Server RAM doubled in 2026 — self-hosting AI now costs a lot more
DDR5 server RAM is up 116% in a year, Q3 contracts lost their price caps, and SK Hynix says the crunch lasts past 2030 — renting won't dodge it.
Copy markdown
The number that moved: server RAM doubled
DDR5 64GB ECC RDIMMs are up 116% year-on-year. A 256GB server config that ran ~$800–1,200 in early 2025 now runs ~$2,000–3,500, and Counterpoint expects that to double again by year-end. If your margins ride on memory, that's your new baseline.
Q3 contracts just lost their ceiling
Samsung is pushing Q3 DRAM and LPDDR contract prices up 20%+ — after roughly 90% in Q1 and 50–60% in Q2 — and SK Hynix has reportedly scrapped the price caps in its long-term supply deals. The number above isn't done climbing.
It's AI eating the wafers — and it won't ease soon
HBM for AI accelerators now consumes ~23% of global DRAM wafers, up from 8% in 2024, with data centers absorbing ~70% of memory output. SK Hynix's CEO says the shortage runs past 2030; analysts see no return to 2025 pricing before 2027.
The self-host math changed
Self-hosting an open-weight model was the hedge against API surge pricing. But the memory bill for a serious inference box — let alone a rig big enough for a ~1.4TB model like Kimi K3 — is now 2–3x last year's. The escape hatch still works; it's just no longer cheap.
What to do this week
Lock memory-heavy cloud reservations now instead of running on-demand, quote new self-host builds against H2-2026 prices rather than 2025 numbers, and re-run the build-vs-buy-API math before you commit. Relief isn't priced in until 2027 at the earliest.