China's CXMT Popped 466% on Its Debut; Memory Stocks Fell

China's Most Valuable Company in a Single Session

On the morning of July 27, ChangXin Memory Technologies — CXMT for short — opened for trading on Shanghai's STAR Market at 8.66 yuan a share. The price cleared 54 yuan intraday. By the close, the company was worth 3.66 trillion yuan, or close to $500 billion. It had become the most valuable listed company in China.

The offering raised 57.92 billion yuan, roughly $8.6 billion — the second-largest IPO on the mainland since Agricultural Bank of China came to market in 2010.

Seven days later, on the morning of Monday, August 3, Reuters reported, citing two unnamed people familiar with the matter, that CXMT was in talks with district authorities about a second memory fab in Beijing. Within the same hour, Micron fell 4% in the premarket to $790.50. Seagate and Western Digital each shed more than 7% during the session.

The link between the two events is not an accident of the news cycle. Memory is a commodity, and in commodities the price is set by whichever producer is most willing to sell.

Three of those four numbers belong to China and one to the United States. Connecting them means first understanding how the memory business makes money — and, more importantly, how it loses it.

From Hefei to Shanghai: Ten Years of a State Project

CXMT was founded in Hefei in 2016 under the name Innotron. In 2019 it was turning out 20,000 wafers a month. Today it runs three fabs, all of them 12-inch, each producing roughly 100,000 wafers a month. Once the new projects in Shanghai and Hefei come online, total capacity is targeted to exceed 600,000 wafers a month. A doubling, in other words.

The shareholder register tells you what kind of company this is. According to the IPO prospectus, the five largest holders are Qinghui Jidian (21.67%), ChangXin Integrated (11.71%), the second phase of the China Integrated Circuit Industry Investment Fund (8.73%), Hefei Jixin (8.37%) and Anhui Investment (7.91%). No single shareholder holds control, but the weight of the list sits with local governments and state funds.

The company's constraints are just as clear. It has been barred from US federal procurement since 2022, was added to the Pentagon's list of firms with military ties in June 2026, and cannot buy the most advanced production equipment. That carries a measurable cost: CXMT burns roughly 30% more wafers than its rivals to produce the same quantity of chips.

Even so, its share of the global DRAM market climbed from 3% in the first quarter of 2025 to 8% in the first quarter of 2026. It is now the world's fourth-largest DRAM maker. The three companies ahead of it — Samsung, SK Hynix and Micron — held close to 90% between them in that same quarter, according to Counterpoint.

The Winter Prices Quadrupled in Three Months

To understand today's unease, look back one winter. The spot price of a DDR5 memory chip was $6.84 in September 2025; by December it had reached $27.20. Counterpoint puts the first-quarter 2026 increase in memory prices at 80% to 90% over the prior quarter.

AI data centers really did buy an enormous amount of memory. But demand did not rise 300%. So why did the price quadruple?

The Arithmetic of the Marginal Unit: Why 10% Excess Supply Cuts Prices 50%

That asymmetry explains the entire memory sector. Price is set not by the size of the gap between supply and demand, but by its direction. A small deficit sends prices vertical; a small surplus collapses them.

So why doesn't a producer cut output the moment it sees a surplus? Because most of the cost has already been paid. Take a simple example: a unit of chip sells for 100, cash production cost is 45, and the fab's depreciation allocated to that unit is 30. Profit is 25.

Now let the price fall 30%. Revenue drops to 70, cost stays at 75, and the loss is 5. Consider idling the fab instead: sales go to zero, but depreciation is still 30. The loss becomes 30, not 5. Selling at 70 means covering the 45 of cash cost and applying the remaining 25 against fixed costs — producing at a loss beats not producing at all.

Because everyone in the industry runs the same calculation, nobody wants to be the first to cut. That is why memory downturns are both deep and long.

July: The Month the Market Remembered the Arithmetic

The real information in that table is not the size of the numbers but their distribution. The Nasdaq finished July down 3%; the memory names lost between a third and a half of their value. The decline did not hit the market broadly — it hit one sector precisely. For semiconductors, July was the worst month since 2008. Micron closed at $823 on July 31, 39% below its high for the year.

The charts below are live: they show not the event itself, but where the stock stands today.

Checking where the index went over the same window is the most practical way to separate a sector-specific move from a market-wide one.

A Seven-Day Chronology

The limits of that last item are worth stating. Reuters reports that the talks are at an early stage and that the company is seeking at least 60 million yuan (about $8.9 million) in support from the district government. That figure is tiny next to the cost of a DRAM fab, which leaves the project's eventual scale and financing structure unclear. On its own, the report is a statement of intent, not a capacity commitment.

The Other Side: Two Readings of the Same News

The last row of that table is where investors go wrong most often, because it runs directly against intuition.

What Is Left

CXMT accounts for roughly one-twelfth of global DRAM output today. It does not yet make the high-bandwidth memory that AI servers demand. The equipment restrictions are real and its technology gap is measured in years. All of that is true.

But price is set by the marginal producer, and the identity of the marginal producer depends not on who is largest but on who can endure the longest. A company that raised its capital in the equity market has to cut output when it starts losing money. A producer whose capital comes from state funds, and whose mandate includes supply security alongside profit, does not feel the same pressure.

What was repriced in July was not Micron's earnings. It was the assumption that memory prices would stay elevated through 2028. That assumption may still prove correct. But it is now an assumption that can be questioned on the strength of one company's fab plans — which was not the case last winter. That is precisely where volatility comes from: news that changes not the price, but the assumption the price rests on.

This article draws on information reported by Reuters, Counterpoint Research, SCMP, TechNode and the company's IPO prospectus. The report of a second plant in Beijing rests on two unnamed sources familiar with the matter cited by Reuters and has not been confirmed by the company. Price and market share figures are as of the date of publication. This is not investment advice.