13 SEP 2026 — CXMT took 10 per cent of the DRAM market in the second quarter, up from 4 per cent a year earlier, and the incumbents' share of the market fell below 90 per cent roughly two years ahead of forecast. It is genuine relief for memory buyers, and it is not relief for the shortage that has been raising AI accelerator prices.

Those are two different products, and the distinction is the whole story.

What CXMT actually took share in

The growth came from expanding DDR5 and LPDDR5 output — commodity DRAM, the memory in servers, phones and laptops. That is where the 10 per cent sits.

High-bandwidth memory is a different product. HBM is the stack bonded next to an AI accelerator, and it is the component whose scarcity we have been tracking since August, when the memory bolted to an AI chip turned out to cost more than most people assume and a maker forecast five more years of shortage.

CXMT is targeting HBM production from the end of 2026. Yields are expected to be low, and low yields mean low volumes. So the company that just took a tenth of the DRAM market has no meaningful presence in the high-bandwidth segment, which is the one that is actually short.

10%CXMT's Q2 2026 DRAM share, from 4%
350kWafer starts per month by end-2026
720k / 595kSamsung and SK hynix, for comparison
End-2026When CXMT targets first HBM output

Third by capacity is not third by capability

At roughly 350 thousand wafer starts per month by the end of this year, CXMT would rank third among memory suppliers if the ranking were wafers alone. Samsung sits near 720 thousand and SK hynix near 595 thousand.

Wafer capacity is the wrong measure on its own. A wafer of DDR5 and a wafer of HBM are not the same economic object: HBM requires stacking, through-silicon vias and packaging that most of that capacity cannot perform, and it commands a multiple of the price. A supplier can be third by wafers and a long way further down by revenue or by margin.

What wafer capacity does measure well is the ability to flood a commodity segment, and that is exactly what has happened.

The incumbents made room

For Samsung and SK hynix, the uncomfortable part is that they helped create the vacancy.

HBM carries far better margins than commodity DRAM, and the rational response to an AI buildout is to move capacity toward it. Doing so leaves the low end less well served, at exactly the moment a new entrant is looking for somewhere to sell volume. CXMT did not have to displace anybody from the segment it grew into; it filled a space the incumbents were walking out of.

That is a familiar shape in semiconductors and it usually ends the same way. The entrant banks revenue and process experience at the bottom of the market and uses both to climb. Whether CXMT climbs from here depends on producing HBM at yield, which it has not yet demonstrated.

So is it good news for prices

For DDR5 and LPDDR5, yes, and it should show up first in the things that use them. Server memory, handsets and laptops are the segment where a fourth supplier at scale changes the negotiation, and there is already evidence of DRAM pricing pressure reaching buyers — we wrote about VMware answering DRAM prices by replacing memory with a drive that wears out.

For AI accelerators, no. The 20 to 50 per cent price rises we have reported are set by HBM scarcity, and CXMT has no HBM volume yet. A Chinese buyer looking at those quotes gets nothing from this quarter's market-share figure.

One exception is speculative. If CXMT's commodity volumes let Samsung and SK hynix shift still more capacity to HBM, the second-order effect is more HBM supply than there would otherwise have been — relief arriving later, and by a longer route than the headline suggests.

The lithography question underneath

Progress in deep-ultraviolet lithography is part of what has raised concern in Korea, and it is the variable that decides how far this goes.

Extreme-ultraviolet machines are unavailable to Chinese manufacturers under export controls, so everything CXMT has achieved has been achieved on DUV. There is real engineering headroom in multi-patterning on DUV and it is expensive in steps, yield and cost per wafer — which is survivable in commodity DRAM and considerably harder in HBM, where the die must also stack.

That DUV constraint, not capital or ambition, is what will set CXMT's ceiling.

What to watch

The first HBM shipments and, more importantly, their yield. A press release announcing HBM production says nothing; a yield figure, or a named tier-one customer, says everything.

And whether the share gain holds once the incumbents respond. Taking ten per cent from a segment competitors were de-prioritising is not the same as holding ten per cent when they decide to defend it.