Sam Altman Calls Local AI a "Terrible" Use of Scarce DRAM: What He Said, the 5% Claim, and What Is Unverified

Sam Altman has said that people running AI models on their own computers are making "quite terrible" use of the world's scarce DRAM. The remark comes from an episode of the Cheeky Pint podcast with John Collison, dated October 29, 2026, and it spread quickly because it lands in the middle of a memory shortage that is raising prices for PCs, consoles and phones. The quote is on the record. The supporting number, that this memory is used "5% of the time," comes with no data. This post separates the two.

For the supply background, see our coverage of Michael Dell calling the memory shortage a four-year problem and Apple's Ternus on the memory crisis.

TL;DR: claimed vs. verified

What exactly did Altman say?

According to the Cheeky Pint episode page, Altman said "A lot of people want local inference. And somehow, they assume that the problem is solved now." He then argued that running models on personal computers is globally inefficient because "it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM" sitting in individual machines, and that "you want maximum utilization in the cloud if you're trying to deal with this very supply constrained environment." The 5% figure refers to how often that memory is in use for this purpose.

Secondary coverage, including TheGamer's write-up, repeats the same lines and pairs them with the broader "RAMpocalypse" framing in which AI data centers absorb global memory supply. Wording differs slightly between outlets ("quite terrible" versus "terrible"); the episode page uses "quite terrible."

Why does Altman think the market is stuck?

In the same stretch of conversation, Altman gave an account of why high memory prices have not produced more supply. He said that for a very long time a DRAM producer "would invest $20 billion of CapEx to make $19 billion." Memory makers, he argued, have lived through boom-and-bust cycles and poor long-term returns, so they stay conservative about capital spending even though they may now each earn around $100 billion in free cash flow in a year. Asked why prices have not yet pulled in supply, he said "the invisible hand is failing right now."

That is a recognizable pattern: producers burned by past gluts are slow to add capacity. It is also consistent with what we reported from ASML on EUV supply and the AI chip supply chain roundup. The disagreement is not about whether DRAM is tight. It is about who should get it.

A green chip carried along a track, illustrating the DRAM shortage and AI chip supply chain

Is the 5% number believable?

Treat it as a rhetorical estimate. Three problems:

1. It measures the wrong thing. Personal-computer RAM serves the operating system, browser, games, creative software and development tools, not only AI inference. Idle capacity for one workload is not wasted memory.
2. There is no denominator. The remark does not say whether 5% means time with a model loaded, tokens generated, or memory pages in active use. Without a definition, the figure cannot be checked.
3. Cloud utilization is not 100% either. Data-center fleets also have idle GPUs and memory, especially across bursty demand. "Maximum utilization in the cloud" is a goal, not a measured advantage.

TheGamer's write-up makes a related point in its own voice, arguing that the same logic would make the RAM in a gaming PC a waste. We did not find a measurement of average consumer-RAM utilization for local inference, so the honest status is unverified.

What is the counterargument?

The strongest rebuttals are practical, not about the arithmetic:

• Privacy and control. Local inference keeps prompts and files on the device. For many users that is the product, not an inefficiency.
• Latency and offline use. On-device models work without a connection and avoid round trips.
• No per-token bill. Once you own the hardware, marginal cost is electricity.
• Pooling cuts both ways. Concentrating memory in the cloud raises utilization but also concentrates supply and bargaining power with a few providers, one of which is Altman's company.
• Incentives. The speaker leads a company that buys cloud memory at scale. That does not make the argument wrong, but a reader should note who benefits from it.

Note also that running larger models on consumer devices would raise, not lower, consumer memory demand, which is the opposite of the direction Altman is arguing for.

What else did Altman say that matters?

The same episode, per its notes, contains several claims about the economics of compute:

• Cost per gigawatt. Asked whether dollars per gigawatt of leading-edge AI capacity would rise or fall, Altman said "That should go down." He used about $70 billion as a rough five-year total cost of ownership per gigawatt and said that at $20 billion he would be much happier, with energy, chip and supply-chain costs all needing to fall.
• Test-time compute. A few years ago you could not usefully spend more than about a dollar of compute on a problem. Chain-of-thought changed that, enabling large parallel agent teams.
• A math proof's cost. He estimated the Navier-Stokes result at on the order of $10 million, possibly a bit more, and suggested it could be reproduced at about a hundredth of that a year later. That is his estimate, not an audited figure.
• Internal policy. OpenAI allocates substantial compute to internal productivity, avoids rewarding raw token usage, and has no token leaderboards.
• Own hardware. He mentioned OpenAI's custom chip, its work on building its own data centers, and a robotics program to help build data centers, framed as longer-term.

These figures are his statements from a podcast, so treat the numbers as directional. For how OpenAI frames its own pacing and safety work, see Altman on safety cases before big RL runs.

What does this mean if you build or buy hardware?

• No policy changed. There is no announced restriction on local inference, and the remarks do not alter what your machine can run.
• Expect continued price pressure on memory-heavy configurations. The shortage is the context, and executives from several companies expect it to last. If you need a local-model workstation, memory is the expensive part.
• Right-size models. Smaller, quantized or mixture-of-experts models that fit in less memory are the practical response to scarce DRAM, whichever side of this argument you take.
• Know the trade. Cloud inference trades privacy and per-token cost for utilization and scale. Local does the reverse. Decide on your data and workload, not on a slogan.

Where does the argument fit in the memory squeeze?

The memory shortage has been building through the year, and each side of this argument reads it differently. Producers say AI servers consume high-bandwidth memory and conventional DRAM at a rate they cannot match quickly, because new fabs take years and extreme-ultraviolet tools are themselves constrained. Device makers say that squeezes the consumer products they sell, raising bills of materials for phones, consoles and PCs. Cloud builders say that scarce parts should go where utilization is highest.

Altman's remark is the cloud-builder reading stated bluntly. The consumer-side reading, as TheGamer frames it, is that a computer with 32 GB of RAM is a general tool, and that arguing its memory is wasted would apply equally to any gaming PC. Neither framing is a measurement. What would settle it is data: average memory occupancy of consumer machines running local models, average memory utilization of cloud inference fleets, and the share of DRAM output going to each. None of those has been published in the coverage we found.

There is also a timing point. Altman said the invisible hand is failing now, which implies he expects supply to respond eventually. If DRAM makers do add capacity, the premise of the argument weakens. If they do not, the pressure on consumer hardware prices continues through 2027 and beyond, which is what Dell has said to expect.

How to check claims like this

When a leader gives a utilization figure on a podcast, ask four questions: what is being measured, over what period, on what fleet, and is there a published dataset? If none of those is answered, quote the statement as a claim. That is how we handled it here: the quote is verified because the episode contains it; the 5% is not, because nothing supports it.

Related reading

• Michael Dell: the memory shortage is a four-year problem
• Apple's Ternus on the memory crisis
• ASML: AI demand exceeds EUV supply
• AI chip supply chain roundup, October 2026
• Samsung, TSMC and NVHBM base die: claimed vs. verified
• Sam Altman on safety cases

Quotes are from the Cheeky Pint episode page and press coverage as of October 30, 2026. We did not hear the full audio.