Huawei is reportedly asking some customers to arrange their own high-bandwidth memory (HBM) for its upcoming artificial intelligence processors, highlighting the growing impact of US export restrictions and China’s limited domestic memory production. The move could push buyers towards unofficial supply channels as the company struggles to secure enough memory for its next-generation Ascend AI chips.
According to a report by The Information, cited by The Decoder on November 3, 2026, smugglers are exploiting a loophole in export rules by shipping restricted memory alongside lower-performance server processors. Once the packages reach China, technicians reportedly extract the memory and discard the processors. Smuggled HBM3E is reportedly selling for around $4,000 per unit, creating another challenge for China’s AI chip industry.
Why Huawei Is Asking Customers to Find Their Own AI Memory
Huawei’s reported decision reflects a major shortage of high-bandwidth memory needed to manufacture advanced AI processors.
The company is preparing its next-generation Ascend 960DT AI accelerator, which is expected to launch in early 2027. Huawei reportedly plans to use substantial quantities of HBM manufactured by South Korean companies Samsung Electronics and SK Hynix.
However, US export restrictions prevent these memory manufacturers from directly supplying advanced HBM to Huawei.
China’s domestic memory manufacturers are also struggling to produce enough advanced memory to satisfy demand.
As a result, Huawei has reportedly informed some customers who ordered its AI chips that they will need to arrange the required memory themselves.
This shifts an important part of the procurement process from Huawei to its customers, creating additional sourcing, pricing and regulatory challenges.
It is important to note that the report does not establish that Huawei itself operates or directly participates in smuggling operations. The reported arrangement concerns customers sourcing memory through their own channels.
How Smugglers Are Bringing Restricted HBM Into China
The reported smuggling operation uses a workaround involving high-bandwidth memory packaged alongside relatively inexpensive server processors.
Under the described method, traders combine HBM chips with lower-performance processors and ship the packages into China, including through Hong Kong.
The workaround exploits differences in how export restrictions assess certain packaged semiconductor products.
Once the packages reach China, technicians separate the memory from the processors.
The process involves heating the packages until the solder softens, allowing workers to remove the HBM stacks while discarding the accompanying processors.
This is a delicate operation because excessive heat can damage the memory components.
The process reportedly has a high failure rate, but the potential profits make it attractive to traders.
How the Reported Smuggling Process Works
| Stage | Activity |
|---|---|
| 1. Procurement | Traders obtain HBM from South Korean supply channels |
| 2. Packaging | HBM is combined with lower-performance server processors |
| 3. Transportation | Packages are routed into China, including through Hong Kong |
| 4. Extraction | Technicians heat the packages and separate the memory |
| 5. Distribution | Recovered HBM is sold to customers requiring advanced AI memory |
The reported method highlights how semiconductor export restrictions can create incentives for alternative supply chains and complicated procurement arrangements.
Smuggled HBM3E Prices Reach $4,000 Per Unit
The shortage has reportedly caused black-market HBM prices to increase sharply.
According to a Shenzhen-based broker quoted by The Information, one HBM3E unit that previously sold for several hundred dollars is now trading for approximately $4,000.
An advanced AI processor requires multiple HBM units to provide the memory bandwidth needed for demanding computing workloads.
The reported cost can therefore become substantial.
| Memory Cost Component | Reported Figure |
|---|---|
| Price per HBM3E unit | $4,000 |
| Memory units per AI chip | 6–8 |
| Cost for 6 units | $24,000 |
| Cost for 8 units | $32,000 |
This means memory alone could cost between $24,000 and $32,000 for an AI chip package, based on the reported prices and quantities.
These figures come from an individual broker quoted in the original report and should not be interpreted as an independently verified market-wide price.
The high cost also demonstrates how memory shortages can undermine the economics of AI infrastructure, even when domestic companies successfully develop competitive processors.
Huawei’s Ascend 960DT AI Chip Targets Higher Performance
Huawei’s next-generation Ascend 960DT is expected to become one of China’s most powerful domestically developed AI processors.
The company reportedly plans to introduce the chip in early 2027.
Huawei has indicated that the processor could deliver approximately twice the performance of its current flagship AI chips.
The Ascend 960DT is also expected to use substantial amounts of HBM supplied by Samsung and SK Hynix, according to people familiar with Huawei’s supply chain.
However, the memory shortage could complicate production and limit the number of processors Huawei can deliver to customers.
This is particularly important because advanced AI accelerators depend on more than computing power.
Their performance also relies heavily on memory bandwidth, memory capacity and the ability to move large quantities of data between processors and memory.
Without sufficient HBM, even a powerful AI processor can struggle to deliver its intended performance.
Samsung and SK Hynix Control Most of the Global HBM Market
South Korean semiconductor manufacturers Samsung Electronics and SK Hynix play a central role in the global high-bandwidth memory industry.
Together, the companies account for more than 80% of the global HBM market, according to the reported figures.
HBM is particularly important for AI accelerators because it allows processors to access large amounts of data at extremely high speeds.
Unlike conventional memory modules, HBM uses vertically stacked memory dies connected through advanced packaging technologies.
This design allows significantly greater data bandwidth while maintaining a compact physical footprint.
The technology is widely used in advanced AI processors from companies such as Nvidia and AMD.
However, US export restrictions have prevented direct shipments of certain advanced HBM products to Chinese customers.
This has left Chinese AI chip manufacturers searching for alternative suppliers while domestic memory producers work to improve their technology and manufacturing capacity.
China’s Domestic Memory Industry Struggles to Fill the Gap
China has been investing heavily in semiconductor manufacturing to reduce its dependence on foreign technology.
However, developing advanced HBM remains a difficult challenge.
Domestic manufacturers have made progress in memory production, but manufacturing high-performance HBM requires specialised technologies, advanced packaging capabilities and consistently high production yields.
According to reporting cited by TrendForce, China’s CXMT has been experiencing challenges with HBM3E production yields.
Reuters has separately reported that early HBM3E production yields at CXMT remain significantly below those achieved by leading South Korean manufacturers.
Low production yields mean a larger proportion of manufactured memory fails quality requirements, increasing costs and reducing the amount of usable supply.
This makes it difficult for domestic suppliers to replace Samsung and SK Hynix at the scale required by China’s growing AI industry.
South Korea Reports Rising Semiconductor Export Violations
The growing demand for restricted semiconductor components has also increased concerns about illegal exports.
South Korea’s customs authorities reportedly uncovered trade violations worth approximately $560 million by May 2026, already exceeding the previous full-year record.
Around one-third of the reported violations involved strategic goods, including semiconductor products exported without proper authorisation or through false declarations.
The figures highlight the enforcement challenges created by increasingly complex semiconductor supply chains.
Memory manufacturers, distributors, shipping companies and other intermediaries face greater scrutiny as governments attempt to prevent restricted technologies from reaching unauthorised buyers.
For South Korean companies, the issue is particularly sensitive because their products remain essential to the global AI infrastructure industry.
What This Means for China’s AI Industry
Huawei’s reported decision to ask customers to source their own memory highlights a broader weakness in China’s AI semiconductor ecosystem.
The country has made considerable progress in developing domestic AI processors, but producing a competitive chip requires access to several critical components.
High-bandwidth memory is one of the most important bottlenecks.
If memory remains difficult to obtain, Chinese AI companies could face higher manufacturing costs, slower infrastructure expansion and uncertainty around delivery schedules.
The situation could also encourage more investment in domestic HBM manufacturing and alternative memory technologies.
For AI companies, the issue extends beyond Huawei. Chinese developers building large-scale AI systems need reliable access to accelerators and memory capable of supporting demanding model-training and inference workloads.
The shortage could therefore affect the broader expansion of AI data centres across China.
The Bigger Picture
The Huawei memory shortage demonstrates how semiconductor export restrictions can influence the entire AI hardware supply chain. Limiting access to advanced processors is only one part of controlling AI computing capabilities. High-bandwidth memory, advanced packaging and manufacturing equipment are equally important components of modern AI infrastructure.
The reported black-market activity also shows how shortages can create alternative procurement networks, although these channels come with significant financial, technical and legal risks. For Huawei, asking customers to arrange their own memory could help address immediate supply problems, but it also shifts procurement uncertainty and potential compliance exposure towards buyers.
Ultimately, China’s ability to compete in the global AI chip market will depend not only on designing powerful processors but also on building a reliable domestic ecosystem for memory, packaging and semiconductor manufacturing.
Looking Ahead
Huawei’s upcoming Ascend 960DT launch in early 2027 will be an important test of its ability to overcome memory supply constraints. The company’s reported strategy of asking customers to source HBM independently could provide some flexibility, but the high prices and risks associated with unofficial procurement may limit its effectiveness.
Meanwhile, South Korean memory manufacturers are likely to remain strategically important to the global AI industry. China’s efforts to improve domestic HBM production will be closely watched as companies attempt to reduce dependence on foreign suppliers. The longer-term outcome will help determine whether export restrictions can slow China’s AI hardware expansion or accelerate its investment in semiconductor self-reliance.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
