China's HBM Breakthrough Changes the Game for Korea
CXMT has entered trial production of HBM3E memory chips, narrowing the gap with Samsung and SK Hynix just as US export controls tighten. The move raises stakes for allies racing to close the door.
China Just Built a Second Source for HBM. The Clock Is Ticking.
CXMT, China’s most ambitious memory chipmaker, has entered risk production of HBM3E — the fifth generation of high-bandwidth memory that powers the world’s fastest AI accelerators. The company hasn’t announced any customer orders yet. It doesn’t need to. Risk production is the semiconductor industry’s way of proving you can build something before you sell it. CXMT is telling the world it can build HBM3E.
This matters more than the press release suggests. For years, Samsung Electronics and SK Hynix have treated HBM as an exclusive club. The two Korean giants supply virtually every major AI chipmaker on Earth — NVIDIA, AMD, Intel, and the rising Chinese players like Huawei and Cambricon. CXMT’s move is the first credible signal that a second source is materializing, and it’s happening at the exact moment Washington is trying to wall off China from the very technology this chip enables.
The Gap Is Real, But It’s Closing Fast
CXMT hasn’t disclosed the performance specs of its HBM3E, so we’re working from JEDEC standards. The HBM3E specification calls for a 1,024-bit interface width, data rates between 9.2 and 12.4 gigabits per second per pin, and a maximum bandwidth of 1.2 terabytes per second. At 8 stacked layers, that’s 24 gigabytes. At 12 layers, 36 gigabytes.
Samsung and SK Hynix are already shipping samples of HBM4E — two generations ahead. HBM4E doubles the interface to 2,048 bits, pushes per-pin speeds to 14–16 Gbps, and delivers up to 4.1 TB/s of bandwidth. Capacity tops out at 64 gigabytes with 16 stacked layers, and power efficiency improves by more than 40 percent.
Two generations is a real gap. But it’s not the unbridgeable chasm that Korean executives may have assumed a few years ago. TechPowerUp noted that CXMT has a pattern of moving from trial production to mass output remarkably quickly. The company completed a cleanroom build in 12 months — half the industry norm of two years. That pace of execution is not typical of Chinese semiconductor firms that have historically struggled with follow-through.
The Capacity Bet
CXMT operates two 12-inch fabs in Hefei and Beijing, each producing 100,000 wafers per month for a combined 200,000. By year-end, it plans to reach 350,000 wafers monthly. Once new facilities in Shanghai and Hefei come online, total capacity could triple to 600,000 wafers per month.
That’s not yet in Samsung’s league — Samsung alone produces millions of equivalent 12-inch wafers annually — but it’s enough to supply a meaningful share of China’s domestic AI demand. And domestic demand is where this story really plays out. Chinese AI chipmakers are being cut off from Western HBM supplies through export controls. CXMT isn’t trying to win in Silicon Valley. It’s trying to become the fallback supplier for a market that can’t buy anywhere else.
The Pressure on Korea’s Customers
The immediate consequence of CXMT’s entry isn’t competitive pricing — it’s optionality. Right now, every Chinese AI chip designer that needs HBM faces a binary choice: buy from Samsung or SK Hynix and navigate US licensing requirements, or wait for a Chinese alternative that may or may not meet performance targets. CXMT removing the wait option changes the calculus. Even if its chips are slower or less reliable than HBM4E, they’re available today to buyers who have no other legal path to HBM at all.
For Samsung and SK Hynix, the threat isn’t that CXMT will undercut them on price in the global market. The threat is that CXMT will lock in the one segment neither Korean company can easily lose customers in the long run — Chinese AI demand. Every customer CXMT captures there is a customer that won’t come back when export controls eventually ease, because they’ve already qualified a second source.
The D램 Side Effect
There’s a secondary ripple that most observers are missing. CXMT is simultaneously expanding its general-purpose DRAM output to offset the wafer capacity it’s diverting toward HBM. Industry sources note that this is creating expectations of price stability in the broader DRAM market. But if CXMT pivots more production lines from standard DRAM to HBM, that supply cushion evaporates — and general-purpose DRAM prices could climb.
This isn’t just a CXMT problem. It’s a supply-chain signal. The company that converts the most wafer capacity to HBM gains strategic leverage, and CXMT is converting faster than anyone expected.
What Happens Next
The most likely scenario isn’t that CXMT overtakes Samsung or SK Hynix in HBM performance. It’s that CXMT becomes good enough for the Chinese market, fast enough to matter before US controls harden further. Korean memory companies now face a reality they’ve never had to contend with: a state-backed competitor with domestic market immunity, building capacity at twice the normal speed, and targeting their most profitable growth segment.
The question for Washington isn’t whether CXMT can make HBM3E — it clearly can. The question is whether export controls will be tightened further before CXMT’s risk production becomes volume production, and whether Samsung and SK Hynix can maintain a performance lead large enough that the Chinese alternative remains unattractive even to buyers who don’t need a US license.