CXMT’s HBM3E Pilot Run: Is a ¥4 Trillion Valuation Buying a Breakthrough or Production Yield?

The breakthrough is reported; the revenue evidence is not

A public summary from The Information dated August 31–September 1, 2026 says ChangXin Memory Technologies, or CXMT, has started small-batch HBM3E production and plans to expand output in 2027. The product is one generation behind leading products already in mass production. The conflict is immediate: the technical node advanced, but CXMT has not publicly confirmed HBM yield, qualified-stack volume, customer qualification pass rates, standalone revenue or capacity. Every one of those fields is therefore N/A / undisclosed, not zero.

The report legitimately raises the value of CXMT’s technology option. It does not establish a revenue inflection. Starting wafers in a pilot flow shows that a process is operating; it does not show that the final stack meets a customer platform’s requirements or that cash has returned to the company.

How much expectation sits inside roughly ¥4 trillion

A public intraday quote snapshot on September 2, 2026 was about ¥58.88. Applying the roughly 67.884 billion shares shown on a public market page gives an equity value near ¥4.0 trillion. Applying the 66,880,886,077 shares in the Shanghai Stock Exchange listing announcement gives about ¥3.94 trillion. This is a difference between public share-count conventions, so the valuation is only an order-of-magnitude illustration; the quote also changes.

Mechanically doubling first-half revenue of ¥150.3101 billion and net profit attributable to shareholders of ¥77.6052 billion produces a simple annualized P/S near 13.3x and P/E near 25.8x at a ¥4 trillion value. Those are not forecasts or price targets. They are stress tests that extend one strong half at a high point in the DRAM cycle.

The denominator is qualified stacks, not wafer starts

The investable formula is: Saleable HBM stacks = wafer input × bare dies per wafer × front-end yield × TSV/stacking yield × packaging-test yield × customer-qualification pass rate. CXMT has not disclosed valuation-ready figures for wafer input, dies per wafer, front-end yield, TSV/stacking yield, packaging-test yield or customer-qualification pass rate. Every multiplier is N/A.

Because this is a multiplication chain, loss at one stage flows through every later stage. Wafer starts can rise while saleable output remains small if stacking, thermals, package testing or qualification becomes the constraint. Translating “generation catch-up” directly into revenue mistakes an engineering milestone for a bank receipt.

The leaders have already moved the gate to HBM4

SK hynix’s official second-quarter 2026 results say HBM4 entered mass shipment during the quarter and production will increase in the second half; HBM4E samples shipped in the first half. The company identified high yield, stable supply and cost competitiveness as advantages. Micron’s official fiscal third-quarter 2026 release says HBM4 is in high-volume shipment for major customer platforms and HBM4E is expected to reach production in 2027.

That does not negate CXMT’s breakthrough. It shows that the competitive window for AI memory chips keeps moving. Customers compare deliverable volume, power, reliability, platform qualification, supply stability and cost—not a static generation name. As leaders advance generation and price, the duration of the HBM3E option may contract. Compare Valuefocus’s Micron stock research and HBM memory-cycle analysis.

Separate the CXMT business from the option

A more disciplined expression is: Enterprise value = traditional DRAM business value + probability-weighted HBM option value − incremental capex and working capital required for scale. The traditional business belongs on normalized DDR and LPDDR cycle cash flows. HBM belongs in separate success, delay, low-yield and failed-qualification scenarios.

A Damodaran-style probability-weighted and real-options lens allows a technical milestone to raise value, but only by “change in success probability × present value of cash flow if successful.” It does not capitalize the entire TAM at once. Use Valuefocus Reverse DCF to infer the HBM success probability, margin, capex and timing embedded in the market price. The research methodology and value-investing framework make those assumptions auditable.

Three investor lenses return to the same cash denominator

A Buffett owner-earnings lens does not stop at ¥77.6 billion of net profit. It asks for maintenance capital expenditure, working capital and free cash flow needed to preserve competitive position. CXMT’s company-wide operating cash flow is strong, but HBM cash flow is not separately disclosed; the ¥131.2 billion total cannot be assigned directly to HBM.

Damodaran places the milestone inside a probability tree. Munger-style inversion asks first how the story breaks. Four disconfirming indicators matter: yield below an economic production threshold, delayed customer qualification, constrained advanced packaging, and leading producers cutting price or advancing generation fast enough to compress HBM3E’s window. These are closer to intrinsic value than the words “small-batch production.”

A four-quarter monitoring table

For the next four quarters, investors do not need unsupported capacity rumors. They need disclosed denominators. Unknown stays N/A rather than becoming a false zero.

Valuation still ends with qualified output

CXMT’s reported HBM3E pilot run deserves to be treated as a technological breakthrough and an option. But a roughly ¥4 trillion CXMT valuation is not buying the statement that wafers have started. It is buying the possibility that CXMT can repeatedly produce, qualify, sell and collect cash. The DRAM cycle can amplify current profit and the error in mechanically annualizing it.

Until yield, qualification, standalone revenue and capex are disclosed, the honest answer is N/A. Investment research should not fill the blank with zero; it should identify exactly which blanks the current price requires to become cash.

Sources, scope and research note