RAM Prices Are Up 400% Since 2024. Sam Altman calls local inference a waste.
Sam Altman calls local inference a waste of scarce resources. J.P. Morgan has DRAM up over 400% since 2024. The fabs moved capacity, they did not add it.
On October 29 Sam Altman was asked about running models on your own machine. His answer, as transcribed from the clip, was that local inference is "a terrible allocation of scarce resources."
Altman
"A lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need [the cloud]."
Set aside whether he is right about the engineering. He is partly right about the engineering. Look instead at what the thing he is telling you not to buy has done to its price.
400%
J.P. Morgan's figure for the rise in DRAM prices from the start of 2024 to the end of 2026.
3 to 5x
The margin on high-bandwidth memory for AI accelerators versus conventional consumer DRAM.
17%
Gartner's projected rise in average PC prices this year, driven by memory and storage.
You are not being outbid by other people who want to run Qwen on a laptop. You are being outbid by the people selling you the alternative.
Gergely Orosz said the quiet part
Orosz
"Sam Altman complaining that individuals (like me!) are purchasing RAM for our PCs wanting to run local inference (and do other stuff), because it gets in the way of him buying more RAM to sell inference at 80% margins… this is barking up the wrong tree"
The 80% is Orosz's characterisation, not a published figure, and it is the weakest part of his post. The strong part is the structural claim underneath it, and that one checks out.
There is no separate consumer memory industry sitting idle. The same wafers become either the DDR5 in your desktop or the HBM stacked on an accelerator, and one of those two buyers pays several times the margin. When Altman says scarce, he is describing a real constraint. He is also describing a constraint his own order book created.
The repricing did not arrive gently either. TrendForce forecast DRAM contract prices rising 90 to 95% quarter on quarter in the first quarter of this year alone, and Bernstein measured roughly 20% quarter on quarter in the third. Those are contract prices, the ones OEMs pay, which is why Lenovo, Dell, HP, Acer and ASUS have all warned of 15 to 20% increases on finished systems. The person buying a memory kit is at the back of that queue, paying retail for what is left after the accelerator orders are filled.
Where he is actually right
Per token, local inference is worse. A single GPU in a tower sits idle most of the day, runs a smaller model, and cannot batch requests across hundreds of users. A rack in a datacentre keeps its silicon hot and amortises it across everybody. If the only thing you are optimising is cost per token, the hyperscaler wins and it is not close.
That argument has one hole, and it is the same hole it had when the mainframe people used it. Efficiency is not the only thing anyone is buying.
The companies putting models on their own hardware are not doing it for price. They are doing it because the data cannot leave, or because a regulator wants to know where it sits, or because they refuse to build on a dependency whose price and terms are set by one vendor. None of those are token-cost decisions, and none of them get solved by a cheaper API.
What to do if you were planning on local
Three things changed this autumn and they point in different directions.
The hardware repriced, so redo the arithmetic. If your on-prem business case was built on 2024 memory prices it is now wrong by a multiple. Price the machine today before you quote the project, and remember that OEMs have already warned of 15 to 20% increases on whole systems this year.
The reason to go local did not change. If the driver was sovereignty, residency or vendor concentration, a price rise is a cost input, not a reversal. It makes the decision more expensive, not less correct. Say that out loud in the room, because somebody will try to use the price as a reason to abandon the requirement.
Watch who is telling you not to buy. Altman is making a sincere technical claim with which he also happens to agree commercially. That does not make it false. It does mean the claim deserves the same scrutiny you would give a cloud vendor telling you not to run your own servers, which is a conversation this industry has already had once.
Orosz bought the RAM anyway. So did enough people to show up in the pricing.
Scarce is a choice.

