CXMT Took 10% of DRAM and Shipped HBM3E a Year Early
China's memory champion went from 4% to 10% of global DRAM revenue in a year, started LPDDR6 mass production before Samsung or Micron, and put HBM3E samples in Alibaba's and Cambricon's hands twelve months ahead of every forecast.
CXMT — ChangXin Memory Technologies — took 10% of global DRAM revenue in Q2 2026, up from 4% a year earlier and 8% in Q1. The three-company oligopoly of Samsung, SK hynix, and Micron fell below 90% of the market roughly two years ahead of when analysts expected it to.
That is the headline. Two other dates matter more.
LPDDR6 first, not third
On August 29, CXMT announced mass production of LPDDR6, with Xiaomi's 18 Fold — launching this month — as the first commercial device.
This is the part that should unsettle people. The comfortable Western framing of Chinese memory has been "years behind on process, competitive on legacy nodes, dumping on price." Being first to volume on a new mobile memory standard, ahead of Samsung and Micron, is not a legacy-node story. It is a cadence story: CXMT hit a standards transition on the leading edge of the industry rather than following it.
Mobile memory is also the right place to do it. LPDDR is high-volume, design-win-driven, and dominated by a domestic customer base — Xiaomi, and by extension the entire Chinese handset industry — that has both commercial and political reasons to qualify a local supplier. Winning the first LPDDR6 socket in a flagship foldable buys CXMT the thing capital cannot: production learning at scale.
HBM3E, twelve months early
On September 1, CXMT had delivered small-batch HBM3E to Alibaba's T-Head and to Cambricon for qualification testing. Risk production has started; large-scale mass production is targeted for as early as 2027.
Every major forecast had China's first HBM3E production landing in 2027. Samples in customers' hands in September 2026 is roughly a twelve-month compression of that timeline.
HBM is the chokepoint that matters. AI accelerators are memory-bandwidth-bound, not compute-bound, in most real workloads — which is why HBM supply, not logic capacity, has been the binding constraint on Chinese AI hardware. Export controls restricted advanced HBM to China precisely because it is the piece that makes a domestic accelerator competitive rather than merely functional.
The customers here are the tell. T-Head builds Alibaba's inference silicon. Cambricon builds the closest thing China has to a general-purpose AI accelerator line. Neither is buying HBM3E for phones. This is the memory half of a domestic AI stack going into qualification.
Why the door was open
There is a supply-side explanation for CXMT's share gain that has nothing to do with brilliance, and it deserves to be stated: Samsung pivoted capacity toward HBM, and conventional DRAM supply tightened as a result. When the incumbents chase the highest-margin product, they vacate the floor of the market — and the floor is where a challenger builds volume, yield experience, and customer relationships.
That is the classic disruption geometry, and it is not an accident of 2026. It is what happens every time an industry's margin structure gets steeply tilted by one product line.
The consequence is now visible in the incumbents' behaviour. The read across the Korean industry is that Samsung and SK hynix have to move faster to HBM4 — accelerate up-market rather than defend the base. That is a reasonable response and also the exact move that opens more floor.
What this does not mean
CXMT at 10% of DRAM revenue does not mean parity. Revenue share at the commodity end reflects volume more than value; the incumbents still hold the high-margin HBM business, the advanced process nodes, and the qualification relationships with Nvidia and AMD that matter most. HBM3E samples are not HBM3E revenue, which is still a quarter or two away, and mass production in 2027 is a plan, not a shipment.
It also does not mean export controls failed. Controls were never going to stop China from building memory; they were designed to delay it and raise its cost. A twelve-month compression against forecast is a smaller victory than "we caught up" and a larger one than the controls' architects would like.
The number to watch
Not market share. Yield on HBM3E at volume.
HBM is a stacking and packaging problem as much as a lithography problem — through-silicon vias, thermal management, and known-good-die screening across eight or twelve layers. Yields on early HBM generations were brutal even for SK hynix, which had a decade of head start. CXMT can sample HBM3E and still take two years to make it economically at scale.
If CXMT posts credible HBM3E yields in 2027, the calculus for every Chinese accelerator program changes: domestic compute stops being bandwidth-starved, and the practical ceiling on Chinese AI training moves up sharply. If yields stall, CXMT remains an excellent commodity DRAM company that took share from distracted incumbents.
Both outcomes are live. What is no longer live is the assumption that China's memory industry is a slow follower. It hit LPDDR6 first and pulled HBM3E in by a year, in the same quarter it broke the oligopoly's 90% floor.