China Just Broke Into HBM. The 'Big Three' Can't Ignore It Anymore.

The quiet rupture

China’s largest DRAM maker, CXMT, is now in small-batch production of HBM3E — the fifth generation of high-bandwidth memory, a component that has been one of the most tightly controlled chokepoints in the global semiconductor supply chain. And it is happening under US export restrictions that were explicitly designed to prevent exactly this.

That is not a trivial detail. For years, Washington’s strategy rested on a simple assumption: if China cannot buy advanced memory chips, it cannot produce them fast enough to matter. CXMT’s move suggests that assumption is fraying.

The story got its first international headlines on September 31 when the Information Times, citing two sources, reported that CXMT had begun limited HBM3E output. By early October, Chinese financial media confirmed the company was testing the chips with domestic AI chip designers — Alibaba’s T-Head and Cambricon Technologies — who are evaluating compatibility with their own processors. At least one product featuring CXMT’s HBM3E could ship as early as next year.

Not the first time, but the fastest

Let’s get the context straight, because the numbers deserve scrutiny.

SK Hynix began HBM3E mass production in 2024. Samsung and Micron are now shipping HBM4, with HBM4E samples already in flight. CXMT is roughly one generation behind the “Big Three” — a gap the Korean and American firms estimate at three to five years in raw technical capability.

But speed matters as much as position. Hardware Breakers, a respected hardware analysis outlet, called the pace “unprecedented” — a Chinese DRAM maker reaching the same generational stage as Samsung and SK Hynix in roughly the same timeframe. That framing deserves attention.

CXMT is not racing alone. The same week it announced HBM3E progress, it also confirmed it had begun mass-producing LPDDR6, the latest low-power mobile memory standard. Xiaomi is already shipping the chip in its upcoming Xiaomi 18 Fold foldable phone, paired with the company’s custom 3-nanometer “Xuanjie O3” AI processor. China’s Securities Times called it the world’s first commercial LPDDR6 deployment. Whether that claim holds outside China remains unclear — Samsung and SK Hynix have parallel LPDDR6 programs — but the coordination between chip designer and memory producer is unmistakable.

The yield problem no one is ignoring

Here is where optimism needs a reality check.

CXMT is reportedly struggling with low yields — the percentage of usable chips coming off a production line. Yield rates determine whether a memory product is commercially viable or just a lab curiosity. CXMT has not disclosed its yield numbers, and neither has it explained how it is sourcing the advanced lithography and packaging equipment that HBM production demands, given that the US has blocked access to the most capable tools from ASML and others.

UBS Group acknowledged the momentum but noted in a research note that Chinese DRAM technology remains two to three generations behind global leaders, and that CXMT’s production costs are structurally higher. In other words: the chips may work, but they may not yet be competitive at scale.

This is the honest tension in the CXMT story. Beijing has achieved a breakthrough under extreme constraints. It has not yet solved the harder problems of yield, volume, and cost.

Who wins, who loses

The immediate beneficiary is China’s AI sector. US export controls have created a bottleneck for Chinese cloud providers and AI chip designers who need HBM to train large models. Every month CXMT delivers viable HBM3E, even in small volumes, reduces that dependency — however slightly.

Alibaba and Cambricon are the canaries in this mine. If their processors paired with CXMT memory prove stable at scale, they create a domestic supply chain that operates outside US jurisdiction. That is the scenario Washington’s export controls were meant to prevent.

For the Big Three — Samsung, SK Hynix, Micron — CXMT is not an existential threat today. But it is a signal. Memory markets are commodities at their core; once a competitor proves it can produce at all, the question shifts from “can they?” to “at what price?” CXMT’s stated 2028 revenue target of 129 trillion won — roughly $93 billion — comes from Guoshi Securities. That is a ambitious number, but it reflects the belief that domestic demand in China alone could absorb a significant share of global memory output within a few years.

The loss side is harder to quantify. Chinese buyers of Samsung and SK Hynix HBM will face either higher prices (as suppliers reroute supply away from China) or geopolitical friction (as US pressure intensifies). AI hardware costs could rise in the near term if the Big Three tighten allocation. Over time, CXMT’s entry could pull prices down — the classic commodity market squeeze.

What happens next

The next twelve months will be decisive. CXMT has committed to expanding HBM3E production next year. If yield improves and domestic AI chip designers lock in volume supply agreements, the chip will move from test batches to real products — potentially in data centers and consumer devices by late 2026 or 2027.

If yields stall, the breakthrough remains a PR win without commercial teeth. That is the realistic range of outcomes.

For the Big Three, the strategic implication is clear: they can no longer assume China will remain a captive customer dependent on their supply. CXMT is building a parallel track. Samsung and SK Hynix will need to decide whether to compete on price in the lower tiers, double down on HBM4 and beyond, or accept that China’s memory market is slowly closing off to them — a familiar pattern in Beijing’s industrial policy playbook.

Washington faces an analogous choice. Tightening restrictions further risks driving China’s domestic industry faster toward self-sufficiency — the very outcome sanctions were meant to prevent. Leaving them open acknowledges that the technology gap, while real, is narrowing.

The era of assuming China could not produce advanced memory under sanctions is ending. CXMT’s HBM3E may not yet match SK Hynix in yield or volume. But it exists. And in semiconductor politics, existence is the first step toward dominance.