The selling started in Tokyo. Kioxia Holdings fell 13.3% on Tuesday to close at 15,180 yen, the sharpest one-day drop yet for a stock that had nearly quadrupled over the past year.
In Seoul, SK Hynix lost 10% and Samsung Electronics 7%, a fifth consecutive session of declines for two companies that together account for more than half of the Korean market’s value. The country’s volatility index rose 23% as the losses fed on each other. By the afternoon in New York, Micron was down about 5% and the semiconductor index had given back a month of gains.
One trade, three continents, and a question that has been deferred for two years: how much of the memory-chip boom reflects genuine shortage, and how much reflects a price umbrella that a Chinese supplier is about to close.
Two disclosures in the same week answered part of it.
The first came from Shanghai, where the president of ChangXin Memory Technologies told an industry conference that the company had made meaningful progress on a 4F² cell structure, a denser DRAM design that shrinks the area each memory cell occupies. The company said the approach could raise density by roughly a third and that a server product built on it would ship before the end of the year, according to people who attended the session. It disclosed few specifications, and mass-production yields remain unverified.
The second came from Sydney. Firmus, an Australian data-centre operator backed by Nvidia, withdrew an initial public offering that would have raised about 5 billion Australian dollars after its banks stopped taking bids at the marketed price of 11 Australian dollars a share. The company said in a statement that the proposed terms did not adequately reflect its business, a formulation that in practice means buyers would not pay them. Reuters reported that the failure dealt a blow to Australia’s shrinking listings market.
Read together, the two events describe a market that has begun to price the cost of the AI build-out rather than its volume alone. A dense-cell Chinese DRAM design threatens the supply side of the memory trade. A shelved data-centre float threatens the demand side, because the customers buying memory at today’s prices are the same companies that need capital markets to keep building.
The concentration makes the first threat easier to trace. Samsung, SK Hynix and Micron supply roughly 90% of the world’s DRAM, and the two Korean makers have directed their newest capacity toward high-bandwidth memory for AI accelerators, where demand has outrun supply and prices are set in long-term contracts. That left the ordinary server and consumer DRAM market as the part of the business that a new entrant could attack with a cheaper, denser design.
ChangXin has spent years behind the leaders on the most advanced products. It has also been the beneficiary of a domestic supply chain that the Chinese government has funded on the stated grounds of import substitution, at a time when American export controls have restricted the leaders’ access to Chinese customers. If it can deliver a server-grade DDR5 part in quantity, the pricing discipline that has kept memory margins unusually wide through this cycle gets harder to defend.
Analysts who follow the sector cautioned against reading a single conference presentation as a shipment plan. Nothing in the Shanghai disclosure settled how many wafers the company can yield, which customers have qualified the part, or what it will cost. History offers a warning in both directions: Chinese makers have repeatedly announced designs that took years to reach commercial volume, and they have also repeatedly arrived at lower prices.
The Firmus withdrawal carries a different kind of signal. The float was pitched as one of Australia’s largest in years and would have funded data-centre capacity whose economics depend on electricity contracts, land and the willingness of tenants to sign long leases. When investors balked at the marketed price, they were not voting against AI demand. They were declining to fund capacity at the returns the seller assumed.
That distinction matters for the memory trade. Hyperscale operators have funded their build-outs from cash flow and, increasingly, from debt. As the IMF noted in its financial stability report this week, the share of investment-grade bond issuance taken by hyperscalers has risen to about 10% from less than 1% since 2024. A market that will not absorb a data-centre equity float at the asking price is a market that will scrutinise the debt next.
What changed on Tuesday was less a verdict than a repricing of the same facts. The bull case for memory has rested on three assumptions: that AI capacity demand keeps compounding, that financing stays available, and that no credible competitor presses prices down. The first remains intact. The second was tested in Sydney. The third was tested in Shanghai.
The companies have not revised their guidance, and contract prices for the current quarter were set before either event. The test comes with the next round of negotiations, when buyers who have accepted higher prices for six consecutive quarters sit down with suppliers who now have something new to point at from the other side of the table. Whether they use it is a question neither side has to answer this week.
