CXMT’s 131 Billion Yuan Quarter Has a 1% Problem

CXMT’s preliminary filing landed on October 16: net profit of 131 to 132.5 billion yuan for the first nine months of 2026, against a 5.3 billion yuan loss a year earlier. Third-quarter revenue alone is projected at 114.7 to 119.7 billion yuan, up 589 to 619% year on year and 15 to 20% quarter on quarter. The headline number is large. The arithmetic underneath it is more interesting.

The quarter that did not accelerate

Strip out the nine-month total and look at Q3 on its own. Net profit attributable to shareholders is forecast at 53.395 to 54.895 billion yuan, against 52.843 billion in Q2. That is a sequential increase of 1.0% to 3.8%. Revenue grew 15 to 20% over the same span. The gap between those two rates is the point.

Profit stopped tracking revenue. That is not a demand-side event. TrendForce’s September 30 forecast has conventional DRAM contract prices rising another 10 to 15% in Q4 2026, with server DRAM supply remaining short of demand as fabs keep advanced nodes allocated to high-performance server products. Prices are still climbing. What changed is the rate at which CXMT converts those prices into profit.

Where the margin came from

CXMT’s first-half gross margin reached 84.84%. The company’s interim filing shows depreciation running close to three-quarters of revenue a year earlier, which is why it was losing money then. This year, revenue jumped nearly ninefold. The depreciation bill rose 28%. That bill shrank from roughly 75% of revenue to under 10%. The cost did not fall. The price moved past it.

The 84.84% margin is the distance between a fixed cost and a price spike. Capacity utilization climbed from 87.06% to 95.73%, which spread that fixed cost across more saleable die. Inventory built at lower wafer costs was sold into higher contract prices. Neither mechanism is a technology moat. Both are cyclical arithmetic, and both have boundaries.

The Q3 stall is those boundaries coming into view. Utilization is near its physical ceiling. The inventory-cost tailwind narrows as older, cheaper wafers clear. What remains is the price line itself. The price line is still moving, but the multiplier between price and profit is no longer expanding.

What the cycle says about 2027

The bear case centers on deceleration, not collapse. If the rate of price increase slows, CXMT’s profit base, already annualizing toward 175 to 177 billion yuan, makes any slowdown look like a cliff in year-on-year terms.

Morgan Stanley’s October note makes this argument without hedging: memory stocks look cheapest at the cycle top, because the market extrapolates peak earnings into a multiple that then expands violently when the earnings base rolls over. Samsung and SK Hynix currently trade at 3.5 to 3.7x forward earnings, Micron at roughly 6.1x. Those multiples are not a discount. They are the market pricing a declining earnings path.

The bull case rests on supply that cannot arrive fast enough. HBM consumes 2.5 to 4x the wafer area per bit of conventional DRAM, and fabs keep shifting capacity toward it. Morgan Stanley expects 2026 DRAM capex to jump 82% to $98 billion, with new capacity from CXMT, YMTC and others reaching the market in the second half of 2027. That is the timing problem: the supply response is large, but it is back-loaded. Prices may hold through the first half of 2027 and then meet a wave of new output.

The valuation number that cuts both ways

CXMT’s market capitalization sits near 3.06 trillion yuan. On annualized 2026 profit, that is roughly 17x. On 2025’s 1.875 billion yuan net profit, it is over 1,600x. The same company, two denominators, two incompatible conclusions.

The forward multiple is the one the market is trading on, and it implies that 2026’s earnings are a base, not a peak. If the supply response lands in late 2027 and prices soften, that 17x becomes something closer to 25 to 30x on revised earnings. Not catastrophic, but no longer cheap. The static multiple is a reminder of how thin the profit base was before the cycle turned.

Against global peers, the convergence is real but incomplete. CXMT’s forward earnings multiple has compressed toward the 5 to 6x range where Samsung and SK Hynix trade. Its price-to-sales ratio, however, remains multiples of theirs. That gap reflects the scarcity premium attached to a domestic DRAM champion with 10% global share, not the cycle earnings themselves. That premium is the first thing that reprices when the cycle narrative weakens.

The buyback is not the signal

CXMT’s proposed 1 to 2 billion yuan buyback equals roughly 1.5% of nine-month profit. Samsung’s shareholder return program and SK Hynix’s 40 trillion won buyback-and-cancel plan operate at a different scale because they sit at a different point in the capital cycle: mature fabs, depreciated assets, free cash flow to distribute. CXMT’s fixed assets carried 181.56 billion yuan as of June, with first-half depreciation of 14.51 billion yuan. It is still building. A company still buying fabs does not buy back stock in size. The buyback size is a statement about where CXMT is in its own buildout, not a verdict on shareholder friendliness.

What the audited Q3 report needs to show

The preliminary filing gives ranges. The formal report gives structure. Three line items will decide whether the sequential plateau was a pause or a peak.

Gross margin, quarter on quarter. If it held near 84 to 85% while revenue grew 15 to 20%, the cost structure absorbed the expansion without slipping. If it narrowed, the price wave is no longer outrunning the cost base.

Depreciation as a share of revenue. First-half depreciation was 14.51 billion yuan. If that ratio rose in Q3, new capacity and front-end costs are already compressing operating leverage. If it held or fell, the plateau is a volume-timing issue, not a margin problem.

Capacity utilization language. At 95.73%, the next increment of profit cannot come from spreading fixed costs further. It has to come from mix, meaning DDR5, LPDDR6 and HBM3E risk production, or from price. If utilization is already at the wall, the profit curve’s slope is now a function of pricing and product mix alone.

The semiconductor inflection, if it is confirmed, will not be confirmed by the 131-billion-yuan headline. It will be confirmed by whether that headline was the top of a slope or a ledge on the way to one. Q3’s formal report is the first hard evidence. Until then, the sequential number, 1% to 4%, is the part of the filing that matters most.