Geoffrey Vandiest
Memory prices went up more than four times in a year. Not because datacenters got built. In the US, 16 GW were promised for 2026 and 5 GW were under construction in February. Capacity under construction grew 18% in eighteen months. Memory contract prices grew more than 300%.
So what did drive it? Stock. The big cloud players reportedly placed open-ended orders at any price, signed three-to-five-year take-or-pay contracts and paid deposits. Micron alone holds $32 billion of them. OpenAI reserved up to 40% of world DRAM output for a project that is still mostly a drawing. Lenovo keeps 50% more memory than normal. The chip makers now ask buyers who their end customers are, because they know part of it is hoarding.
Stock has limits. It ties up capital, it fills warehouses, and the chips are a generation old by the time the building is finished. Nobody can stock memory indefinitely. Meanwhile the chip makers, who kept their capex at +14% and enjoyed 80% gross margins, are now adding cleanrooms for 2027. Higher prices bring more production, and that production will arrive when the hoarders have stopped buying. That is how every memory cycle has ended: 2008, minus 50%. 2011, minus 85%. 2023, NAND minus 39%.
The quarterly rises already tell the story: +93% in Q1, +60% in Q2, +15% in Q3, +12% forecast for Q4. The makers say the shortage lasts until 2028. The makers also sell memory.
Two nuances. The crash will show up in the spot market first, where you and I buy, because the multi-year contracts have price floors and the cloud giants can carry their stock longer than anyone else. And part of the rise is here to stay: wafers moved to HBM, which eats three to four times more silicon per gigabyte, so the floor will be higher than last time.
My bet: the hoarders will be the ones giving discounts. If your AI server can wait a year, let it wait.
First published on LinkedIn.