the runoff
For forty years consumer volume paid for the leading edge and the serious work picked up what was left. That direction reversed, and memory prices are where you can watch it.
AlexNet was trained on two GeForce GTX 580s. Gaming cards. Krizhevsky used consumer hardware in 2012 because that was where the floating point was, and the floating point was there because millions of people who wanted better shadows in their shooters had spent a decade paying down the tooling.
That’s the real mechanism under every claim that chip progress trickles down. It never trickled. Consumer scale amortized the fab, the halo part became the midrange in eighteen months because the node moved and the line needed volume, and the frontier work bought the runoff at retail.
Nvidia closed fiscal 2026 with $193.7 billion in data center revenue against $16.0 billion in gaming. Twelve to one.
Whoever buys the volume sets the price. That used to be us.
Memory is the clean case, because memory is fungible. Same fabs, same wafers, and the wafer becomes an HBM stack or it becomes the DIMM in your desktop. Q1 2026 DRAM industry revenue rose 81% quarter over quarter. Q2 contract prices rose 58–63% for DRAM and 70–75% for NAND. No factory burned down. Suppliers pointed capacity at high-capacity RDIMMs for AI servers, and consumer parts got what came after.
The tell is Q3. Prices kept rising, but only 13–18%, and TrendForce says plainly why: consumer device makers can’t absorb any more. Supply did not improve. The price stopped climbing because the bottom of the market ran out of money. That isn’t a market clearing. That’s a market getting outbid and the result being filed as weak consumer demand.
2021 was the rehearsal. Miners outbid gamers for identical silicon with a better ROI story, retail price stopped meaning anything for two years, and everyone called it a shortage. It was an auction. It ended when the bid collapsed, not when the supply arrived.
Call it tickle-down chiponomics: the spend at the top is supposed to reach you eventually, and the part that reaches you first is the invoice.
Trickle-down in silicon was never generosity. It was a clock. You got a 3080 in the end because the process advanced and somebody had to sell you one. The clock is broken in one specific place: a rack of accelerators never becomes your part. There is no consumer descendant. It depreciates inside a building you can’t enter and gets resold into a market you’re not in.
The honest counter is that the trickle does arrive, just not as hardware. Cost per token has fallen by roughly an order of magnitude a year at fixed capability, and a hobbyist today can do work that was a funded lab’s budget in 2020. That’s real. It is the strongest version of the argument and it’s correct.
It’s also metered. You get the capability and you don’t get the substrate — what’s for sale is an hour of someone else’s allocation, priced by whoever holds the wafer contract. Krizhevsky went to a store and bought two cards and owned them. The equivalent move now is a credit card on file and a rate limit.
Supply-side had exactly one honest version in this industry: wait, and it gets cheap. It still does.
It just doesn’t come down anymore. You go up to it, by the hour.
Sources: TrendForce on 1Q26 DRAM revenue ↗ · TrendForce on 2Q26 contract prices ↗ · TrendForce on 3Q26 moderation ↗ · Memory price surge cools as consumers hit affordability limits, Tom’s Hardware ↗ · NVIDIA Q4 and fiscal 2026 results ↗