Key points
- Micron (MU) is down 33% and SanDisk (SNDK) 46% since June 25. CoreWeave and Vertiv fell about half as far.
- DRAM contract prices never fell. The quarterly increase went from 93-98% in Q1 to 58-63% in Q2 to a forecast 13-18% in Q3.
- Monday's bounce lifted CoreWeave 15% and Nebius 12%. Micron was still red and Seagate fell 6%.
- CXMT listed in Shanghai on July 27. Micron's three worst days of the drawdown were July 27, 28 and 29.
- Microsoft put $25B of its $190B capex guide on memory costs. Amazon raised its own to $220B.
Memory chip stocks opened Monday down about 4%, fell as much as 8% inside the first half hour, and spent the rest of the morning climbing part of the way out. The rest of the artificial intelligence complex went straight up. Micron (MU) traded as low as $770.15 in the first minutes and was at $810.87 by 11:34 a.m. Eastern, still down 1.5% on the day. SanDisk (SNDK) fell to $1,121.33 and clawed back to a 2.9% gain. Seagate (STX) was down 5.9% and Western Digital (WDC) 4.9%.
Over the same hours CoreWeave (CRWV) rose 15%, Nebius (NBIS) 12%, Meta (META) 7%, Vertiv (VRT) 6%, Oracle (ORCL) 6% and Amazon (AMZN) 5%. The Nasdaq 100 was up 1.3%. The VanEck Semiconductor ETF (SMH) was flat, and the reason a semiconductor fund is flat on a day like this is memory.
The six weeks behind that gap are worth laying out properly, because the popular version of this story is wrong. It is not that the AI trade has been rising while memory sank. Almost all of it sank.
| Stock | Since June 25 | Monday, 11:34 a.m. |
|---|---|---|
| SanDisk (SNDK) | -46.5% | +2.9% |
| Micron (MU) | -33.2% | -1.5% |
| Western Digital (WDC) | -23.3% | -4.9% |
| Vertiv (VRT) | -21.5% | +5.8% |
| Seagate (STX) | -21.4% | -5.9% |
| CoreWeave (CRWV) | -16.2% | +15.3% |
| Oracle (ORCL) | -9.9% | +5.8% |
| Nasdaq 100 (QQQ) | -2.7% | +1.3% |
| Nvidia (NVDA) | +5.8% | +3.2% |
Two things come out of that table. The high-beta AI names all de-rated together from late June, so a good deal of what happened to memory was the same air coming out of the same balloon. And memory fell about twice as far as the neoclouds and the power names, which is the part that needs its own explanation.
Demand did not crack. The shortage is still there, and the people closest to it keep saying so.
The prices are still going up. That is the problem.
Memory is a commodity, and commodity producers are not valued on the level of the price. They are valued on the direction of the next move. TrendForce, which tracks contract prices quarter by quarter, has that number falling off a cliff even as the price itself keeps setting records.
Conventional DRAM contract prices rose 93% to 98% in the first quarter of 2026. In the second quarter they rose 58% to 63%. For the third quarter TrendForce forecasts 13% to 18%.
Read that as a sequence rather than three separate facts. Prices are still climbing, and the rate at which they climb has fallen by a factor of five in six months. A memory maker whose selling price rises 95% in a quarter is printing money it did not have to earn. One whose price rises 15% is running a normal good business. The stock adjusts to the second version long before the first version stops being profitable, which is why the group topped in late June while every headline about demand kept getting better.
TrendForce is blunt about where the deceleration comes from, and it is not servers. Server DRAM stays undersupplied in the third quarter and server shipments hold up through 2027. The give is in phones and PCs, where "record-high contract prices mean customers from consumer markets, such as PCs and smartphones, are reaching their affordability limit." Add the arithmetic of a high base and the increases have to slow.
Then July happened
Three things landed inside five weeks, and none of them was about whether the world wants memory.
On July 27, ChangXin Memory Technologies listed on the Shanghai STAR Market and closed its first session up 465.82%, raising 57.92 billion yuan, about $8.6 billion. Its prospectus puts its 2025 share of the global DRAM market at 7.67%. We covered how quickly US funds found a way to own it. Micron's three worst days of the entire drawdown were July 27, 28 and 29, when it fell from $920.95 to $739.00.
On Monday, Reuters reported that CXMT is planning a second Beijing plant, in Yizhuang, and is in early-stage talks on state funding. That is the headline memory opened down on. A newly capitalized Chinese producer announcing more capacity one week after a blowout listing is the specific thing that ends a memory cycle, and the market knows the history.
Second, the biggest customers stopped absorbing quietly. Apple chief executive Tim Cook, on the company's June-quarter earnings call, said Apple had "reluctantly raised prices" because "we're in what I would characterize as a 100-year flood on the memory pricing, with exponential increases in memory prices." Asked about sourcing, he went further. "Primarily the DRAM market has three suppliers. Obviously if there were more suppliers, that would be good, and it would help us on the supply side and perhaps the pricing side," he said. "We're evaluating all options."
That last line is aimed at a negotiating table rather than at the public. Micron chief executive Sanjay Mehrotra said on June 30 that certain customers had led industry-wide price talks and pushed prices down. Apple has been willing to qualify Chinese memory, and CXMT going public gives that option a share price.
Third, the hyperscalers priced the bill in public. Microsoft (MSFT) guided to $190 billion of 2026 capital spending against a consensus near $152 billion, and finance chief Amy Hood attributed $25 billion of it to memory and component costs. Amazon raised its own number to $220 billion, and Andy Jassy named memory prices as the reason.
Those two disclosures are the cleanest statement of the problem. The memory windfall is a line item on the buyer's income statement, the buyers have now put a number on it, and the buyers are the ones with the balance sheets. Microsoft is up 38% since June 25 while its own capex guide went up, because the market reads that spending as buying future revenue. Memory's price increase is the same money viewed from the other end, and it arrives as something the customer is actively working to reverse.
Monday's economic data cuts the other way
The Institute for Supply Management reported at 10 a.m. Eastern that its manufacturing index came in at 55.6% for July, against expectations near 54 and June's 53.3%. It is the highest reading since May 2022. New orders were 56.7%, production jumped to 58.5% from 52.2%, and the employment index hit 52.8%, in expansion for the first time in 33 months.
Memory turned up hard in the half hour after the release. Micron went from $787.57 at 10 a.m. to $816 by 10:30, and Seagate from $791 to $819. So the print did help. But note what it means. Memory is the most cyclical product in semiconductors, more levered to a manufacturing upturn than logic chips or networking gear. An ISM at a four-year high is a straightforwardly good number for Micron, and Micron was still red in late morning while the neoclouds ran 15%.
That is the tell. If memory's drag were macro, the best manufacturing number since 2022 would have gone a long way toward fixing it.
The rest of Monday is the Federal Reserve selloff unwinding. The Nasdaq fell six straight sessions into July 29, when the Fed held rates and the Dow dropped more than 1,100 points on worries the central bank was behind on inflation. It has now risen three straight, starting with a 2.8% jump on July 30 behind Microsoft's earnings. The AI names that fell hardest for macro reasons in the six red days are snapping back hardest in the three green ones. Memory is snapping back least, because a good part of why it fell was never macro.
What would change it
The bull case does not require prices to reaccelerate. It requires the deceleration to stop where TrendForce currently has it rather than running through zero into outright declines. SK Hynix (SKHY) chief executive Kwak Noh-Jung expects the shortage to run past 2030 and calls 2027 the worst year of it. Its Seoul-listed shares still fell almost 9% on Monday, for reasons that had more to do with profit taking after a record session than with anything new about supply.
The things to watch are narrow. Whether CXMT's Beijing funding talks turn into committed capacity, and on what timeline. Whether TrendForce's fourth-quarter number holds above zero. Whether Apple signs second-half DRAM contracts with the three incumbents or moves volume to a Chinese supplier. And whether hyperscaler capex guides keep rising, because as long as Microsoft and Amazon are raising budgets specifically because memory costs more, the demand argument is settled and only the split of the profit is in question.
Six weeks in, the market has not decided that AI is over. It has decided it would rather own the companies spending the money than the ones collecting it.
Sources
- TrendForce contract price forecasts for 1Q26, 2Q26 and 3Q26, published June 1, March 31 and July 3, 2026.
- Reuters exclusive on CXMT's second Beijing plant and funding talks, August 3, 2026.
- Apple's June-quarter earnings call for Tim Cook's remarks; CNBC for Microsoft's $190B and Amazon's $220B capex guides.
- Institute for Supply Management, July 2026 Manufacturing PMI report, August 3, 2026.
- Price data from exchange records, June 25 closes through 11:34 a.m. Eastern, August 3, 2026.
