AI data centers are gobbling up memory as fast as it's manufactured.
Vertigo3d/Getty Images
As companies push all their chips into the AI basket, GPU makers are shifting their resources to support the boom. The resulting memory shortage has had dire consequences for the consumer tech industry: Component prices are skyrocketing, supplies are dwindling, and everyday consumers are left holding the bag.
Known colloquially as "chipflation," the rise in GPU prices has vastly outpaced economy-wide inflation levels. According to a report by JP Morgan, by the end of 2026, dynamic random access memory (DRAM) prices will have risen by 400 percent since 2024. The Tom's Hardware chips tracker reflects this increase, finding that Nvidia's most powerful consumer GPU, the GeForce RTX 5090, currently retails for $4,920 above the $1,999 suggested price when it launched in January 2025. Foreign markets are also feeling the squeeze. In Japan, for instance, some gaming GPUs have risen roughly 41 percent above their American counterparts.
The knock-on effecs across the consumer electronics market has been shocking. According to that JP Morgan report, the Consumer Price Index for software and accessories has increased by nearly a quarter over the last two years. A 2026 report by The Economist, meanwhile, found similar rises in the prices of video game consoles and laptops, prompting an uncharacteristic fall in PC shipments. Opportunistic resellers have taken advantage on the resale market, where companies like GameStop market Sony's PlayStation 5 Pro for roughly $500 above MSRP.
Some expect the crisis to extend into the 2030s. Many chip producers already have sold much of their 2027 supplies, with some stating that they are now focusing their sales upon 2028 and beyond.
Inside the GPU shortage
IM Imagery/Shutterstock
At its core, the GPU shortage is a memory problem. Of course, peaks and valleys in chip supplies are nothing new. Already, the industry was shifting from DDR4 to DDR5 memory formats, stalling supplies as manufacturers transitioned to new production lines. But the rapid data center buildout has worsened the shortage, necessitating vast stores of high-bandwidth memory (HBM), a specialized form of DRAM with faster connections and more bandwidth. AI developers need HBM chips to both train and run their AI models, so HBM demand far outpaces current market supplies.
Unfortunately, HBM competes for the same wafer capacity as DRAM and other memory products. With resources limited, suppliers have steadily reoriented production toward advanced HBM chips, following not only demand but profit margins, selling HBM for roughly 30 percent more than traditional DRAM. To make matters worse, suppliers are transitioning to a more advanced generation of HBM, namely HBM4 and HBM4E, further limiting supplies as companies transition production.
As a Samsung executive told Reuters in September, HBM could account for nearly 30 percent of total DRAM production in 2027, up from 20 percent the year before. Some expect that percentage to jump to almost 40 percent by 2029. More broadly, AI data centers are projected to consume roughly 70 percent of global DRAM supplies this year.
A highly concentrated market further exacerbates the issue. Today, three companies account for roughly 90 percent of global DRAM production: Micron Technologies, Samsung Electronics, and SK Hynix. All three are expecting long-term shortages despite making major expansions of their production capacity. Moreover, the trio continue to pivot towards datacenters at the expense of consumer GPUs. Micron, for instance, ceased direct-to-consumer sales last year to dedicate resources toward AI production. Cue continued RAM price hikes.
Until 2030?
Kathrine Andi/Shutterstock
Industry leaders paint a worrisome picture. Micron CEO Sanjay Mehrotra said in a recent earnings call to expect supply to be "much tighter in calendar 2027 and 2028" than in 2026. SK Hynix CEO Kwak Noh-jung echoed this prediction, telling Reuters in July, "we forecast that next year will be the worst year in the industry's history from the supply perspective." SK Hynix's chairman Chey Tae-won, meanwhile, said at a Nvidia event that it will take "at least four to five years" to build adequate wafer supply, with production lagging 20 percent behind demand until at least 2030 (via Reuters).
Several major suppliers have already sold most of their production capacity for next year. Micron, for instance, has reportedly sold 75 percent of its 2027 HBM output. According to Mehrotra, via Yahoo Finance, these sales were struck at prices "much higher than 2026 prices." Samsung, meanwhile, has sold almost 80 percent of its 2027 memory output, according to Korean investment firm KB Securities, with roughly 70 percent tied up in five-year contracts.
Some are skeptical that all this will continue. Acer CEO Jason Chen, for instance, told DigiTimes prices could plateau next year, citing a potential influx of cheap Chinese alternatives. Others question whether the AI industry will generate enough growth to warrant memory suppliers' massive buildout. One South Korean publication caused memory stocks to tumble in October 2026 when it quoted experts predicting that a "flood of new supply" in 2029 would transform the memory market "into a red ocean," per a CNBC translation. Nvidia head Jensen Huang, though, waved off the possibility of a supply surplus, telling CNBC, "One of the things that's really great about memory is that you can always use more of it."
