Altman and the 5 percent: is local AI a waste of DRAM?

On the podcast Cheeky Pint (published 29 October 2026), Sam Altman calls it a poor use of the world's scarce DRAM when that memory sits in private computers that use it 5 percent of the time. He gives no source for the figure. This post shows the exact wording, the price picture according to other sources, and what the argument means for an in-house AI stack.

Local inference means a language model runs on computers the company owns or rents, and requests and data stay on its own network. The alternative is sending each request to a provider's cloud, where you pay per request or per token and the data leaves the building. With local inference you buy the hardware once; with the cloud you keep paying for compute time.

Altman runs OpenAI, which sells compute time from data centres. Anyone weighing up a machine of their own for AI should know this sentence and what it does and does not establish.

What this post is based on

The source is a podcast conversation, available as a video and as a transcript. The host is John Collison, co-founder of Stripe. Episode 39 of Cheeky Pint came out on 29 October 2026 and runs 52:45 (Apple Podcasts lists PT52M45S, YouTube 3,165 seconds). We read the transcript on Cheeky Pint's Substack and did not listen to the episode. The video is on YouTube and the episode is also on Apple Podcasts. Everything attributed to Altman here comes from that conversation.

What Altman says

In the "Compute" chapter (from 00:25:17), Collison raises the high memory prices. Altman says many people want local inference and assume that settles things: there will be open models, they will run on your own machines, and the data centre will no longer be needed. Then comes the passage in question. The transcript reads:

"But from a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer. And, being used … 5% of the time. You want maximum utilization in the cloud if you're trying to deal with this very supply constrained environment."

He does not say where the 5 percent comes from, and Collison does not ask. We treat the figure as an estimate.

In the same chapter Altman says the invisible hand is failing at the moment. In his account, memory makers long invested $20 billion to earn $19 billion and are therefore cautious about new fabs. That is his explanation; the conversation offers no measurement behind it.

Near the end, in the chapter "Token pricing is fake" (from 00:49:27), the topic is something other than prices below cost. Altman calls tokens "a kind of fake unit" because buyers must choose between a small model at 10 cents per million tokens and a large one at $10, and must also guess how long the chain of thought will run. He calls that cognitive overhead without value. On whether prices cover costs, he says nothing in that passage.

The price picture according to other sources

Altman describes the shortage; he does not measure it. The figures below come from third parties and are not re-measured by us.

(source: TrendForce News, 20 Oct 2026, citing a Commercial Times report)

Quantity	Value	Unit	As of	Source (not measured by us)
DRAM contract price vs prior year	171.8	percent	Q3 2025, reported 4 Nov 2025	Tom's Hardware, citing CTEE
DDR5 kit 2×16 GB, 6000 MT/s (Corsair Vengeance RGB)	183 vs 91	US dollars	Newegg 4 Nov 2025 vs PCPartPicker July 2025	Tom's Hardware
DRAM price vs prior year	171	percent	article of 13 May 2026	Data Centre Review, opinion piece by the CTO of Park Place Technologies
Contract price, rise vs previous quarter	approx. 80 / 40 / 20	percent	Q2 / Q3 2026, Q4 2026 expected	TrendForce News, 20 Oct 2026, citing Commercial Times

The numbers show a real shortage, but they are secondhand: Tom's Hardware cites CTEE, TrendForce News cites a Commercial Times report. We have not checked whether both mean the same source, and we did not read the originals. The 171 percent in the Data Centre Review piece has no measurement date and may be the same underlying figure, so it does not count here as a second independent measurement. All values are US or Taiwan market data. We did not collect prices for Austrian or German purchasing.

On the cause, Data Centre Review writes that the largest manufacturers are redirecting limited production capacity from standard DRAM to high-bandwidth memory (HBM) for AI accelerators. It expects further price rises and longer lead times through the end of 2026 more than relief. TrendForce News describes a slower rise for Q4 2026, citing a higher comparison base, weaker end-market purchasing power and price ceilings in long-term agreements.

OpenAI sells compute time from data centres. That is part of the standpoint behind Altman's remark. (Quelle: OpenAI / Wikimedia, CC BY-SA 4.0)

Where the argument stops

The arithmetic is easy to follow: a computer that sits idle 95 percent of the time uses its memory poorly. Altman is arguing about how hardware is distributed across the world. A single company decides by other questions.

For whoever has bought a machine, utilization is a cost factor they control themselves. Cloud compute, by contrast, is bought again with every request. Which side is cheaper depends on what a request costs on your own machine over its life and what it would cost at the provider. Altman does not address that. This is our own reading of the situation.

Several points never appear in his argument. Regarding data sovereignty, a model in a data centre sees your input before it answers. Response time depends on the line to the outside. And whoever buys compute from a provider reveals what they are computing. Depending on the data, these are exclusion criteria or line items in the calculation. The utilization argument does not dispute them; it just does not cover them.

That memory can be a bottleneck for local inference shows in a measurement on our own host kipc5090: how DRAM bandwidth slows a host with four RTX 5090s. How token prices and GPU rental drift apart is covered in a16z figures on falling token prices and expensive GPU rental. More posts are collected under local AI and GPU hardware. If you are planning an inference stack of your own, our services around in-house AI are the place to start.

In practice: measure your own utilization by having the stack log requests and tokens per day, and set that number against the cloud bill before adopting Altman's 5 percent for your own case.

Correction of 5 November 2026

The first version contained errors that we have corrected here. The "Token pricing is fake" chapter was misrepresented: Altman criticizes tokens as a unit of measure and says nothing there about prices below cost. The 171.8 percent comes from contract prices reported by CTEE, not from a "Typically AI price index". A flash-memory figure from a Wall Street Journal report (behind a paywall) and a Samsung profit report could not be checked at the source, so we removed them. Statements about our own token-cost tracking had no measurement behind them and are deleted. New is the TrendForce News forecast of quarterly price rises.