Smugglers are packaging high-bandwidth memory (HBM) made in South Korea together with low-end server processors to ship it into China, The Information reported late on 2 November US time, citing sources. Once the packages arrive, technicians remove the memory and discard the processors, according to accounts of the report by Seoul Economic Daily and TrendForce. A broker told the outlet that a single HBM3E unit now costs about $4,000, up from several hundred dollars a year ago.
The broker said one AI chip package needs six to eight memory units, so the memory alone adds $24,000 to $32,000. By our arithmetic, that equals up to 65% to 86% of the more than RMB 250,000 (about $37,000) that Huawei recently suggested for its Ascend 950DT, a price The Information also reported.
Business Pill · The Memory Wall
An evergreen Business Engineer explainer on why memory bandwidth, not chip speed, sets the pace of AI. It teaches the general idea only and does not discuss Huawei, smuggling or export controls.
The key insight: By our arithmetic, at the broker’s quoted $4,000 per HBM3E unit, the six to eight units one AI chip package needscost $24,000 to $32,000, equal to up to 65% to 86% of the more than RMB 250,000 (about $37,000) Huawei suggested for its Ascend 950DT.
The Information’s story, titled “How Smugglers Get Expensive Memory Chips to China from South Korea” and labelled exclusive, is behind a paywall; we read its free opening. It says the technique “has begun to catch on among smugglers involving a key type of memory chip.”
The opening explains why. “While U.S. export regulations ban the shipment of high-bandwidth memory chips on their own to China, including from countries like South Korea, those rules have a loophole that allows the sales of those memory chips if they’re packaged together on plates with cheap server processors whose performance falls below the export restriction threshold,” it says. “That’s because the computing power of the package gets determined by the server processor, not the memory chips, according to the rules.”
According to Seoul Economic Daily’s account of the report, the packages are shipped as a single product and classified as low-end server processors. The work in China fails often, the paper says, because memory stacks warp or crack if heat is applied too quickly or for too long, but smugglers accept the risk because the profit is large even after ruined units.
The $4,000 figure resolves a headline that circulated this morning on a news aggregator, “Smugglers exploit $4,000 HBM loophole”: it is the broker’s quoted price for one HBM3E unit, as relayed by Seoul Economic Daily.
Memory for one AI chip package at a broker’s quoted $4,000 per HBM3E unit (six to eight units, per The Information as relayed by Seoul Economic Daily) against the more than RMB 250,000 (about $37,000) Huawei suggested for its Ascend 950DT (The Information, via TrendForce). Shares are our arithmetic and upper bounds.
Kyle Chan posted this summary of The Information’s report on 3 November at 11:29 UTC. It says Huawei is telling some customers it does not have enough HBM and that they need to source their own memory, and describes the packaging route; this piece reads the US rule text behind the loophole the report describes.
Why Huawei Needs Korean Memory
The smuggled HBM flows mainly to Huawei’s customers, Seoul Economic Daily reports. People familiar with Huawei’s supply chain say Huawei recently told some customers that its next flagship AI chip, the Ascend 960DT, will contain substantial amounts of HBM from Samsung Electronics and SK hynix. Unable to buy memory directly from the Korean suppliers, Huawei has told some customers ordering its AI chips to source the memory themselves, the paper says.
Kyle Chan summarised the same point on X on 3 November: “Huawei is telling some customers it doesn’t have enough HBM and that they need to source their own memory.”
TrendForce, citing The Information, says the Ascend 960DT is expected in the first quarter of 2027 and is set to double the memory bandwidth and capacity of Huawei’s current AI processor, and that Huawei is working with DeepSeek to optimise the chip for pre-training. Fujian Jinhua Integrated Circuit, a US-sanctioned memory maker that has supplied Huawei for years, lacks the capacity to meet Huawei’s customer demand, according to the report.
Reuters reported, citing sources, that CXMT’s HBM3E yields remain at 20% to 30%, against about 70% at SK hynix, TrendForce says. Samsung and SK hynix together control more than 80% of the HBM market, according to Seoul Economic Daily.
What the US Rule Text Says
The controls date from an interim final rule the Commerce Department’s Bureau of Industry and Security published on 5 December 2024, with a compliance date of 31 December 2024 for the HBM changes. It added a new export control classification, ECCN 3A090.c, for HBM with a ‘memory bandwidth density’ above 2 GB per second per square millimetre. The rule says “All HBM stacks currently in production exceed this threshold.”
The rule’s technical note carves out packaged chips. It reads: “This control does not cover co-packaged integrated circuits with both HBM and logic integrated circuit where the dominant function of the co-packaged integrated circuit is processing.” The preamble adds that such chips “may be controlled by other ECCNs such as 3A090.a or 3A090.b depending on their Total Processing Performance (TPP) and performance density.”
That is the gap The Information describes: a package whose dominant function is processing falls outside the HBM control and is judged on its computing performance, which here is set by a low-end processor. The rule states its purpose plainly: “As indigenous PRC advanced computing ICs rely upon imported HBM, new ECCN 3A090.c implements restrictions to slow PRC attempts to indigenize advanced AI chip production.”
The Price of Scarcity
HBM shortages are also pushing up Huawei’s chip prices, TrendForce reports, citing The Information: Huawei told some customers it was raising the suggested price of the Ascend 950DT, currently its most powerful AI chip, to more than RMB 250,000 (about $37,000). TrendForce adds that the 950DT delivers roughly half the raw computing performance of Nvidia’s H100, launched in 2022.
Other Chinese chipmakers face the same pressure. Citing Reuters, TrendForce says Cambricon has raised indicative pricing for its next-generation 690 chip by 20% to 30%, and that MetaX and Iluvatar CoreX have reportedly made similar increases.
One smuggler who has been ordering memory chips from Korea told The Information that the routes the smuggler uses are now being used by other dealers, and that the smuggler plans to sell most of a stake in the company next year, expecting enforcement in Korea and the US to get tougher, according to Seoul Economic Daily.
Where SK Hynix and Samsung Stand
Seoul Economic Daily says the two Korean suppliers could come under US scrutiny if their memory is found inside Chinese-made AI chips. It recalls that Taiwan’s TSMC faced a US investigation in 2024 after one of its chips turned up in a Huawei product.
SK hynix said it strictly complies with all applicable laws and regulations, including US and Korean export controls, and has not supplied products to Huawei since Washington imposed export restrictions in 2020, the paper reports. The reports we read carry no comment from Samsung. None of them says either company sold memory to the smugglers.
As we read the rule text, the December 2024 control is written around the HBM stack, while a package whose dominant function is processing is judged on its computing performance; The Information says smugglers use that line by pairing the memory with a low-end processor.
As we read the reports, the price is the signal: a broker says one HBM3E unit went from several hundred dollars a year ago to about $4,000, and Huawei is asking some customers to find the memory themselves.
As we read the supplier side, SK hynix says it has not supplied Huawei since 2020, and none of the reports says either Korean company sold to the smugglers; Seoul Economic Daily says the risk for both is US scrutiny if their memory turns up in Chinese AI chips.
The Information, 2 November 2026
“That’s because the computing power of the package gets determined by the server processor, not the memory chips, according to the rules.”
Three Implications
FOR HUAWEI AND ITS CUSTOMERS Seoul Economic Daily says Huawei told some customers ordering its AI chips to source the memory themselves, and TrendForce says the 950DT’s suggested price rose to more than RMB 250,000.
FOR SK HYNIX AND SAMSUNG Seoul Economic Daily says they could come under US scrutiny if their memory is found inside Chinese-made AI chips; SK hynix says it complies with US and Korean export controls.
FOR US EXPORT POLICY The December 2024 rule excludes co-packaged chips whose dominant function is processing from the HBM control and points them to the performance-based ECCNs 3A090.a and 3A090.b.
The framework puts it this way: “Memory is becoming the new compute. HBM growth is outpacing accelerator demand, making memory the rate-limiting factor in the AI scaling equation.”
As we read it, the smuggling report shows the same constraint from the other side of the export line: the reported price of one HBM3E unit, about $4,000, is what a buyer outside the legal supply chain is said to pay. The framework does not discuss Huawei, smuggling or export controls.
What Is Not Established
We read only The Information’s free opening; the rest of its account comes to us through Seoul Economic Daily, whose English text is AI-translated from Korean, and TrendForce. The volume of HBM moved this way has not been reported, and the $4,000 price comes from one broker.
The comparison with the 950DT is ours. It sets the broker’s memory cost for a generic AI chip package against a reported suggested price for a different, current Huawei chip; the 950DT’s own memory content and cost are not given. Because the price is ‘more than’ RMB 250,000, the 65% and 86% shares are upper bounds. We did not contact Huawei, SK hynix, Samsung or the Bureau of Industry and Security.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
The Information reports that smugglers ship Korean HBM into China co-packaged with low-end processors, using the December 2024 rule’s carve-out for chips whose dominant function is processing. A broker prices one HBM3E unit at about $4,000; at six to eight units per AI chip, that memory equals up to 65% to 86% of the more than $37,000 Huawei suggests for its Ascend 950DT, by our arithmetic.
A note on sourcing. This piece is based on the free opening of The Information’s report of 2 November 2026 (US time), the accounts of that report by Seoul Economic Daily (English edition, AI-translated from Korean) and TrendForce on 3 November, a post by Kyle Chan on X, and the Bureau of Industry and Security’s interim final rule of 5 December 2024 in the Federal Register. The 65% to 86% comparison is our arithmetic. We did not contact Huawei, SK hynix, Samsung or BIS.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
