AI data-center demand for high-bandwidth memory is crowding out conventional DRAM used in PCs and consumer graphics cards, and chipmakers already have sold much of next year's output, Engadget reported on Nov. 7¹.

The piece, by Alec Hively, frames the GPU crunch as a memory problem rather than a shortage of GPU logic chips¹. GPU makers have shifted capacity toward AI servers, it said, while consumer component prices have risen faster than overall inflation, a pattern the industry has nicknamed chipflation¹.

Engadget cited a JP Morgan research note projecting that DRAM prices will have risen 400 percent from 2024 levels by the end of 2026¹. A Tom's Hardware price tracker, as described in the article, put Nvidia's GeForce RTX 5090 about $4,920 above the $1,999 suggested price from its January 2025 launch¹. The story said some gaming GPUs in Japan were selling about 41 percent above U.S. prices¹.

Knock-on effects in consumer electronics include a nearly 25 percent rise over two years in a software-and-accessories consumer price index, according to the same JP Morgan report as summarized by Engadget¹. A 2026 Economist graphic, also cited, pointed to higher console and laptop prices and a drop in PC shipments¹. Resale listings, including GameStop marketing of Sony's PlayStation 5 Pro about $500 above MSRP, were presented as another sign of tight supply¹.

The shortage sits on top of an industry shift from DDR4 to DDR5, which already slowed output as factories changed lines¹. Data-center buildouts then added demand for HBM, a stacked DRAM used to train and run large AI models¹. HBM uses the same wafer capacity as other DRAM products and, Engadget said, sells for roughly 30 percent more, so suppliers have reallocated toward it¹.

A further squeeze comes from the move to HBM4 and HBM4E, which again takes capacity offline during the transition¹. Engadget cited a Samsung executive, via Reuters in September, saying HBM could be nearly 30 percent of industry DRAM capacity in 2027, up from 20 percent the prior year, with some forecasts near 40 percent by 2029¹. AI data centers are projected to consume about 70 percent of global DRAM this year, the article said¹.

Three companies — Micron, Samsung Electronics and SK Hynix — account for about 90 percent of DRAM production in Engadget's account¹. All three are expanding plants but still expect long shortages and are tilting toward data-center customers¹. Micron stopped direct-to-consumer sales last year to focus on AI production, the story said¹.

On an earnings call, Micron CEO Sanjay Mehrotra said supply would be much tighter in calendar 2027 and 2028 than in 2026, according to prepared remarks linked by Engadget¹. SK Hynix CEO Kwak Noh-jung told Reuters in July that the following year could be the worst on record from a supply standpoint¹. SK Group chairman Chey Tae-won, speaking at an Nvidia event, said building adequate wafer supply would take at least four to five years and that production could lag demand by about 20 percent until at least 2030, Engadget reported via Reuters¹.

Micron has sold about 75 percent of its 2027 HBM output at prices higher than 2026, Mehrotra said, as relayed by Yahoo Finance in the Engadget piece¹. Samsung has sold almost 80 percent of 2027 memory output, with about 70 percent in five-year contracts, according to KB Securities as cited there¹.

Acer CEO Jason Chen told DigiTimes, as cited by Engadget, that prices could plateau next year if cheaper Chinese supply arrives¹. A South Korean report in October 2026 that a flood of supply in 2029 could turn memory into a crowded, low-margin market briefly knocked memory stocks, Engadget said, citing a CNBC translation¹. Nvidia CEO Jensen Huang told CNBC that extra memory can always be used, dismissing the idea of a lasting surplus¹. What remains unknown is how fast new fabs and HBM4 lines will actually add consumer DRAM, and whether 2029 capacity will overshoot demand¹.

Sources

1. EngadgetThere is no end in sight for the GPU supply problem8:45 PM

Read the original at Engadget ↗