This month, mainstream media have been warning consumers that electronic devices may get pricier because of rising demand for dynamic random access memory (DRAM), a key component. The surge in DRAM costs, estimated to have risen by 50% during the final quarter of 2025, can largely be traced back to a specific cause: the AI industry’s appetite for high-bandwidth memory (HBM). This demand has led memory-makers to shift production away from standard DRAM chips and toward HBM.
While its contribution is often overshadowed by those of processors it supports, HBM plays a vital role in training and running advanced AI systems, so much so that it now accounts for half the production cost of an AI chip. Companies’ determination to secure this lesser-known component proves its value. In December 2024, the US announced new export restrictions on the sale of HBM chips to China. In the month before the restrictions came into effect, Huawei and other Chinese companies reportedly stockpiled 7 million Samsung HBM chips, a haul likely worth over $1 billion.
The December 2024 controls specifically targeted HBM in order to slow China’s domestic AI chip-production efforts. Targeting HBM in this way is possible because it is manufactured separately from GPUs and then fused with them in a subsequent packaging step. Yet, while these controls are having an impact, they contain significant gaps. In the year since, Chinese companies have continued to acquire HBM directly via loopholes, and they have also purchased tooling needed to develop HBM domestically. If US policymakers are serious about slowing China's AI chip production, they should close the gaps in HBM controls.
When people discuss AI chips, they usually talk about processing speed, measured in the number of floating point operations per second (FLOP/s). Processing speed is an important metric, as it tells us how quickly an AI chip can perform calculations when it has data to work with. But in practice, AI chips often sit idle, waiting for data to be fed to them.
Just as a factory’s output depends not only on how fast workers can assemble parts but also on how quickly parts reach the assembly line, so does memory bandwidth constrain overall AI chip performance. If AI chips are a factory, HBM is both the stockroom and the conveyor belt, storing and delivering parts to the workers fast enough to keep the assembly line moving.
HBM improves AI chip performance. In 1994, William Wulf and Sally McKee coined the term “memory wall” to describe problems arising from improvements in processor speeds far outpacing improvements in memory bandwidth. This observation proved prescient and drove efforts to increase memory bandwidth. In the mid-2000s, AMD began working on memory innovations that eventually led to HBM.
Traditional DRAM sits on separate chips connected to processors via the motherboard through relatively narrow channels. AMD’s key innovation was a series of procedures for stacking multiple DRAM dies on top of one another and using tiny, vertical electrical connections called through-silicon vias (TSVs). This configuration allows the layers to communicate. (A “die” is a piece of silicon circuitry that gets packaged alone or with other dies to create a chip.) These stacks are incorporated directly onto the same silicon wafer (an “interposer”) as the processor (for example, a GPU). At the bottom of the HBM stack sits a logic die that interfaces between the memory stack and the processor.
Because of the HBM stack’s position immediately beside the processor, the two can communicate across many more parallel channels than traditional memory allows. The result is dramatically higher data throughput while drawing less power.
HBM impacts AI progress. In December 2013, SK Hynix, which had worked with AMD to develop HBM, produced the world's first HBM chip. In 2015, the company began high-volume production at its facility in Icheon, South Korea.
The first product to use HBM was the AMD Fiji GPU, in 2015: a consumer chip mainly used for video gaming. But HBM soon started appearing in data center GPUs and other AI chips, first with Nvidia's Tesla P100, announced in April 2016. As AI models grew in size and complexity through the late 2010s and into the 2020s, HBM became effectively mandatory for high-performance AI accelerators. And as time went on, each successive generation of HBM provided AI chips with greater memory bandwidth and more memory capacity. In 2024, Nvidia announced its Blackwell architecture, which coupled each data center GPU with multiple HBM stacks.
Today, every major AI chip maker — including Huawei — uses HBM in its products. Those stuck with an older generation of HBM produce an inferior product, no matter how fast their processors are.
HBM production is highly lucrative and highly concentrated. Three companies control 97% of HBM wafer production today: SK Hynix and Samsung, of South Korea, and Micron, of Boise, Idaho. Because HBM is so vital to the AI industry, these companies now enjoy extraordinary market valuations. SK Hynix, the leading HBM producer, has seen its stock rise sevenfold since November 2022, when ChatGPT launched.
This concentration in HBM production reflects decades of global competition that left South Korean and US companies in command; Chinese players like ChangXin Memory Technologies (CXMT) are still racing to catch up.
China is a latecomer to HBM manufacturing. Just as cellular network technology has gone from 4G to 5G and so on, HBM technology progresses in generations. CXMT, China’s leading DRAM manufacturer, is currently manufacturing second-generation HBM chips — comparable to those first produced by South Korean companies in 2016. However, it is now reportedly skipping an interstitial generation to develop third-generation HBM — technology that is approximately three to four years behind SK Hynix, Samsung, and Micron (see table below).
CXMT is attempting to close this gap by using large quantities of imported equipment, most notably stockpiled immersion deep ultraviolet (DUV) photolithography machines from the Dutch firm ASML. CXMT’s production also relies on equipment made by US firms like Applied Materials and Lam Research, as well as Japanese companies like Tokyo Electron.
/inline-pitch-cta
Despite CXMT’s manufacturing advances, Huawei primarily uses more advanced HBM stockpiled from Samsung and, to a lesser extent, from SK Hynix. At its Connect conference in September 2025, Huawei announced plans for “proprietary” HBM, possibly to be fabricated primarily by CXMT or by Huawei itself with collaborators.
China is ramping up domestic HBM production. It remains to be seen how CXMT and other Chinese manufacturers may ramp up production over the coming years. Our best guess is that they will be able to produce approximately 7 million (primarily HBM3) dies in 2026. That is sufficient for about 600,000 AI chips of comparable performance to Nvidia’s H100 accelerator. This estimate assumes that each AI chip uses eight HBM stacks (as the Huawei Ascend 910C accelerator does) and a 70% yield (i.e., 7 in 10 packaged AI chips prove functional). However, these estimates are highly uncertain.
Even if these targets are met, China’s domestic production will likely fall far short of demand, leaving its AI ambitions fragile and highly susceptible to effective Western restrictions.
Because HBM is a critical enabling component for AI chips, and because AI capabilities have direct national security implications, in December 2024, the Bureau of Industry and Security (BIS), under the US Department of Commerce, rolled out export controls on HBM and related tooling. Its goal was to cut off China’s ability to both acquire and produce HBM more advanced than what the country can currently manufacture.
However, the rules and their application contain important gaps, detailed below.
Coverage for key equipment is incomplete. The December 2024 controls included a mechanism called Foreign Direct Product Rules (FDPRs) to target certain semiconductor manufacturing equipment (SME) — but with a key upgrade. Usually, FDPRs give BIS jurisdiction over foreign-produced items manufactured using US software or technology (e.g., designs, specifications, and know-how). The 2024 FDPR, however, gives BIS jurisdiction over any SME that “contains” a chip which was itself manufactured using US technology. These rules effectively capture the entire non-Chinese supply chain.
Importantly, though, the 2024 FDPR contains exemptions for firms headquartered in Japan, the Netherlands, and other nations with export-control regimes similar to those of the US. While the new rules reinforced restrictions of the most advanced immersion DUV lithography machines — crucial for much of HBM’s development to date — the informal agreements they rely on still allow Japanese and Dutch companies to sell older immersion DUV machines to China. Though these machines are less advanced, Chinese companies can use a technique called multi-patterning to make more advanced dies than they could otherwise, at the cost of higher defect rates and slower production.
The 2024 controls also contain gaps that allow China to scale up its semiconductor tooling industry. For example, the controls do not target equipment needed for hybrid bonding, a technique for reducing the height of the HBM stack to comply with industry standards, increase performance, and lower power consumption. Hybrid bonding will likely be critical for developing future generations of HBM.
An implementation gap allowed stockpiling. As previously described, the US rules’ one-month gap between announcement and implementation gave Chinese firms ample time to stockpile inventory. By then, the US government had been signaling that HBM controls were imminent for nine months. Chinese firms utilized this nine-month window aggressively. Huawei and Baidu stockpiled 6 million HBM stacks from Samsung — not counting the 7 million in the month between the rules’ announcement and their implementation. These stockpiles are enough to allow Huawei to manufacture about 1.6 million Ascend 910C chips (with memory bandwidth performance comparable to the Nvidia H100, introduced in 2022).
Stockpiling became rampant for chip-manufacturing equipment as well: according to one analysis, CXMT has likely acquired enough SME for HBM production through 2026 or 2027, after which it will encounter obstacles both in ramping up production volumes for current HBM generations and in developing more advanced HBM generations.
Equipment manufacturers received carve-outs. When introducing the 2024 rules, BIS also announced that it was adding 140 companies to the Entity List, which names foreign companies, organizations, and individuals that the US believes pose national security or foreign policy risks. US companies cannot do business with these entities without special licenses, which are difficult to obtain.
However, BIS diluted the final rules in response to diplomatic and economic pressures. For example, it placed CXMT on a narrower list that bars only the US Department of War from doing business with it, thanks to extensive Japanese lobbying on behalf of an equipment manufacturing company, Tokyo Electron, that relies on the Chinese market for 45% of its revenue. Such pressure from Japan illustrates how loopholes in the 2024 FPDR, described above, allow allied nations autonomy that may pose conflicts of interest when paired with US national security and chip export-control concerns.
To make matters even more complicated, Japanese regulation lacks sweeping abilities (which the US claims) to impose restrictions on specific companies that receive exports. As long as Japan can largely make its own rules under the 2024 FDPR, impeding CXMT’s progress will remain a challenge.
/odw-inline-subscribe-cta
Current HBM controls incentivize extraction and smuggling. The December 2024 controls targeted the sale of raw HBM stacks, but they did not restrict HBM that had already been integrated into AI chips. In late 2024 and early 2025, several Taiwanese companies reportedly exploited this loophole by loosely attaching Samsung HBM to very simple processor chips. These chips were designed to appear compliant with export restrictions while allowing the HBM to be easily extracted once in China.
SemiAnalysis judges that this breach has now been contained, based on the offending firms’ revenues returning to normal levels. But the structural incentive to smuggle Western HBM stacks is high, because of their far superior performance and greater supply. Given that AI chip smuggling appears rife, it seems likely that bad actors will engage (or are already engaging) in similar diversions of HBM. Meanwhile, BIS’s enforcement, which remains heavily resource-constrained, cannot counter the strong incentives for these activities.
To make the controls more effective, policymakers can target each of the loopholes mentioned above.
Expand and modernize controls on key equipment. The US should pressure the Netherlands to ban the servicing of all existing ASML lithography tools in China, thus degrading their long-term viability. Depending on diplomatic momentum, the US could also work with the Dutch government to ban the sales of remaining immersion lithography tools to China, further narrowing the pipeline of chip-making tools available for HBM production.
Looking forward, controls should also target the next technical frontier: hybrid bonding. As the industry approaches the next few generations of HBM, the US should consider putting trade restrictions on companies that manufacture hybrid bonding systems, as well as overlay metrology (for measuring the alignment of one die layer to another) and plasma dicing (an etching process to perfectly separate dies from wafers), among other tools over which it has meaningful leverage.
Close the implementation gap. In addition, future controls should be issued with minimal advance notice to prevent the massive stockpiling China achieved ahead of the 2024 controls. While there is a trade-off between speed and allies’ buy-in, speed is essential for an effective export-control regime and should be prioritized over airtight allies’ approval. As evidence emerges in support of new controls, the US will need to act as quickly and with as little signaling as possible, while substantially closing the gap between action and diplomatic approval.
Stand firm on chip-maker sanctions. In its next round of controls, BIS should explicitly target the companies leading China’s domestic HBM efforts. CXMT, as the prime mover among those companies, should be added to the BIS Entity List so it cannot as easily acquire needed tools from foreign companies to advance HBM in quality and quantity.
Mindful of the diplomatic sensitivities of this decision, the US must convince Japan to either adapt its domestic export laws to mirror US restrictions or find alternative customers for Tokyo Electron. As a last resort, BIS could unilaterally revoke Japan’s exemptions from the relevant FDPRs. BIS should also target the broader ecosystem supporting Huawei by adding companies involved in chip-packaging to the Entity List and applying FDPRs to those making semiconductor equipment.
End component extraction and smuggling. BIS should issue “red flag” warnings to chipmakers, outsourced semiconductor assembly and test companies, and third-party vendors to watch for orders of advanced HBM memory packaged in ways that allow the memory to be easily removed from finished chips after export. It should also investigate whether Taiwanese or South Korean companies knowingly violated the controls by supplying Chinese companies with HBM stacks via lightly assembled products. Punishing the offending companies will send a strong signal to other would-be smugglers that semiconductor-related transactions with China are not worth the trouble.
Furthermore, Congress should authorize a whistleblower incentive program for individuals who expose export-violation schemes, such as HBM smuggling. This program should offer large, penalty-financed rewards, alongside strong protections such as anonymous reporting channels. Such a program has been proposed in the bipartisan and bicameral Stop Stealing Our Chips Act, first introduced in April 2025 by US Senators Mike Rounds (R-SD) and Mark Warner (D-VA).
Finally, to improve BIS’s enforcement capacity, Congress should grant the President’s budget request of $303 million for the agency, a 60% increase over its current budget.
Export controls must track China’s real capabilities. Ultimately, the US must calibrate its controls to the best chips and tools China can realistically make at large quantities domestically, not what America can achieve at the frontier. Otherwise, the controls risk allowing exports that ease China’s key domestic bottlenecks. By conducting regular audits of China’s capabilities, the BIS’s internal evaluators can ensure that export controls remain flexible enough to keep US firms relevant in the global market, while strict enough to prevent China from acquiring HBM capabilities that far exceed its domestic production potential.
Tooling controls must be less vulnerable to diplomacy. The efficacy of future HBM controls is complicated by the current diplomatic climate. In November 2025, for example, BIS suspended its “Affiliates Rule,” which would have extended Entity List restrictions to majority-owned subsidiaries of listed firms. This suspension, which lasts one year as part of broader US-China trade talks, will enable listed firms to continue accessing controlled technologies through overseas subsidiaries and affiliates. The suspension came after reports, in May 2025, that BIS was considering adding CXMT to the Entity List, a move that now seems to have been shelved for similar diplomatic reasons.
The most recent example of the political complications of new chip restrictions is the Trump administration’s decision to allow China to buy Nvidia H200 accelerators and comparable chips, signalling a more dovish approach to controls on Chinese AI chip-making. The H200 offers considerably higher memory bandwidth than current Chinese alternatives, allowing Chinese companies to enjoy better chip performance without having to import HBM stacks for domestic AI chips.
However, the H200 decision was motivated in part by a desire to counter Huawei’s AI chip efforts. The strategic rationale is that Chinese companies with access to US chips won’t want to buy inferior Chinese competitors. Strong HBM controls are essential for buttressing that strategy. Despite the aforementioned issues with the HBM controls, the technical gap remains a significant barrier for China’s AI industry. While Huawei and CXMT are aggressively pursuing domestic HBM development, limited domestic Chinese production, for now, will curtail the amount of advanced AI chips Huawei can produce, and thus limit China’s ability to surpass the newly available H200s.
Every loophole grants China another opportunity to close the technological gap. This gap may shrink faster than anticipated unless the US fully commits to tightening its regime. It may still be possible to contain China’s chip-making ambitions, but, to do so, the US must first close the exits.
See things differently? AI Frontiers welcomes expert insights, thoughtful critiques, and fresh perspectives. Send us your pitch.
Traditional insurance can’t handle the extreme risks of frontier AI. Catastrophe bonds can cover the gap and compel labs to adopt tougher safety standards.
The White House is betting that hardware sales will buy software loyalty — a strategy borrowed from 5G that misunderstands how AI actually works.
