators can actually be built. Anyone building Chinese hardware into their plans should keep a fallback option in the cloud or on other chips, and spell out in contracts who is responsible for components.