CXMT's $488B debut: DRAM prices are now your hardware budget

A commodity memory maker just became the most valuable company listed on mainland China. CXMT closed its first Shanghai trading session up 466%, at a market capitalization around $488 billion — for a business that makes DRAM, the least glamorous component in any machine you own. That valuation is the market pricing one fact you should already be budgeting around: DRAM prices are not coming back down soon, and that shows up in your hardware line, not somebody else's.

What actually happened

Per Reuters, CXMT priced at 8.66 yuan and closed at 49 yuan on July 27, after raising 57.92 billion yuan (about $8.6 billion) in the largest mainland Chinese semiconductor offering on record. The close put its market cap at 3.3 trillion yuan — roughly $488 billion — ahead of ICBC.

The financials behind it are the real signal. The Hefei-based company, the world's fourth-largest DRAM producer at under 8% global share, guided first-half revenue of 110 to 120 billion yuan, more than seven times the prior-year figure, with net profit of 66 to 75 billion yuan after a loss a year earlier. Reuters quotes an analyst expecting the memory market to stay tight with price increases continuing through the end of 2027.

That matches what buyers on the other side are already reporting. Microsoft attributed roughly $25 billion of its calendar-2026 capex purely to higher component pricing on its FY26 Q3 earnings call. When the largest buyer in the world can't negotiate its way out of a memory bill, neither can you.

Why DRAM prices matter for your business

If your plan for controlling AI costs involves owning hardware — a box under the desk running an open model, a beefier VPS, a laptop refresh for the team — memory is the line item that moved. RAM is the single component with the least substitution: you can pick a cheaper CPU or a smaller SSD, but a model that needs 64GB needs 64GB.

Three practical calls. Buy the RAM now if the refresh is coming inside twelve months — spec it high on the machines you're already ordering, because upgrading later at 2027 prices costs more than over-provisioning today. Price the self-hosting plan against hosted inference honestly, including the hardware amortization, not just the token savings; the break-even moved when memory repriced. And treat quantized models as a budget lever, not a compromise — dropping a model to a smaller memory footprint is often the difference between one machine and two.

The broader read: everyone watching AI cost curves is watching model prices, which fall. The hardware underneath them is going the other direction, and it's the half of the equation nobody quotes you.

Sources: Reuters, Microsoft FY26 Q3 earnings call.