The Sanctions Trap: How CXMT Turned US Export Controls Useless

The Chinese industry cannot keep up with Western companies when it comes to manufacturing the most advanced dynamic random access memories (DRAMs). This analysis rests on the assumption that the semiconductor restrictions imposed by the United States limit China’s ability to obtain the latest generation of fabrication nodes – 7 nm, 5 nm, and 3 nm. Such an assumption becomes less and less relevant as time goes by.

Illustration by The Geostrata

This is the logic that has dominated the policy debate for the past five years. However, it is increasingly outdated. Changxin Memory Technologies (CXMT), which is currently the only producer of domestic DRAM in China, does not try to overcome the equipment restrictions. Rather, company strategy is reflected in supply chain reports and announcements that show CXMT's strategic pivot toward HBM.

CXMT currently lags just 3-4 years behind global players in HBM, planning to manufacture HBM3 by 2026 and HBM3E by 2027 with significant technological advancements in DRAM.

The latter technology is not yet covered by restrictive measures, while China needs it for the development of its AI infrastructure.

ACKNOWLEDGING THE PLATEAU

The sanctions success story is built on a lie that equipment control equals outcome control. In 2019, when CXMT found itself under the early pressures of sanctions, the company adjusted its design, cutting back dependence on the slight tactical move. The United States prevented Advanced Semiconductor Materials Lithography (ASML), a 5 nm company, from selling its Extreme Ultraviolet (EUV) lithography machines, which are important for the creation of sub-10 nm semiconductors. It is a restriction, but very real.

CXMT's current production plateau stands at about 240,000 wafers per month for these markets without breaking any American laws concerning exports to China, with all manufacturers using 16nm-19nm nodes and older lithography technology.

CXMT cannot advance past this plateau without EUV – on paper, this seems conclusive; in reality, it misses the crucial misjudgement by the US – DRAM as the battleground, while CXMT had already moved on from the battle.

The production of traditional DRAM in constrained nodes is still profitable in markets outside of the United States, such as Europe, Southeast Asia, and India, which all need commodity memory chips. CXMT’s talent, capital, and strategic focus have been redirected elsewhere.

IP THEFT AS COMPRESSED R&D

The acquisition of intellectual property via espionage. In 2023, ten ex-Samsung employees were taken into custody on charges of stealing intellectual property that had taken five years to develop and which cost the company more than $1 billion. At the same time, SK Hynix engineers were detained for trying to expose their HBM process technology. It was the compressed research and development pipeline of CXMT. With the help of stolen Korean technology – nm, 5 nm, architectural blueprints, process guidelines, and yield information – CXMT gained five to ten years of research work.

The sanctions have no power to restrict information after it has crossed the border and made its way inside the company. Controls on equipment are related to machines, but theft is related to brains. The imbalance is perfect. As the United States considered increasing export controls, CXMT was busy reverse engineering the Korean HBM architecture, compressing a period of a decade into months.

HBM PIVOT 

When there was no longer any possibility of expanding DRAM through conventional means, CXMT had already started working on developing its HBM production capacity based on the stolen specs. This product is completely different from DRAM: more bandwidth, more expensive, higher margins, and most importantly, under less developed export controls.

The United States has been controlling EUV lithography for DRAM exports with surgical precision. There are HBM export controls, yet they lag behind technological development by 2 to 3 years.

Approximately 60,000 wafers have been set aside per month for HBM3 production, with this production expected in Q4 of 2026 and reaching HBM3E-level capabilities by 2027. While yields currently stand at 50%, the yield curve is quite steep. The reason for such urgency lies in Huawei. Unable to import Korean HBM because of the sanctions imposed by the US government, Huawei has been pushing CXMT to provide faster HBM3 deliveries. CXMT has already started delivering samples of HBM3 since mid-2025.

DISENTANGLEMENT FROM KOREAN DEPENDENCE

About the AI chips used by Huawei in its efforts to build China’s AI infrastructure, Korean-made HBM has now become a chokepoint. The HBM produced by CXMT constitutes direct substitution. By 2026, Huawei will no longer be fully dependent on Korean companies. And by 2027, China's AI infrastructure deployment will no longer require Korean industries at all. Currently, Korea makes about 30,000 HBM wafers each month through Samsung and SK Hynix. And CXMT’s capacity of 60,000 wafers constitutes 20% of Korea’s current capacity. This is not domination but substitution. As it stands, access to Korean HBM for sanctioned Chinese AI processors is now restricted.

AMBIGUITY IN REGULATION

The ambiguity of designation. CXMT received an entry on the US Department of Defence list of companies (section 126H List of Communist Chinese Military Companies) at the beginning of 2025, but CXMT was delisted in January 2026. However, delisting is not an indication of any easing of the technology restrictions. The export control of equipment continues. All that is achieved by de-listing is an insurance policy. CXMT looks like a “normalised” company, fit for doing business, no longer designated as a defence firm. At the same time, the relevant technology restrictions (on equipment, not on designation) continue to apply.

LEGITIMACY LAUNDERING VIA SHANGHAI IPO

On June 25th, 2026, CXMT launched the biggest tech IPO in China that year, for $26 billion. There is no mention of any controlling shareholder in the prospectus, thus hiding the state funding behind the corporate blanket. That’s legitimacy laundering. The transformation into a company funded by stock markets allows CXMT access to the commercial capital markets, allows foreign investment, and provides political cover through the "civilian enterprise”. The funds of $7.5 billion for upgrading production lines, $13 billion for DRAM technology, and $9 billion for R&D are directed to both DRAM upkeep and HBM development, but the stock market is about commercial competitiveness, not technological substitution.

SUBSTITUTION ARCHITECTURE

CXMT does not avoid the sanctions using tricks. Instead, it is executing an intelligent substitution strategy that makes some of the restrictions irrelevant. The equipment controls will not stop what is no longer the focus of attention. The sanctions are directed against DRAM capacity expansion, but CXMT does not engage in it anymore. The regulations relate to the use of advanced nodes, while CXMT operates profitably in legacy nodes and turns to HBM, which is under immature controls.

QUESTION FOR 2027

The outcome can be measured by 2027. The question changes from “Can we build advanced DRAM?” to “Is Korea still able to control our AI infrastructure?” 

The answer, by 2027, will be no. This is not an expectation that CXMT will become the dominant player in the world market of HBM. Korean manufacturers will continue to be more advanced.

But what this shows is the loss of effectiveness of sanctions at the point of their strategic need. Sanctions are based on the assumption of forcing through the lack of scarcity. CXMT demonstrates that, through capital, intellectual property, and strategic pivoting, the unpleasant realisation for semiconductor policy is that the sanctioning of equipment works only if the sanctioned entity is unable to change the target of production.

---

BY HEMA

TEAM GEOSTRATA

info@thegeostrata.com