CXMT sues the Pentagon for the right to be an ordinary DRAM supplier

Take Quartz's figures at face value and the net margin on CXMT's first half is 51.6% ^[19], which is an awkward number for the tidiest version of the company's own story. Run the growth rate backwards and the base is tiny: revenue up 874% puts the year-earlier half near 15.4 billion yuan ^[20], call it $2.3 billion at the exchange rate implied by the company's own dollar translation ^[21]. A supplier annualising toward $44.8 billion ^[22] while keeping half of it as profit is collecting the shortage rent rather than dissolving it, and that is the first thing I would put to anyone arguing that letting the fourth player in would cool DRAM prices.

The bit arithmetic points the same way. The three incumbents held about 90% of Q2 2026 DRAM revenue ^[8], leaving at most a tenth of the industry's money for CXMT and every other minor supplier combined ^[23], while CXMT already moves close to 9% of global bits ^[10]. Dollars per bit at or below the industry average ^[24] is what you would expect from conventional DRAM rather than the high-bandwidth parts AI servers are queueing for, and Counterpoint's forecast is that conventional prices climb anyway, because agentic AI and server builds keep absorbing the capacity that used to make them ^[11].

What is being litigated, then, is a customer's political cost, or rather the transferable version of it. The Pentagon's 1260H list is not the Commerce Entity List and American companies remain free to buy CXMT chips ^[6]; the harder edge is the FY2023 NDAA clause barring federal agencies from using goods containing CXMT semiconductors from December 2027 ^[7], which reaches through any manufacturer with a government line of business. CXMT told the District Court for the District of Columbia ^[2] that the determinations are unsupported by the factual record and that it "is not affiliated with the Chinese military" ^[4], selling into civilian and commercial channels ^[14]; the Pentagon said it cannot comment on pending litigation ^[5]. Apple, meanwhile, is weighing CXMT memory ^[15] after AI-driven memory costs pushed some iPad and MacBook prices up by as much as 20% ^[16], and a bipartisan group of senators has told Apple to refuse, on the explicit theory that approval by so demanding a buyer would carry the rest of the industry with it ^[17].

That senatorial theory is the strongest evidence for my read, which is that the marginal DRAM bit available to Western buyers is now gated by a designation rather than by a fab. This is probably wrong in one specific way: if the binding scarcity sits in high-bandwidth memory and packaging, then admitting the only credible fourth supplier ^[18] relieves phones and PCs and leaves AI server pricing exactly where it was. Two other paths are open. CXMT wins the case and discovers the 2027 procurement clause still keeps regulated buyers away, making the victory reputational; or Apple qualifies it, other original equipment makers follow, and conventional contract prices flatten, which is the only outcome that would show policy was setting the price. The falsifier is cheap to check: if CXMT volume reaches Western buyers and contract prices keep rising anyway, the fourth supplier was never the constraint.

One caveat on the record. All of this arrives through a single publisher relaying Reuters on the filing ^[1], Counterpoint on share ^[8] and Quartz on the accounts ^[12], and a $22.4 billion half-year at a company whose bit share is under a tenth is large enough that I would want the underlying statements in hand before leaning on the margin.