---
title: "RAM Prices Are Up 400% Since 2024"
url: https://nativefirst.ai/blog/ram-400-percent-altman-local-inference/
description: "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."
author: John Tan
category: "AI Infrastructure"
published: 2026-11-02
updated: 2026-11-02
publisher: nativefirst.ai
format: Markdown copy of the web page at the url above, same text.
---

# 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.

John Tan · October 31, 2026 · 4 min read

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."

> "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 **\[the cloud\]**."
>
> **Sam Altman** · Relayed by @morqon · October 29, 2026

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

> "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"
>
> **Gergely Orosz** · X · October 31, 2026

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.

**Figure 1. Why your RAM and his RAM are the same RAM · The fabs did not add capacity. They moved it.**

*Chart: A margin comparison between two uses of the same fab capacity. Conventional consumer DRAM is shown as a baseline bar at one times margin. High-bandwidth memory for AI accelerators is shown as a bar three to five times longer, with the upper end dashed. The note below reads that Samsung, SK Hynix and Micron shifted capacity toward high-bandwidth memory, that SK Hynix has said its DRAM, NAND and HBM output is essentially sold out for all of 2026, and that a consumer buying a memory kit is bidding against an accelerator order at several times the margin.*

Margin multiple as reported by memory-industry analysts, captured 2 November 2026. Figure by nativefirst.

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](https://nativefirst.ai/blog/on-prem-ai-european-companies/) 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.**

[John Tan ↗](https://www.linkedin.com/in/johntans/)

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources

1 Sam Altman's remarks on local inference, 29 October 2026, as relayed and transcribed by @morqon on X: "sam: local inference is a terrible allocation of scarce resources" followed by the quoted passage beginning "a lot of people want local inference". This is a third party's transcription of an interview clip, not an official transcript, and the bracketed closing words are an inference from a sentence the relay cut off.

2 Gergely Orosz on X, 31 October 2026, quoting the above: "WTH: 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% figure is Orosz's characterisation and is not sourced to any published margin disclosure.

3 J.P. Morgan Global Research on the AI-driven memory shortage: DRAM prices projected to have risen more than 400% from the start of 2024 to the end of 2026. The window runs to year end, so the figure includes a forecast tail rather than being fully realised as of this writing.

4 Gartner's projection that combined DRAM and SSD prices climb 130% for the full year, pushing average PC prices up 17% versus 2025, and OEM warnings from Lenovo, Dell, HP, Acer and ASUS of 15 to 20% system price increases for 2026. Reported figures, reached through secondary coverage rather than Gartner's own publication.

5 The 3 to 5x margin advantage of high-bandwidth memory over conventional consumer DRAM, and SK Hynix's statement that its DRAM, NAND and HBM capacity is essentially sold out for all of 2026. Memory-industry analyst figures reported in secondary coverage, captured 2 November 2026; TrendForce separately forecast DRAM contract prices rising 90 to 95% quarter on quarter in Q1 2026, while Bernstein measured roughly 20% quarter on quarter in Q3.
