Huawei Tells Some Customers to Source AI Chip Memory Themselves, Report Says

The Chinese company plans to launch the Ascend 960DT early next year and says the chip will have double the performance of its current processors. But Huawei's Chinese suppliers cannot make enough high-bandwidth memory, or HBM, the stacked memory placed next to AI processors. US export rules also stop it from buying directly from Samsung and SK Hynix, which together hold more than 80 percent of the global market, The Decoder reported.

How the Memory Reaches China

According to the report, smugglers get around the rules by bundling HBM with cheap server processors whose performance is below the US export threshold. The packages go to China through Hong Kong, where workshops separate the memory chips from the processors. The Decoder describes the process as fragile, with a high error rate, but says the margins make it worth the risk.

Prices have risen sharply. A broker in Shenzhen said smuggled HBM3E memory has gone from a few hundred dollars to around $4,000 per unit in a single year. Each AI chip needs six to eight units, so memory alone can cost up to $32,000 per chip.

Smuggling is only part of the supply. The report says older Nvidia GPUs taken apart for their components also feed the market, as does surplus stock from chip designers that scrapped projects or cut orders. By May, South Korea's customs authority had uncovered trade violations worth $560 million, already more than the previous full-year record. About a third of that total involved strategic goods such as chips that were exported without proper authorisation or with false declarations.

Why It Matters

The report suggests memory is now a bottleneck for China's effort to build alternatives to Nvidia's accelerators, alongside the processors themselves. Huawei's Ascend line sits at the centre of that effort. If customers have to source HBM themselves, they take on the cost, supply risk and legal exposure of buying through informal channels, and the price of deploying a complete Ascend system rises.

For policymakers, the bundling method points to a gap in the rules. The Decoder calls it a loophole in US export controls: processors below the threshold can be shipped with restricted memory attached, and the memory is removed once the package arrives. South Korea's record customs figures suggest the trade is not limited to Huawei's supply chain.

The Decoder's report did not include a response from Huawei. For executives buying AI computing capacity, the report shows that memory supply, not just access to processors, now shapes the cost and availability of AI hardware, especially where export restrictions apply.