On 29 October the Cheeky Pint podcast published an interview with Sam Altman, and one answer in it has travelled further than the rest. Asked why high memory prices were not fixing themselves, OpenAI's chief executive said that "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". Within a day, Sam Altman's RAM comments were being read as a case against owning a PC at all. If you searched for "sam altman pc ram comments", this is the context.
This piece sets out what he actually said, the memory shortage behind it, why many people think OpenAI is the wrong messenger, and what any of it means for a business buying laptops, servers or AI.
What Sam Altman actually said about PC RAM
The exchange comes about 29 minutes into the video, in a section the show labels "Compute". The interviewer, John Collison, observed that memory had become very expensive and asked why new suppliers were not flooding the market. Altman's first answer was about the memory makers, not about anyone's PC. For years, he said, a DRAM producer "would invest $20 billion of CapEx to make $19 billion", and after many near-bankruptcies "they all remember their history". Collison summed this up as makers being "conservative in their CapEx", and Altman agreed.
Then he turned to local AI. The full passage, from the show's own transcript:
"A lot of people want local inference. And somehow, they assume that the problem is solved now. Because, we are going to have open-source models. We're going to be able to run them on our computers. And so we're not going to need all the stuff in the data center. 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. And somehow everyone's... Many people's intuition is the opposite of that."
Three things are worth noting. He was talking about running AI models locally, not about gaming, spreadsheets or a console. The "5%" figure was given without a source. And he did not say anyone should be prevented from buying memory; he argued that, while supply is short, shared data centres use it better than idle personal machines.
The DRAM shortage of 2026 behind the remark
The shortage itself is not in dispute. The market researcher TrendForce said on 30 September that conventional DRAM contract prices would rise a further 10–15% quarter on quarter in the fourth quarter of 2026, and that server DRAM "is therefore expected to remain undersupplied in 4Q26". A later TrendForce news item put the fourth-quarter rise at about 20%, a slowdown "from approximately 80% in the second quarter and 40% in the third". The two figures measure slightly different things; the direction is the same.
The cause is where the memory goes. TrendForce's September bulletin summarises it in a few lines: "Suppliers favor HBM and server DRAM", "Consumer devices weaken; CSPs keep buying for AI", and long-term agreements "shrink negotiable supply as new fabs lag". HBM is the high-bandwidth memory stacked beside AI accelerators, and every wafer that goes to it is a wafer that does not become a laptop's RAM.
One maker went further. In December 2025 Micron announced it would exit the Crucial consumer business, "ending the sale of Crucial-branded memory and storage products", describing it as a reallocation toward faster-growing markets.
The effect has reached finished machines. In June the analyst firm IDC forecast that global PC shipments would fall 11.3% in 2026, with average selling prices up 17%, said it saw no relief "before the end of 2027", and warned that pricing "is unlikely to return to 2025 levels" even as capacity grows.
Why critics say OpenAI helped cause the memory shortage
The reaction was sharp, and much of it rested on one point: OpenAI is among the largest buyers of the memory it says is scarce.
On 1 October 2025 SK Group announced that SK hynix had signed a letter of intent with OpenAI to supply High Bandwidth Memory for the Stargate data centre programme, with production able to meet demand for "up to 900,000 DRAM wafers per month", which SK described as "more than double the current HBM industry capacity". Samsung signed a letter of intent too, as The Register reported. Some coverage, such as TechCrunch's, described the figure as chips rather than wafers; SK's own release says wafers.
Two details matter for a fair reading. These were letters of intent, and SK's release says "specific supply volumes and arrangements will be determined through further discussions". The widely repeated claim that this amounted to about 40% of world DRAM output is an estimate: The Deep Dive attributes it to unnamed analysts, and argues that the letters carried "no enforceable purchase obligation" while contract prices "surged in the aftermath". How much of the price rise those deals caused, against everyone else's AI spending, is not something anyone has measured.
The commentary that followed the podcast was less measured. Futurism read it as an argument that people should not own computers, and Kotaku noted that devices "from gaming consoles to iPhones" now cost "hundreds higher than they were even a year ago". Notebookcheck made the more useful criticism: the argument concerns local inference and leaves out the many other reasons people need memory in a PC.
Where Altman has a point, and where to be sceptical
The utilisation argument is sound as far as it goes. A server shared by thousands of users runs far busier than a laptop that sits idle overnight. That is why cloud computing exists, and it is true of memory as it is of processors.
He has a commercial interest. OpenAI sells AI that runs in its data centres. An argument that scarce memory belongs in the cloud is also an argument for OpenAI's business model, and he is making it while his company competes for the same supply.
"5% of the time" is not evidence. It may be a fair guess for some machines, but no figure was offered to support it, and memory in a working laptop is doing other jobs too.
His own diagnosis points elsewhere. In the same answer Altman blamed makers' caution about building fabs, and said "the invisible hand is failing right now". If supply is slow to grow, the fix is more capacity, not fewer PCs. The memory makers' profits are part of that story; this piece is about building and buying systems, not about anyone's shares.
The narrow reading is the useful one: memory will stay expensive for a while, the cloud and your own hardware now compete for the same chips, and the people making the case for one side often sell it.
What the DRAM shortage means for a business
Budget for higher hardware prices through 2027. IDC's forecast of no relief before the end of 2027, and prices that do not return to 2025 levels, is the planning assumption to use. A laptop or server replacement budgeted at last year's prices will come up short.
Buy on need, not on a hope of falling prices. IDC noted that commercial buyers already pulled purchases forward in early 2026. Waiting for a dip is a bet against the forecasts. If machines are failing or out of support, replace them; if they work, stretching their life by a year is now worth pricing out.
Specify memory once, properly. With memory the expensive part, it is tempting to buy the base configuration. For staff who will run local AI tools, or simply keep many applications open, an under-specified machine is a cost you pay every day for years.
Get server quotes early, with validity dates. TrendForce expects server DRAM to stay undersupplied this quarter, with long-term agreements soaking up supply. For on-premises servers, lead times and quote expiry matter as much as the list price.
Decide cloud or local on the task, not the headline. The shortage does not settle this either way. Cloud providers buy the same memory, and TrendForce says they "keep buying for AI".
| Cloud AI | Local AI (your laptops or servers) | |
|---|---|---|
| Privacy | Data goes to a provider; depends on its terms and retention | Data can stay on your own hardware |
| Cost shape | Pay per use; no hardware outlay | Paid up front, now at high memory prices |
| Utilisation | High, shared across many customers | Often low, as Altman says |
| Latency | Network round trip on every request | No network hop; limited by your hardware |
| Model quality | Access to the largest models | Limited to models that fit in your memory |
| Maintenance | The provider's job | Yours: updates, testing, security |
For many businesses the sensible design uses both: a cloud model for heavy or occasional work, and local or self-hosted models where data must not leave the building. If the larger question is whether AI spending is overheated, our earlier piece on what an AI bubble would mean for business covers how to plan either way.
The short version
Sam Altman did not say people should give up their PCs. He said that while DRAM is scarce, memory in busy shared data centres does more work than memory in idle personal machines, a point with some logic and an obvious commercial interest behind it, from a company whose own supply deals are part of the story. For a business, the practical consequence is unchanged by who is right: memory is expensive and will stay that way into 2027, so plan hardware budgets accordingly, buy when you need to, and choose between cloud and local AI on the task and the data, not on the argument of the week.