A Buyback Disclosure From China's Memory Champion Carries a Tenfold Data Error

A report on ChangXin Memory Technologies that circulated through blockchain-focused news channels this week contained a profit figure inflated by a factor of ten, according to a line-by-line audit of the original text. The disclosure, which also announced a share buyback with a ceiling of 72.69 yuan per share, described a company that had swung from losses to profit across the first three quarters of 2026. Both the profit number and the corporate identity require correction before any downstream analysis can be trusted. The arithmetic resolves once stated. The English word "billion" denotes one thousand million, or ten to the ninth power. The unit transliterated as "yi" denotes one hundred million, or ten to the eighth power. Taken literally, 132.04 billion yuan equals 1,320.4 yi. The corrected figure is 132.04 yi, or 13.204 billion yuan. The published number overstated profit by a factor of ten, and the error compounded when the second reading accepted the inflated total without testing it against the balance sheet of any comparable company. Three checks confirm the inflation. First, a 132 billion yuan profit across nine months implies roughly 176 billion yuan annualized, which would place a single Chinese memory maker at two-thirds of TSMC's entire 2024 net income of about 36.5 billion US dollars, despite TSMC generating more than ten times the revenue. No semiconductor company in China earns at that scale. Second, the unit confusion is a textbook machine-translation artifact rather than a clerical slip. Third, the resulting valuation would be inconsistent with every primary-market round the company has ever raised. The entity behind the filing is almost certainly ChangXin Memory Technologies, the only DRAM manufacturer in China operating at scale, founded in Hefei in 2016. The obstacle is that ChangXin spent its entire history as a private company, while a centralized auction buyback with a stated price ceiling is an action reserved for listed issuers. Two explanations survive. Either ChangXin completed a listing on the Shanghai STAR Market between 2025 and 2026 and this is its first post-listing repurchase, which would make the 2026 timestamp coherent, or the name refers to a different listed company altogether. The market capitalization implied by the buyback terms supports the listing hypothesis. A ceiling of 72.69 yuan per share against a share count of roughly two billion implies an equity value near 145 billion yuan. That number sits almost exactly on top of the 140 to 150 billion yuan valuation ChangXin carried in primary markets during 2025, which would be an improbable coincidence for an unrelated company. The assumption is therefore workable but unproven, and every conclusion that depends on it deserves a lower confidence band. DRAM does not obey the nanometer narrative used for logic chips. Process generations are described by half-pitch and labeled 1x, 1y, 1z, 1alpha, 1beta and 1gamma. ChangXin shipped 1x around 2019 against a 2016 industry debut, reached 1y in 2021 and moved 1z into volume in 2023 and 2024. The current workhorse is 1z at roughly fifteen nanometers for DDR4 and DDR5, while 1alpha sits in risk production. Samsung, SK Hynix and Micron all began 1alpha in 2021 and 2022. The gap is two to two and a half generations, or three to five years. The binding constraint is the absence of extreme ultraviolet lithography. Samsung and SK Hynix introduced EUV at the 1alpha generation for critical layers. ChangXin works exclusively with deep ultraviolet scanners in the ASML NXT:1980Di and NXT:2000i class, relying on self-aligned double and quadruple patterning. That path doubles mask counts, raises cost per wafer, tightens overlay budgets and caps yield. Yield estimates put ChangXin's 1z DDR4 output at seventy to eighty percent against eighty to ninety percent for the incumbents, with early DDR5 materially worse at fifty to sixty-five percent. In a cost-driven commodity, ten points of yield translate into twelve to fifteen percent higher unit cost. High-bandwidth memory is where the story either breaks or compounds. HBM requires through-silicon vias and 2.5D advanced packaging that no DRAM maker can complete alone. ChangXin displayed an HBM2 sample in 2024 and has HBM3 in development, roughly three years behind SK Hynix. Domestic packaging capacity at JCET and Tongfu is still ramping, and the equipment needed for TSV formation falls inside the December 2024 export controls. Without HBM, ChangXin sells DDR5 into a cycle. With it, the equity story shifts from cyclical to structural. Supply chain exposure is concentrated to a degree that is easy to understate. Lithography is one hundred percent imported with no domestic substitute; Shanghai Micro Electronics remains at ninety nanometers. Etch, deposition and chemical mechanical polishing have partial local suppliers including AMEC, Naura and Hwatsing. Metrology still leans on KLA. ArF and KrF photoresists are dominated by Japanese vendors, with Nata and Red Avenue qualifying slowly. Twelve-inch wafers come from National Silicon and TCL Zhonghuan. Electronic specialty gases come from Huate and Yoke. Equipment localization on a DRAM line sits at fifteen to twenty-five percent, materials at twenty to thirty percent, and EDA above ninety percent imported. Capacity across Hefei Fab1, Fab2 and Beijing Fab3 is estimated at 250,000 to 300,000 twelve-inch wafers per month in 2025, running above ninety percent utilization. Each additional 10,000 wafers per month of DRAM capacity costs six to eight billion yuan, which puts capital expenditure above fifty percent of revenue, well beyond TSMC's thirty-five to forty-five percent band. New lines carry depreciation from the first quarter of operation on a five to seven year straight-line schedule, dragging gross margin by ten to twenty points. That arithmetic explains why the swing to profit arrived with the price upcycle rather than with a technology milestone. Demand composition favors the incumbents. Servers and data centers absorb about thirty-five percent of DRAM revenue and are growing above twenty percent on AI buildouts. Mobile takes thirty percent, PCs twenty, automotive and industrial fifteen. A single H100-class accelerator carries six HBM stacks totaling roughly eighty gigabytes, and an AI server consumes six to eight times the DRAM of a general-purpose machine. ChangXin's mix is weighted toward PCs, consumer devices and handsets, which makes servers and HBM both its largest gap and its largest addressable upside. The cycle position matters more than any single quarter. DRAM has traced a three to four year rhythm: an upcycle from 2016 to 2019, a downturn into 2020, expansion through 2022, contraction through 2023, and an advance from 2024 into 2026. On that cadence the current expansion is in its middle-to-late stage, with peak pricing risk concentrated in the second half of 2026 and 2027. ChangXin holds five to eight percent of global share, which makes it a price taker: it captures outsized leverage on the way up and absorbs disproportionate damage on the way down. A profit peak in 2026 is the base case, not a pessimistic one. Policy shapes both the ceiling and the floor. ChangXin is not on the BIS Entity List as of public records, but the December 2024 package placed HBM and a range of DRAM manufacturing tools under stricter control, and the company is monitored in practice. ASML's NXT:2000i has required a Dutch license for China shipments since September 2023, and the NXT:1980Di followed in September 2024. Japan's July 2023 rules cover twenty-three equipment categories including resist and cleaning. Beijing's countermeasures on gallium, germanium, graphite and antimony create friction without equivalence, because substitution exists for each. The third phase of the National Integrated Circuit Industry Investment Fund, registered in May 2024 with 344 billion yuan, is the offsetting force. Market share reinforces the same picture. Samsung holds about forty percent of DRAM, SK Hynix thirty to thirty-two percent with more than half of HBM, Micron twenty-two percent, and ChangXin five to eight percent as the fourth global player and the only scaled Chinese one. Research intensity runs fifteen to twenty-five percent of revenue during the catch-up phase, above the ten to fifteen percent the incumbents spend, but the same money buys less capability when the tools and EDA licenses are restricted. Bargaining power is weak upstream and moderate downstream. Barriers to new entry remain prohibitive, and no credible new DRAM entrant has emerged globally in two decades. On the corrected numbers, the conditional valuation exercise is uncomfortable. A profit of 13.2 billion yuan against an equity value near 145 billion yuan produces a price-to-earnings ratio of roughly eight to eleven times on annualized 2026 estimates, against fifteen to twenty times for Micron and ten to fifteen for SK Hynix. That discount is not generosity from the market; it is the standard cyclical trap, in which earnings peak exactly when the multiple troughs. The buyback authorizes one to two billion yuan, between seven-tenths of one percent and one and four-tenths percent of market capitalization, which is a signal rather than a commitment of balance sheet capacity. The risk ledger is short and heavy. The data-source failure has already occurred and would corrupt every downstream metric if left uncorrected. A cycle top in the second half of 2026 carries sixty percent or better odds on timing that nobody can pin down. An escalation to servicing and spare-parts restrictions on installed tools, which would freeze expansion and eventually erode utilization, sits at thirty to forty percent. A failure to bring HBM3 to volume by 2027 sits near forty percent and would collapse the growth narrative back into a commodity one. Against that, localization is the durable upside: Chinese DRAM self-sufficiency remains under ten percent, and moving share from five to eight percent up toward fifteen percent would double revenue on policy alone. For anyone who encountered the original item on a Web3 news feed, the lesson is narrower than it looks: aggregation pipelines that translate and republish industrial filings without a domain check will keep producing errors of exactly this shape, and the correction will rarely arrive on its own.