Articles & Essays

MEMORY IS EXPENSIVE

GPU and DRAM prices have climbed 2-3x in a year because AI data centers now absorb 70% of the world's memory output - and the people who can least afford to wait it out are the ones running compute as a business, not a hobby.

GPU and DRAM prices have climbed 2-3x in a year because AI data centers now absorb 70% of the world's memory output - and the people who can least afford to wait it out are the ones running compute as a business, not a hobby.

The RTX 5090 launched at $1,999 in January 2026. Street prices hit roughly $5,000 by September 2026. Desktop DDR5 modules averaged $15.27 per gigabyte by mid-August 2026, and TrendForce's July forecast called for conventional DRAM contract prices to climb another 13-18% quarter-over-quarter in Q3 2026 alone. NAND and SSD pricing is moving the same direction: 2TB NVMe drives that cost $120-150 a year ago now run $300-480.

The cause isn't gaming demand and it isn't manufacturers being greedy. AI data centers are projected to absorb roughly 70% of the world's memory chip output in 2026, up from 20-30% in 2022. That's not a speculative bubble showing up in a spreadsheet - it's real capital, building real data centers, buying real silicon at a rate the memory supply chain didn't plan capacity for. The fabs aren't rationing on purpose. They're running flat out against demand that grew faster than anyone forecast a few years ago.

This isn't a single dated event with a headline that fades in a week - it's a structural shift, which is exactly why it's easy to undercover. There's no press release to react to, just a slow-motion repricing of every component that touches memory. The specific numbers here - $5,000 for a 5090, $15.27/GB DDR5 - will be stale within months, and probably wrong in the direction of worse before they're wrong in the direction of better. What won't go stale on the same timeline is the mechanism: AI data center demand is now large enough to out-compete consumer and prosumer demand for the same physical fabs, and nothing about that mechanism resolves on a predictable schedule.

The take running through TechPowerUp, Tom's Hardware, and TechRadar is straightforward and not wrong: building or upgrading a PC right now is brutal, GPU and RAM prices are genuinely insane, and AI data centers eating the world's memory supply is the actual, verifiable cause. Anyone frustrated at this year's build costs has every right to be.

But "it's brutal for gamers" undersells who's actually being punished, because it treats this as a hobby-cost story. Gamers can wait out a GPU generation - skip a build cycle, play the backlog, the hobby survives a pause with no real damage. A streamer or small studio running local inference, video encode, and a print-farm control rig on one box cannot do that. That hardware is a cost of doing business, not a discretionary purchase, and the operators who need it most are disproportionately the ones without the balance sheet to absorb a 2.5x price increase. The people getting priced out of the hardware are, in large part, the exact operators the "AI will democratize compute" story was supposed to help.

I use an RTX 5090 in the main computer for Red Warden Studios. It is front loaded to do just about everything that needs raw computer power for mathematical reasoning. I bought it mid-2025 because of the goals I have personally set for myself and looking at today’s prices – I don’t regret it. Just about every single division touches it: MythrilWorks for game development and art. Bastion for software creation and LLM training. RedWardenCDR uses it for playing games and live streaming.

I'm not selling mine into this market, and if you're sitting on similar hardware right now, you probably shouldn't either - not while data-center demand is still climbing and there's no clear signal of when supply catches up.

If my card died tomorrow, at this point its hard to justify spending that much money for just the graphics card. I have several cards in the background I can use instead; one being an RTX 3080. If I truly needed as much horsepower behind a 5090, I could always invest in a systems integrator like Starforge Systems.

There's a real tension underneath this, and it's worth naming plainly: being genuinely excited about what open, local models make possible, while watching the actual hardware needed to run them drift further out of reach for anyone who isn't a hyperscaler. That's not an argument against AI, and it shouldn't be read as one. It's an argument that "AI will democratize compute" is mostly a claim about software - open weights, cheaper inference algorithms, smaller models doing more with less - running headlong into a hardware market currently doing the opposite. If you build on local models, self-host anything, or run a compute-heavy small business of any kind, the memory famine isn't background noise anymore. It's a line item you have to plan around now, the same way a print shop plans around resin and filament costs.

The GPU and DRAM shortage will end eventually - fabs catch up, AI capex cools, or both, on some timeline nobody can honestly call right now. Until then, the useful math isn't "here's how bad it is." It's simpler than that: don't sell hardware you already own into a market this distorted, and don't budget your next build as if 2024 prices are coming back, because there's no evidence they are.