Huawei is asking some customers to source high-bandwidth memory for its AI chips independently, giving businesses that use its hardware a way to assess supply and procurement risks before expanding deployments. The company faces U.S. export restrictions and a shortage of HBM, a critical memory component, as it prepares to launch the Ascend 960DT early next year.
The shift places procurement responsibility on buyers at a time when smuggled HBM3E units reportedly cost about $4,000 each. Each AI chip requires six to eight units, so the memory alone can cost as much as $32,000 per chip.
Why Is Huawei Asking Customers to Find Their Own Memory?
Huawei plans to introduce the Ascend 960DT early next year and says the chip will double the performance of its current products. Its Chinese suppliers cannot produce enough high-bandwidth memory to meet demand.
Direct purchases from Samsung and SK Hynix are also restricted by U.S. export rules. According to the report, those two companies control more than 80 percent of the global HBM market, leaving Huawei’s Chinese supply chain with limited access to the memory required for advanced AI accelerators.
Huawei is asking certain customers to arrange their own memory. Buyers planning to deploy Huawei-based AI infrastructure now need to check HBM availability, cost and compliance as part of the hardware purchasing process.
How Are Smugglers Moving HBM Into China?
Smugglers may have developed a network around the gap between restricted HBM supply and demand inside China. According to the report, smugglers pair HBM with low-end server processors that fall below the applicable export threshold, then ship the packages through Hong Kong.
Workshops in China then separate the memory from the processors. The report describes the process as fragile, with a high error rate, but the strong margins make the operation attractive to those involved.
The black market also receives components from older Nvidia GPUs that have been stripped for parts. It can also draw on excess inventory from chip designers that cancelled projects or reduced orders.
How Much Has the Memory Shortage Affected Prices?
According to a broker in Shenzhen, prices for smuggled HBM3E memory rose from a few hundred dollars to around $4,000 per unit in one year. With six to eight units needed for each AI chip, the memory component can reach a cost of $32,000 per chip.
These prices are for smuggled HBM3E units, not memory bought through authorized suppliers. Buyers can compare the reported $4,000 per unit with the $32,000 memory cost for one AI chip before expanding deployments.
What Does the Export Crackdown Mean for Business Visibility?
South Korean customs authorities had uncovered trade violations worth $560 million by May, exceeding the previous full-year record. One third of that total involved strategic goods, including chips exported without proper authorization or with false declarations.
A Huawei-based AI project may reportedly need six to eight HBM units for each AI chip. Announced chip performance does not confirm component availability or compliant sourcing.
For Huawei’s Ascend 960DT, a shortage of HBM and restricted sourcing can affect component availability. Product and service pages should state actual availability and avoid presenting deployment dates or capabilities as confirmed when essential components remain uncertain.
What to Do Now
- Map the full hardware requirement. Include memory quantity and supplier dependencies in each planned AI deployment. Huawei’s reported requirement of six to eight HBM units per AI chip is a useful starting point for procurement planning.
- Check authorization before purchasing. Consider asking suppliers for documentation covering the origin and export status of HBM and other strategic components. The reported use of false declarations and unauthorized exports shows why procurement records matter.
- Model several memory prices. Compare the reported smuggled-unit price of about $4,000 with the potential $32,000 memory cost for one chip. Use the scenario that reflects the chip configuration and supplier terms relevant to the project.
- Align launch plans with supply. Treat the Ascend 960DT target of early next year as a company plan, not a guaranteed delivery date for every buyer. Confirm component availability and the broader system schedule before promoting a launch.
- Publish accurate availability details. Update product, service and launch pages when supply changes. Clear information helps customers, search engines and AI systems distinguish confirmed availability from a projected rollout.
What Is HBM?
High-bandwidth memory is a memory component placed close to an AI accelerator so that processors can transfer data more efficiently during AI workloads. HBM3E is a version of this memory described in the report as increasingly valuable because of supply constraints.
How Does Memory Availability Affect AI Chip Performance?
An advanced AI accelerator also needs sufficient, compatible and properly sourced memory to operate as intended. Huawei says its planned Ascend 960DT will double the performance of its current chips, while the report indicates that each chip requires six to eight HBM units. The claimed performance gain and the availability of its components are separate procurement questions.
FAQ
Why Is Huawei Asking Customers to Source HBM Memory?
Huawei is asking some customers to find their own high-bandwidth memory because Chinese suppliers cannot produce enough and U.S. export rules restrict direct purchases from Samsung and SK Hynix.
How Much Does HBM Memory Cost in the Black Market?
A broker in Shenzhen said smuggled HBM3E units rose from a few hundred dollars to around $4,000 each in one year. At six to eight units per AI chip, the memory alone can cost up to $32,000.
When Does Huawei Plan to Launch the Ascend 960DT?
Huawei plans to launch the Ascend 960DT early next year. The company says the chip will double the performance of its current chips.
Related coverage
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.