Skip to content

China's Biggest IPO Since 2010 Is a Memory Company. Your RAM Bill Explains Why.

DS
LDS Team
Let's Data Science
10 min
ChangXin Memory Technologies closed up 466% on its Shanghai debut on July 27, making it the most valuable company listed on a mainland exchange at roughly 3.3 trillion yuan. Its revenue rose more than 700% year over year in the first quarter as AI data centers drained the world's DRAM supply. U.S. lawmakers now want American companies barred from buying its chips.

Check the price of a 32GB DDR5 kit before you read any further. Tom's Hardware put the floor at $375 this summer, and DRAM prices have hit a 15-year high.

Every practitioner who has speced a workstation, sized a Kubernetes node, or read a cloud invoice in the last twelve months has already felt this. Memory is the constraint nobody planned for. The GPU shortage got the headlines; the DRAM shortage quietly repriced everything the GPUs plug into.

On Monday, July 27, that shortage produced a number in Shanghai large enough to reorder a stock exchange.

Shares of ChangXin Memory Technologies, China's largest memory chipmaker, opened for the first time on the Shanghai Stock Exchange's STAR Market and finished the day up 466%. The company ended its first session as the most valuable company listed on any mainland Chinese exchange, with a market capitalization of roughly 3.3 trillion yuan, according to the Associated Press.

A DRAM manufacturer founded ten years ago in Hefei is now worth more than any other listed company in China. That is not a story about Chinese equity markets. It is a story about what AI did to memory.

The Debut Was the Largest Mainland Listing in Sixteen Years

CXMT priced its offering at 8.66 yuan per share and raised at least $8.6 billion, making it mainland China's second largest IPO ever. The only bigger one was Agricultural Bank of China's dual listing in Shanghai and Hong Kong in 2010, which raised 22.1 billion dollars.

The financials underneath the pop are the part worth reading twice. CXMT reported revenue of 50.8 billion yuan in the first three months of 2026, a rise of more than 700% year over year, driven almost entirely by demand from AI infrastructure.

Seven-hundred-percent revenue growth is not a semiconductor number. It is what happens when a commodity becomes scarce and a producer with capacity discovers it can name a price.

"CXMT plays a critical role in China's AI push, particularly in the face of U.S. export controls," said Kyle Chan, a fellow at the Brookings Institution who studies Chinese technology policy.

The Shortage Is the Whole Story

DRAM is the working memory that sits beside a processor. High-bandwidth memory, or HBM, is a stacked, much faster variant bolted directly onto AI accelerators. HBM is what an H-series or Instinct-class GPU uses to hold model weights and activations, and it is the component every accelerator vendor has been fighting over since 2023.

Samsung, SK Hynix and Micron responded to that fight by moving production toward the higher-margin server and HBM products. That decision starved everything else. Conventional DRAM for laptops, desktops and non-accelerated servers has been in shortage ever since, and the price data shows no clean exit.

SegmentQ3 2026 contract price moveSource
Conventional DRAMUp 13% to 18% quarter over quarterTrendForce
NAND FlashUp 10% to 15% quarter over quarterTrendForce
Prior quarter (Q2 2026)Roughly 60% quarter over quarterTrendForce

TrendForce reads the slowdown from 60% to the mid-teens as a demand-side event rather than a supply-side one. Consumer electronics makers hit the ceiling of what they could absorb and stopped paying. Supply did not improve. Buyers simply ran out of room.

On the enterprise side, TrendForce expects demand to stay healthy through 2027 as improving CPU availability supports more AI server deployments built around registered DIMMs. Server DRAM is expected to remain undersupplied through the third quarter, with increases moderating only because a slice of purchases sits under long-term supply agreements.

Samsung and SK Hynix have both warned the shortage could run past 2027. Goldman Sachs, cited by 24/7 Wall St., estimates a 2026 DRAM supply-demand gap of 4.9%, which it calls the most severe shortage in 15 years.

What CXMT Can Build, and What It Cannot

The bull case is straightforward: a fourth major DRAM supplier scaling aggressively is the only structural fix for a shortage caused by three suppliers reallocating capacity.

The constraint is equally straightforward. American-led export controls bar China from importing HBM, and they restrict CXMT's access to the most advanced chipmaking tools, forcing it to depend on Chinese equipment makers. Chan called the company China's best shot at developing its own advanced HBM to power Chinese AI models, while noting it faces real bottlenecks in scaling manufacturing capacity. Those models are no longer an afterthought, given that Chinese open-weight releases overtook American ones on the open leaderboards this spring.

DRAM makerShare of global shipments, 2025
Samsung Electronics36%
SK Hynix29%
Micron TechnologyAbout 24%
CXMTRoughly 8%

Those figures come from Counterpoint Research. CXMT accounted for approximately 9% of global shipments in the first quarter of 2026, and Counterpoint forecasts about 11% by 2028. The same firm estimates CXMT will likely need at least 15% global share to be competitive over the long run.

"Trade restrictions on tools are remaining as the key challenge for CXMT," said MS Hwang, a Counterpoint research director who specializes in memory semiconductors.

The product mix matters as much as the share number. More than 98% of CXMT's revenue comes from commodity DRAM with effectively no HBM presence, according to figures compiled by 24/7 Wall St., which also cited analysts describing the company as at least one generation behind in high-bandwidth memory. A fourth large supplier of the memory that is already in shortage is real relief for laptops, phones and general-purpose servers. It is not relief for the accelerator racks where HBM is the bottleneck.

That gap between 11% and 15% is the entire investment thesis, and it is being priced today at a valuation larger than any other listed company in China.

The Other Side of the Trade

Three arguments cut against the debut, and none of them are fringe.

  • The market cap is smaller than it looks. Despite becoming the most valuable listed company on a mainland exchange, CXMT is still worth less than Samsung Electronics, SK Hynix or Micron. A first-day pop on a domestic exchange with heavy retail participation is not a global repricing of memory.
  • A ban is actively on the table. Some U.S. lawmakers have called on the Trump administration to block American companies from buying CXMT memory over national and economic security concerns. CXMT is among the Chinese companies the Pentagon says have links to the Chinese military, a designation Beijing has rejected in most cases. A company that cannot sell into the largest AI market on earth relieves that market's shortage only indirectly.
  • Capacity is not a press release. Chan's caution about supply chain bottlenecks is the operative one. Adding DRAM fab capacity takes years and the best lithography tools, and CXMT has restricted access to both.

The incumbents did not shrug it off. 24/7 Wall St. reported that SanDisk fell about 12%, SK Hynix about 8% and Micron about 5% as the debut rattled memory stocks, and that Micron closed at 900.20 dollars on July 27 after slipping roughly 13% over the prior week.

Worth noting alongside the debut: South Korea's SK Hynix completed a 26.5 billion dollar share offering on the Nasdaq earlier in July. Memory is where the capital is going on both sides of the export-control line.

What This Changes for Teams That Buy Compute

The practitioner takeaway is not about a Shanghai ticker.

  • Memory is now a first-class line item in cost models. Inference cost per token has historically been reasoned about in terms of accelerator hours. With conventional DRAM up double digits quarter over quarter and server DRAM still undersupplied, the memory attached to those accelerators is moving fast enough to matter in a budget.
  • On-device and local-first plans get more expensive. Running models locally trades cloud spend for hardware spend, and the hardware side just got repriced. Anyone planning a local agentic workstation buildout should price DRAM before committing to a design.
  • Efficiency work has a harder dollar value than it did last year. Quantization, KV-cache compression and activation offloading all buy back memory. When memory prices climb 13% to 18% a quarter, those techniques stop being research and start being procurement. DeepSeek's engineers made their models 85% faster without new chips, and that kind of work is now worth more than it was.
  • Supplier diversification is happening at every layer. Labs are already spreading accelerator risk, as Anthropic did when it agreed to run Claude on AMD hardware. Memory is the next layer down, and it has fewer credible suppliers than accelerators do.

The Bottom Line

A memory company just became the most valuable listed business in China because the world cannot make enough of a component that costs a few dollars per gigabyte and gates everything above it.

That inversion is the honest summary of this AI cycle. The scarce thing was never the idea, and increasingly it is not the accelerator either. It is the boring, high-volume, capital-intensive substrate that the accelerator sits on, and the industry spent a decade treating it as a commodity that would always be there.

CXMT's investors are betting that a company with restricted tools and no access to imported HBM can close the distance between roughly 8% of global shipments and the 15% Counterpoint says it needs, while three incumbents with better equipment defend their positions. Counterpoint's own forecast has CXMT reaching about 11% by 2028. The market priced it as though the gap is already closed.

The next time a training run stalls on host memory, or a quote for a 512GB server node comes back higher than the GPUs inside it, remember which exchange decided that was worth 3.3 trillion yuan.

Sources

Practice interview problems based on real data

1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems