A GeForce RTX 5090 that launched at $1,999 now sells for $4,920 above that price, according to the Tom’s Hardware chip tracker. That is not a typo, and it is the clearest picture yet of what AI demand has done to consumer hardware. Gamers are the visible casualties. The real story is that memory makers have found a better customer and aren’t looking back.
The shortage was reported by Engadget this week, and the numbers are bleak. JP Morgan expects DRAM prices to be up 400 percent by the end of 2026 versus 2024. In Japan, some gaming GPUs cost about 41 percent more than in the US. Console and laptop prices are climbing too, and PC shipments are falling, which almost never happens in a market this size.
Why memory is the choke point
This is a memory problem more than a GPU problem. AI developers need high-bandwidth memory (HBM) to train and run models, and HBM is made on the same wafer capacity as ordinary DRAM. Suppliers sell it for roughly 30 percent more, so they’re shifting production toward it. Simple incentives.
The numbers show how far that has gone. A Samsung executive told Reuters that HBM could reach nearly 30 percent of DRAM output in 2027, up from 20 percent the year before. AI data centers are projected to consume about 70 percent of global DRAM this year. The move to HBM4 and HBM4E also eats capacity while lines are retooled, and the earlier DDR4 to DDR5 transition had already tightened supply.
Concentration makes it worse. Micron, Samsung and SK Hynix make about 90 percent of the world’s DRAM. All three are expanding, and all three expect shortages anyway. Micron stopped direct-to-consumer sales last year to put more resources into AI.
Sold out through 2027
The executives are not hedging. Micron CEO Sanjay Mehrotra said supply will be “much tighter in calendar 2027 and 2028” than in 2026. SK Hynix CEO Kwak Noh-jung called next year potentially the worst in the industry’s history for supply. Chairman Chey Tae-won said it will take at least four to five years to build enough wafer capacity, with production 20 percent behind demand until at least 2030.
Contracts back that up. Micron has reportedly sold 75 percent of its 2027 HBM output at prices well above 2026 levels. Samsung has sold nearly 80 percent of its 2027 memory, with about 70 percent locked into five-year deals. When supply is committed that far out, consumer buyers are fighting over scraps.
The skeptics
Not everyone buys the long-shortage story. Acer CEO Jason Chen thinks prices could plateau next year if cheap Chinese memory arrives. A South Korean outlet rattled memory stocks in October by quoting experts who expect a “flood of new supply” in 2029 that turns the market into a “red ocean.” Jensen Huang dismissed the idea of a glut, saying you can always use more memory.
Huang has an obvious interest in that view. Still, the skeptics have history on their side, because memory is a notoriously cyclical business and big buildouts have ended in oversupply before. The difference now is the five-year contracts, which protect suppliers if demand cools.
Does it matter?
Yes, and not just for gamers. If you’re a founder budgeting for hardware, or a developer planning a local-inference rig, assume current prices are the floor through 2027. Cloud GPU capacity will probably stay expensive too, since the same memory feeds it. Buyers who can wait may get relief around 2029 if the glut predictions hold. Everyone else should plan for chipflation as the normal state of things.
