
There's a story doing the rounds, and it's a good one: America's data centres ordered mountains of RAM, then half the data centres didn't get built, and all that memory is about to flood the market. Prices crash back to sanity. Everyone with a home lab cheers.
We went looking for the flood. Half the story holds up. The flood isn't here, and nothing in the market says it's coming this year.
The half that's true
The cancellations are real, and they're big. Estimates from earlier this year put roughly half of the US AI data centres planned for 2026 as cancelled or delayed, with a 7 GW capacity gap opening under a buildout valued around $650 billion US. The causes are unglamorous: power grid interconnection queues running five to seven years, local opposition hardening into this year's moratorium wave of state and local pauses, and plain construction bottlenecks. Some analysts put the slippage at 30 to 50 per cent of planned capacity sliding to 2027 or 2028, and even the optimists concede the grid queues.
So the buildings didn't happen. The question is what happened to the chips ordered for them.
The half that isn't
If cancelled projects were dumping pre-bought RAM, you'd see it in spot prices first. You don't. A 64GB DDR5 kit jumped 485 per cent to $1,118 US, about $1,610 AU, by August. TrendForce has conventional DRAM contract prices up 13 to 18 per cent quarter on quarter for Q3 2026. And the three memory giants, Micron, Samsung and SK Hynix, are posting some of the most spectacular profits in semiconductor history, as of this week.
That is not what a flood looks like. A flood looks like the opposite.
Why the water stayed away
The reasons aren't mysterious once you look. Supply discipline is the big one: the memory makers learned from every previous crash cycle, and 2026 has been described as unprecedented on that front. HBM is another: high-bandwidth memory for AI chips now eats a fifth of the big three's wafer input and rising, so even fabs running flat out produce less of the ordinary stuff. And delayed isn't dead. Hyperscalers mostly didn't have warehouses full of chips sitting idle; they had allocations, which get moved to the sites that did get built. The demand slid right rather than vanishing.
Worth saying plainly: even the cancellation numbers have a sceptic. SemiAnalysis has pushed back on the "half cancelled" framing as exaggerated, noting the hyperscalers' self-builds keep rolling. Some of this is a scheduling story, not a demand collapse.
What would actually crack it
The spot-contract spread is the early warning. Spot prices have been surging while contract prices lag behind, and that gap is where a turn would show up first. CXMT is the second watch item: the Chinese challenger the incumbents are now openly worried about, building capacity while the giants milk record margins. Third, whether the demand marked "delayed to 2027" actually returns. If it doesn't, that's not a delay. That's a cancellation with better manners.
Why we care
This corner of the internet has started watching local models and the affordable hardware that could run them. A 128GB unified-memory box that runs 70B-class models currently lands just under $5,000 AU, and RAM is most of the reason why. If the flood ever does arrive, that box gets cheap in a hurry and the local AI case reopens. Until the data says otherwise, though, the shortage is the story.
The rumour was half right. Half is more than most rumours manage.