Korean HBM reportedly diverted to China for Huawei customers | Magica
Korean HBM reportedly reaches Huawei customers through processor packages
Editorial Team
••📖7 min read
Traders reportedly extract Samsung and SK hynix memory from processor packages shipped into China, selling it mainly to Huawei customers. The account highlights the memory constraint behind Huawei’s AI chip rollout, but leaves trade volumes, recovery yields and transaction legality unresolved.
Traders reportedly ship Korean high-bandwidth memory alongside low-end processors, then strip out the memory in China.
Some Huawei customers have reportedly been asked to obtain their own memory for its next flagship AI processor.
A broker’s $4,000 HBM3E price illustrates the procurement pressure; the volume of usable memory getting through remains unknown.
Traders are reportedly moving high-bandwidth memory made by South Korean chipmakers Samsung Electronics and SK hynix into China inside packages containing inexpensive server processors. Technicians then extract the memory and sell it mainly to customers of Huawei Technologies, the Chinese telecommunications and IT company behind the Ascend family of AI processors.
The account comes from November 2 reporting by The Information, relayed in a November 3 report. Five Korean news accounts draw on that same underlying reporting, rather than independently confirming the transactions.
The immediate commercial question is whether Huawei’s customers can secure enough memory to support its next chip rollout. People familiar with Huawei’s supply chain said the company told some customers its upcoming Ascend 960DT would rely substantially on Samsung and SK hynix HBM, and asked some buyers to obtain the memory themselves, according to the supply-chain account. Neither shipment volumes nor the proportion of customer demand met by the reported trade is disclosed.
David Wang speaks at HUAWEI CONNECT in Shanghai on September 17, 2026, where Huawei announced its company-reported Ascend roadmap. Photo: Huawei. Source: Huawei.
Huawei’s processor roadmap still needs memory
High-bandwidth memory, or HBM, stacks multiple dynamic random-access memory dies vertically to move large quantities of data quickly. Its role is distinct from the processor’s: insufficient memory capacity or bandwidth can prevent an advanced processor from delivering its full capabilities in AI training and inference.
Huawei already supplies AI computing systems. In its September 17 announcement, the company said more than 1,000 Atlas 900 A3 SuperPoDs—systems that tightly connect multiple computing nodes—had been deployed, and that its Atlas 950 systems were seeing large-scale commercial use. Those are company-reported deployment claims, but they establish that the memory issue concerns an existing infrastructure business as well as a future chip.
Xiaomi says its humanoid robot sorted car parts for a full shift and corrected errors autonomously. The result extends its factory trials, but published evidence does not yet establish deployment costs or reliability across repeated shifts.
Editorial Team
Huawei said the Ascend 960DT would be available in the first quarter of 2027, three quarters ahead of its original roadmap. It also said Ascend 960 performance was doubling as planned. That is a company performance claim, not an independently verified benchmark, and the announcement does not confirm the memory procurement arrangements described by supply-chain sources.
The reports describe limited Chinese HBM production and a global market dominated by Samsung and SK hynix. They do not quantify how much domestic Chinese memory could substitute for imported supplies. Developing a domestic processor therefore does not, on this evidence, establish a fully domestic supply chain.
Packaging changes the classification, not every obligation
According to the reported method, traders package HBM alongside low-end server processors on a board and ship the combined product under a processor classification. In China, technicians heat the packages to melt the solder, recover the memory and discard the processors. Poorly controlled heat can warp or crack the memory stacks; the reports describe high failure rates without providing a measured recovery yield.
The U.S. Commerce Department’s Bureau of Industry and Security, which administers the Export Administration Regulations, added HBM-specific controls in its December 2024 rule. It also brought qualifying foreign-produced HBM within the scope of its advanced-computing foreign-direct-product rule, explaining how the controls can reach memory manufactured outside the United States. BIS said the HBM restrictions were intended to slow China’s efforts to produce advanced AI chips using imported memory, citing military, intelligence and surveillance risks.
The rule draws a distinction relevant to the reported workaround. Certain integrated circuits combining HBM and logic are excluded from the HBM-specific classification when processing is their dominant function. But BIS explicitly says such chips may fall under other advanced-computing classifications, depending on processing performance and performance density.
The same rule’s temporary authorization for certain HBM-related activities also says subsequent exports, reexports or transfers must comply with other applicable license requirements, including those tied to end use and transaction parties. An exclusion from the standalone-memory category therefore does not establish that a particular shipment or later transfer is permitted.
The accounts do not identify processor models, export documents or individual transactions sufficient to determine how those requirements applied. They describe an alleged diversion mechanism, not a regulatory finding that the route is lawful or that specific parties violated the rules.
A costly recovery route meets a wider capacity squeeze
A broker interviewed by The Information said a single HBM3E unit traded in China for about $4,000, up from several hundred dollars a year earlier. The account of those prices describes AI chip packages using six to eight memory units. Multiplying those counts by the broker’s price produces an illustrative memory bill of $24,000–$32,000, before the processor, servers and electricity.
That is not a representative market survey or a verified bill of materials for the Ascend 960DT. The reports describe substantial profits, but do not supply the acquisition costs, processing expenses or recovery yields needed to calculate a margin. The quoted price alone cannot establish the economics of each shipment.
Export restrictions are only part of the supply problem. Research firm TrendForce’s June 2026 analysis estimated that HBM would consume a much larger share of the three leading suppliers’ DRAM wafer input than its share of their DRAM bit supply:
Period
HBM share of total DRAM wafer input
HBM share of total DRAM bit supply
End of 2026
About 22%
About 9%
End of 2027
About 30%
About 13%
These are TrendForce’s projections, not observed year-end results or measures of diverted trade. Wafer input measures manufacturing allocation; bit supply measures memory output. The figures illustrate HBM’s disproportionate demand on factory capacity, rather than a directly interchangeable volume of conventional memory.
Suppliers also have competing incentives. TrendForce estimated that HBM’s per-wafer revenue and profitability had fallen below those of DDR5 64GB registered memory modules in the first quarter of 2026, using die size, yield and per-gigabit pricing in its analysis. It expected suppliers to adjust production allocation between HBM and conventional DRAM depending on HBM pricing negotiations. That limits any assumption that higher demand automatically produces enough additional HBM supply.
TrendForce’s June 2026 projections compare HBM’s shares of DRAM wafer input and bit supply among the top three suppliers. The measures have different denominators and do not quantify diverted shipments. Source: TrendForce.
What would establish the route’s significance
The reported trade raises a tracing problem for the Korean manufacturers. The accounts warn that finding their memory in Chinese AI chips could invite U.S. scrutiny, citing the 2024 U.S. investigation of Taiwanese contract chipmaker TSMC after chips it manufactured were found in Huawei products. That precedent does not establish liability for Samsung or SK hynix in the reported HBM transactions.
SK hynix said it complies with U.S. and South Korean export-control rules and has not supplied Huawei since U.S. restrictions took effect in 2020, according to the reported company response. The accounts do not establish that either Korean manufacturer knowingly participated in the diversion.
One broker described the trade as service to China; another said widening use of the same routes increased detection risk and planned to sell most of the business’s equity next year. These are individual traders’ accounts of their incentives and expectations, not evidence of state direction or an announced enforcement action.
Huawei’s planned first-quarter launch makes the missing supply evidence consequential. Shipment records, tested recovery yields and verified deliveries to customers would show whether the route supplies a marginal quantity of usable memory or a meaningful share of the rollout’s needs. Regulatory findings would be needed to establish responsibility for particular transactions. Until then, the reports describe a possible source of scarce memory, without showing that it can reliably sustain Huawei’s production plans.
Railway’s CEO apologized after an engineer’s AI agents targeted rival Primcloud during testing. Primcloud says attempts to obtain customer secrets failed, leaving questions about the agents’ instructions and Railway’s controls.
Cortex Training gives developers custom training APIs on managed GPUs. Snowflake is extending its existing model-customization tools, with enterprise-data integration and a cost claim limited to LoRA fine-tuning.
Apple’s developer preview extends Siri AI beyond predefined app categories and adds guidance for complex tasks. Testing is limited to development builds and internal testers, with App Store submissions planned for spring 2027.
Jill Hazelbaker will join OpenAI as chief corporate affairs officer after a decade at Uber, bringing experience in technology communications and political campaigns to a role focused on reputation, positioning and public trust.
The watchOS 27.3 developer beta brings Live Rewind and Siri Recap to Apple Watch Series 12 and Ultra 4. The opt-in tools offer brief transcripts and conversation summaries, but require a compatible iPhone and carry Apple warnings about accuracy and sensitive information.
Creative-industry groups are challenging OpenAI and Microsoft’s arguments against licensing AI training material, as a New York court weighs fair use. Their briefs share an emphasis on licensing but differ over what the judge should decide.
Traversal's Series B backs its expansion from diagnosing software outages toward autonomous production operations. Company-reported adoption, cloud competition and unfinished access controls frame the next tests.
The startup formerly called TesterArmy is replacing its testing agent with an open-source framework, betting that adoption will generate demand for paid infrastructure and enterprise services.
Fal has announced Sequoia-led and Bessemer-led funding rounds for its 2027 expansion, adding post-trained media models to an established inference business. The announcements leave the new rounds’ size and comparative serving economics undisclosed.
SK hynix bought existing Penguin Solutions preferred shares from an SK Telecom investment vehicle. The deal brings director nomination rights and a convertible stake in an existing AI infrastructure partner, while sending the purchase price to the seller.
Google plans to switch off Gemini and Search for school-managed K–12 student accounts unless administrators enable them. Schools will gain control over access, but Google says keeping Search without its AI features will not be an option.
Anthropic is expected to exceed 3.5 GW of computing capacity in 2026 and aims to double it in 2027. Its filing shows $895.4 billion in commitments, substantial financial resources and an immediate cost from scarce capacity.
Independent tests reproduce MIT researchers’ repair of Qwen’s repetitive reasoning. The edited model still beats a stronger unedited configuration, but by much less; broader claims about creating capabilities remain unproven.
Gartner forecasts that open-weight models will execute 90% of enterprise operational agent steps by 2030. Its cost estimates favor routine, high-volume work, but infrastructure costs and the choice of comparison model leave enterprise savings to be proved.
FTAI Aviation is financing a planned 100-turbine production run in 2027, while ProEnergy has a 650 MW equipment order from Crusoe. Converted aircraft engines offer another supply route for data-center power, but delivery targets, affiliated financing and unproven operating economics qualify the opportunity.
Meta is marketing dollar bonds in a plan to borrow at least $20 billion, including an expected euro offering. Capital spending exceeded operating cash flow last quarter, but the debt is part of a wider financing strategy that also includes equity offerings and a proposed non-voting share-sale program.
A nonbinding opinion targets real-time retrieval and recognisable reuse of journalism, while rejecting a monopoly over news facts and recommending that questions about model training be dismissed as hypothetical.
A study of Lovable projects found passport details, customer records and private messages accessible without login, raising questions about database permissions and the limits of automated security checks. Its sample does not establish a platform-wide leak rate.
CrowdStrike is extending Falcon Guardian into Claude Cowork’s virtual machine and AWS’s AI guardrail layer. The Cowork beta addresses a visibility gap, while existing AWS safeguards and unfinished acquisitions limit what buyers can infer about the broader platform’s capabilities.
