Gartner says memory prices climb 130% this year and AI is eating 70% of supply

Around 130 percent. That is Gartner's combined forecast for DRAM and SSD price increases by the end of 2026, and the firm has given it a name: memflation. If you have been waiting for the answer to when RAM prices go down in 2026, that is the answer, and it is not the one anyone wanted.

TrendForce has the near-term detail. Conventional DRAM contract prices are up 58 to 63 percent quarter over quarter. NAND flash is up 70 to 75 percent over the same period.

Where the memory is going

AI data centres will consume roughly 70 percent of global memory output in 2026. In 2022 that figure was 20 to 30 percent.

That single shift explains the rest. Manufacturing capacity has moved towards high bandwidth memory because that is where the margin lives, and HBM wafers are not client DRAM wafers. When hyperscalers lock in NAND supply ahead of time, manufacturers stop producing the client grade SSDs that go into ordinary laptops. Vendors then either pay a premium for TLC or quietly ship you QLC instead and say nothing about it on the spec sheet.

Why your laptop got worse without getting cheaper

The substitution is the part most buyers miss. A drive advertised at the same capacity and the same price can be a different class of flash than it was eighteen months ago, with lower endurance and slower sustained writes. Nothing on the box changes.

The same pressure runs through system memory. Anyone sizing a machine to run AI models locally is buying into the worst memory market in a decade, and the capacity tiers that make local inference worthwhile are exactly the ones under most pressure.

When RAM prices actually go down

The timeline is where people get it wrong. New fabs from Micron, SK Hynix and Samsung come online around 2027 to 2028. A large share of that capacity is already committed to AI and enterprise customers before a single wafer ships.

That puts the earliest credible relief window in late 2027. Anyone holding off on an upgrade until prices normalise is planning to wait roughly fifteen months, on a component that is currently appreciating.

So: if you need the memory, buying now genuinely beats waiting. A 32GB DDR5 kit bought this month costs less than the same kit next quarter on every forecast currently published. If you do not need it, there is no scenario where 2027 hardware is cheaper in real terms, so buy for the machine you have rather than the one you are hoping to build.

The part that stings

This is the clearest case yet of AI infrastructure costs landing directly on people who are not buying AI. The buildout that produces compute fabric funding rounds and national sovereignty programmes is paid for at the top by hyperscalers and at the bottom by whoever is replacing a dead SSD in a five year old laptop.

The number to watch is HBM allocation as a share of total wafer starts. When that stops climbing, client memory gets its capacity back. Nothing published so far says it is close.

Some links in this article are affiliate links. We may earn a small commission at no extra cost to you.