ChangXin Memory Technologies (CXMT, 688825.SH) released a voluntary disclosure on its earnings forecast for the first three quarters of 2026 on October 17. The company expects revenue of roughly 265 billion to 270 billion yuan for the period, with net profit attributable to shareholders of about 131 billion to 132.5 billion yuan. For comparison, CXMT posted revenue of 32.084 billion yuan and an attributable net loss of 5.28 billion yuan over the same period in 2025.

The company disclosed something else on the same day: on October 16, its board received a letter from chairman Zhu Yiming proposing a share buyback.

The first thing to note is that the company itself names AI computing power as the primary driver of the swing. From the filing: "Since the second half of 2025, AI computing power demand has grown rapidly, the global DRAM industry has moved into an upcycle, products are in short supply, and prices have risen sharply."

How Big Is the One-Year Turnaround

Set the numbers side by side: attributable net profit went from a 5.28 billion yuan loss a year earlier to 131 billion–132.5 billion yuan for the first three quarters, which the filing describes as an increase of 136.28 billion to 137.78 billion yuan. The fact that the increase exceeds the profit itself is exactly what a jump from loss to massive profit looks like.

The revenue line is just as steep: from 32.084 billion yuan to 265 billion–270 billion yuan, an increase of 232.916 billion to 237.916 billion yuan by the company's own accounting. Roughly speaking, revenue for the first three quarters came in at more than eight times the year-earlier figure.

Price increases are not the only reason the filing lists. The company cites "steady progress in capacity build-out, product R&D and process technology improvements … with scale effects continuing to show through," along with "ongoing optimization of operational management, active market expansion, and deeper coordination with upstream and downstream partners," and states plainly that "capacity utilization has remained at a high level."

The phrase worth dwelling on is "scale effects." Memory is a textbook capital-intensive business, and once capacity utilization climbs, thinner unit costs amplify both the revenue and the profit line at the same time. That is why the company broke this point out separately.

Buyback Disclosed the Same Day: A Chairman's Proposal, Not a Board Resolution

The timing of the second filing sits right against the earnings forecast. CXMT's board received Zhu Yiming's buyback proposal letter on October 16, 2026, covering a plan to use the company's own or self-raised funds to repurchase a portion of its issued A-shares through centralized bidding on the Shanghai Stock Exchange trading system.

The stated rationale, per the filing: "Based on confidence in the company's sustained and stable future development and recognition of its long-term value, in order to protect the interests of the company's investors, particularly minority investors, and to strengthen investor confidence."

According to Sina Finance, the buyback would total no less than 1 billion yuan and no more than 2 billion yuan, with the repurchased shares earmarked for employee equity incentives or an employee stock ownership plan, or for a reduction in registered capital. The execution window runs 12 months from the date the board or shareholders' meeting approves the plan, and the buyback can only begin after the stock has been listed for six months. The repurchase price is capped at 150% of the average trading price over the 30 trading days preceding the board resolution approving the plan. The filing also confirms that Zhu had not traded company shares in the six months before making the proposal, has no plans to increase or reduce his holdings during the buyback period, and will vote in favor of the proposal at the board. The buyback still requires approval and remains uncertain.

Yicai, citing industry observers, notes that buybacks by STAR Market-listed companies have historically clustered in the 200 million to 800 million yuan range, putting CXMT's plan in the top tier by size. A 2 billion yuan buyback would cover close to 1% of the current free float, and with unrestricted tradable shares accounting for just 6.63% of the company's share capital at this stage, the "shrinking-float effect and the sentiment floor it provides will be amplified." That assessment is market analysis, not part of the filing.

How the AI Computing Power Bill Travels Down the Chain

What gives this news its weight is that it lays out a complete chain in the open: rising AI computing power demand leads to rising memory demand, which leaves DRAM in short supply and pushes prices up, which in turn shows up as earnings at memory makers.

For the past two years, the AI computing power narrative has centered on GPUs and the model layer, with the "memory wall" treated mostly as a technical talking point. CXMT's filing puts the industrial payoff on the table: for the first time, tight computing power supply registers on a memory maker's quarterly accounts at the scale of trillion-yuan revenue and hundred-billion-yuan profit.

It is the same logic Suanmiao Technology bet on when it announced funding in October. That company's 3D TokenPU architecture is built to attack the "memory wall," using hybrid bonding to stack memory wafers vertically onto logic wafers and shorten the data movement path. The two events landing within two weeks of each other suggests the inference-side memory bottleneck has moved from a technical question to an industry consensus.

The filing also makes the capacity side of the equation explicit, listing "capacity utilization has remained at a high level" alongside "scale effects continuing to show through" in the same section on drivers. That means the thing to watch next is not just price, but whether supply can keep up.