Neither, actually. CXMT's first quarterly report since its July IPO is a boom in ordinary DRAM caused by everybody else abandoning ordinary DRAM. And China's HBM3E push is a supply-security policy for domestic AI chip designers, not a price cut coming your way.

Here is what the number is made of.

CXMT, Q3 2026
Revenue ¥118.71bn (about $17.8bn), +613.2% YoY, +19.3% QoQ
Net profit ¥54.4bn (about $8.2bn), from a loss a year earlier
Gross margin 89.53%
R&D, first nine months ¥11.07bn (¥4.2bn of it in Q3)
Cash on hand ¥285.16bn
HBM3E Limited production within two months, 20% to 30% yield
Share price Down 32% from the August peak, and 2.5% lower on results day

The money is a vacancy, not a breakthrough

CXMT is China's largest DRAM supplier by revenue, with 8% of global DRAM revenue in Q1 2026 and customers that read like a map of Chinese compute: Alibaba Cloud, ByteDance, Tencent, Lenovo.

None of that produced ¥54.4bn. Samsung, SK Hynix and Micron pulled wafers toward HBM, where margins are fat, and the conventional DRAM market thinned out. CXMT filled the gap and the price went to the seller. That is the whole mechanism.

The wafer math explains it better than any press release. Making one bit of HBM costs roughly three times the wafer capacity of ordinary DRAM, because the die is bigger, the stacking is harder and yield losses compound. The big three's HBM wafer starts went from about 18% of their total at the end of 2025 to 22% at the end of 2026, while HBM's share of total bit supply crawled from 8% to 9%. Nearly a quarter of the wafers produce about a ninth of the bits. CXMT is selling into the space everyone else vacated.

Check the guidance against the result. On 17 October CXMT told investors to expect Q3 net profit of ¥5.34bn to ¥5.49bn, according to Chosun. It delivered ¥54.4bn, roughly ten times the top of its own range, while Bloomberg notes net income was largely unchanged from Q2. This was price, not capability, and the company could not predict it three weeks earlier.

You are not getting cheaper memory

This matters if you buy phones, laptops or servers.

TrendForce's 30 September forecast has conventional DRAM contract prices up 10% to 15% quarter on quarter in 4Q26 and NAND Flash contracts up 15% to 20%. PC DRAM is up 13% to 18%, and PC OEMs keep buying because they expect supply to stay tight in 2027. Mobile DRAM is still going up, but the pace moderates: earlier increases were already priced in, and phone makers are burning down cheap inventory while ordering almost nothing new. NAND wafer prices have reached historic highs that module buyers refuse to absorb. DRAMeXchange data put commodity PC DRAM module prices at record highs in September, with standard DDR4 8Gb modules averaging $26 after 18 straight months of increases.

So no. Every category of memory is going up, and the reason is the same HBM pull that made CXMT's quarter. Storage has gone from 10% to 15% of a finished device's bill of materials up to 30% to 40%, which makes memory the single most expensive part of a phone, a laptop or a server, ahead of the processor.

Nomura expects the storage shortage to run through 2028. Citi sees a DRAM supply gap in 2027 and 2028. Waiting out a memory price collapse is a 2028 story.

Chinese HBM is a policy, not a product

Reuters reported that CXMT plans limited HBM3E production within two months, with yields of 20% to 30% against about 70% at SK Hynix. Samples have already gone to domestic AI chip designers. In testing, performance lands closer to SK Hynix's previous-generation HBM3 than to current HBM3E, and the expected price is above overseas equivalents.

At a 20% yield, three wafers in five get scrapped or reworked, and wafer, packaging and test cost spread over the survivors. US export controls list CXMT's chip facilities as controlled facilities, so Chinese AI builders cannot simply buy SK Hynix stacks. HBM3E is a supply-chain insurance policy, and the admission ticket has to be paid in engineering years.

The timing is brutal. Samsung started HBM4 mass production in February 2026 and has climbed from under 60% yield to about 80%; SK Hynix holds roughly 54% of the HBM4 market. TrendForce now expects HBM's blended ASP to jump 121% year on year in 2027, with 8-Hi stacks becoming the priority configuration because they use fewer dies. CXMT is entering a generation late, at a fraction of the yield, into a market that is repricing upward.

Nobody serious is buying Chinese AI boxes for cheaper HBM in the next 18 months. They are buying them because the alternative is no HBM at all.

What the balance sheet buys

¥285bn of cash is the real headline. CXMT is funding fabs in Beijing, Shanghai and Hefei projected to double capacity to 600,000 12-inch wafers a month. It has validated its sixth-generation DRAM process, plans server DRAM on a 4F² cell layout by end-2026, and is preparing to enter NAND. Its ¥58bn IPO was Asia's largest this year; the stock surged more than five times on its Shanghai debut, made CXMT China's most valuable listed company in July, and has since given back 32%.

The market is not pricing the quarter. It is pricing 2027 and 2028, when the AI capital-expenditure assumption gets tested, and whether the profit came from product gains or a temporary price spike. ¥4.2bn of R&D against a ¥54.4bn profit says the company has ammunition. It does not say it has caught up.

The verdict for buyers

Do not wait for CXMT to make memory cheap. It is not trying to. It is trying to be the biggest conventional DRAM supplier in a market where the three companies that set the price are busy elsewhere, and it is spending a windfall on HBM yield.

If you are buying a phone or a laptop this quarter, buy the one you want and assume memory is already in the price. If you are specifying servers for an AI build-out, Chinese HBM does not move your cost curve yet. Price it as a 2028 supply option.

¥54.4bn is ammunition. 20% is a road sign. The gate is still ahead.

Sources