The AI memory crunch is now a tax on everything you ship.

Memory is the story of 2026 so far. DRAM prices rose by roughly 90 percent in the first quarter against the last quarter of 2025, with some consumer modules up past 100 percent and SSDs climbing further still. This is not a demand spike from buyers. It is supply being pulled away from the rest of us.

Where the memory went

The three firms that make most of the world's memory, Samsung, SK Hynix and Micron, have pointed the bulk of their capacity at high-bandwidth memory for AI data centres. Estimates put the combined shift at close to nine tenths of production. HBM earns far more per wafer than the DDR5 in a laptop, so the ordinary stuff is being rationed. IDC expects the PC market to contract by about 11 percent this year as those costs reach the shelf.

You can see it in this month's launches

Lenovo's SteamOS Legion Go 2 arrives in June at 1,199 dollars, a handheld priced like a laptop. Valve has been unusually candid that the Steam Frame headset's price and timing are both in flux, and a memory-heavy standalone headset is exactly the kind of device this squeeze hits hardest. When a company that wants your money will not commit to a price, believe the supply chain, not the marketing.

Why a studio cares

• 01Bills of materials move under you. A device costed in 2025 can be a different margin in 2026 on memory alone. Re-cost late, not early.
• 02Cloud is memory too. Managed caches, in-memory stores and large instances track the same shortage. The cheapest performance win this year is using less RAM, not renting more.
• 03On-device budgets tighten. Shipping a feature that assumes generous headroom on a mid-range phone or a kiosk is a riskier bet than it was a year ago.

The calls we are making now

• 01Spec memory toward the top of a sensible range while it is in the build, not as a later upgrade. Adding RAM after the fact is the expensive path now.
• 02Profile for footprint, not just speed. We treat peak memory as a first-class budget in reviews, the way we already treat latency.
• 03Buy the machines the team needs this cycle rather than waiting for a better price. The earliest credible easing is late 2027, and most new capacity is already promised to AI.
• 04Design the degradation path. A system that still works on less memory is worth more than one that assumes plenty.

"

When the cheapest part on the board doubles in price, it stops being an afterthought and becomes an architecture decision.

None of this is a reason to panic, and it will not last forever. But it is a real constraint, and constraints are where the interesting engineering lives. The teams that treat memory as a budget this year, on the device and in the cloud, will ship things that still make sense when the bill arrives.