Sam Altman Says Your Laptop Is Hoarding the World's Memory

Sam Altman told Stripe cofounder John Collison that local AI ties up scarce DRAM in machines that use it about 5% of the time, and that the memory belongs in data centers running flat out. The math is real, but it leaves out why companies run models locally, and it comes from the executive whose memory deals helped start the shortage.

OpenAI CEO Sam Altman says the people counting on local AI to escape the memory shortage have it backwards. Running models on personal computers is, "from a global allocation perspective, actually, it's quite terrible," he told Stripe cofounder John Collison on the Cheeky Pint podcast, which published the episode and a full transcript Thursday.

Altman's case is that DRAM is in short supply, a laptop uses the memory it holds only a small fraction of the time, and a data center can keep the same memory busy around the clock. In a shortage, he argued, the memory should go where it gets used.

At State of AI, we think the utilization point is correct and worth taking seriously, including inside your own company. We also think it is the wrong frame for deciding where enterprise AI should run, and Altman is not a neutral party on the question. OpenAI sells cloud inference, and its own memory deals helped set off the shortage he is describing.

Altman says the market isn't fixing the shortage

The remark came up when Collison asked why high memory prices hadn't yet pulled in new supply. Altman said the fix was going badly and that the invisible hand isn't working in this market.

His explanation started with the chipmakers. For decades, he said, a DRAM producer would spend about $20 billion on capex to bring in about $19 billion, through repeated booms and busts. Now each of the big ones is headed for roughly $100 billion in free cash flow in a year. Altman said they remember coming close to bankruptcy after past overbuilding, so they are being careful about new fabs.

Then he turned to local inference. Many people, he said, assume open-weight models running on their own computers will reduce the need for data centers. He argued the opposite: memory sitting in personal machines, used around 5% of the time, is a bad allocation when supply is this tight, and the world should want maximum utilization in the cloud instead.

The claim is narrower than it may sound. Altman didn't say local AI should go away, and he framed it as a problem of this particular shortage. But the conclusion points one way: in a tight market, the efficient place for your model is a shared server.

The utilization math is real

The argument is the same one that built the public cloud. When hardware is expensive and demand comes in bursts, shared machines beat private ones. A server answering thousands of requests keeps its memory full, while a laptop holding a large model mostly holds it while its owner answers email.

The hardware is expensive right now. DRAM prices were up 171% year on year by May, Park Place Technologies CTO Chris Carreiro wrote in Data Centre Review. On Wednesday, The Wall Street Journal's Rolfe Winkler reported that hyperscaler demand has quintupled flash memory prices and shifted bargaining power from Apple to suppliers like Kioxia.

Local AI also needs a lot of exactly what's scarce. NVIDIA's new RTX Spark laptops offer up to 128GB of unified memory, which NVIDIA says is enough to run a 120-billion-parameter model on the machine, according to Acer. HP's versions come in 32GB, 64GB and 128GB configurations. Across millions of buyers, that adds up to a real claim on the same supply.

OpenAI helped create the scarcity it's describing

Altman is describing a market his company reshaped. In October 2025, he signed letters of intent in Seoul with Samsung and SK Hynix to supply memory for OpenAI's Stargate data centers, TechCrunch reported. Both chipmakers put Stargate's projected demand at 900,000 DRAM wafers a month, with no timeline, according to Data Center Dynamics. The two companies hold about 70% of the global DRAM market, Reuters reported.

Analysts put that volume at roughly 40% of global DRAM supply, and the agreements carry no obligation to buy, The Deep Dive noted. Contract prices jumped anyway. Two months later, Micron announced it would exit its Crucial consumer business to improve supply for its larger strategic customers.

So the company that reserved a large share of future memory output now says the memory left in consumers' hands is being used inefficiently, and that the efficient fix is renting AI from companies like OpenAI. Contrary Research has asked whether the threat of on-device inference gives AI labs a reason to keep memory tight. That's still an open question, and we aren't treating it as settled. Altman's comments make it a fair one to ask.

Laptops didn't cause the shortage

The shortage runs the other way from how Altman's framing suggests. Manufacturers moved wafer capacity toward high-margin HBM for AI accelerators, and that's what squeezed conventional DRAM, Altium's analysis of the crunch explains. Laptops are absorbing a shortage that data center demand created.

Memory in a laptop also isn't as idle as Altman's 5% figure implies. It runs the browser, the IDE and the video call, and a local model borrows memory the company already paid for. By the same logic, every phone, workstation and car is a misallocation too.

Utilization also leaves out most of what enterprises run local inference for. A model on the device doesn't send contracts, patient records or source code to a third party, which is why some regulated teams run models locally in the first place. It also works offline and doesn't get rate-limited at peak hours.

Then there's cost. A laptop is a capital purchase, while cloud inference is a variable bill set by a seller in the supply-constrained market Altman just described. Data centers aren't fully utilized either, since they're built for peak demand, and Altman offered a figure for idle laptops without one for the alternative.

A trillion AIs on a few landlords' servers

The same conversation makes a different argument about who should hold power. Explaining why he's against a single superintelligence, Altman told Collison he would rather see a trillion slightly superhuman AIs join society than one AI with a trillion IQ. He said society is robust because many independent actors keep each other in check, and he wants AI to work the same way.

That's a decentralized picture of intelligence. His picture of the hardware is centralized: put the world's memory in a few clouds and keep it fully used. A trillion AIs that all run on servers owned by the same few companies are only as independent as those companies let them be.

OpenAI's own plans point the same direction. In the interview, Altman said the company has built its own chip, is learning to build its own data centers and has a robotics program aimed at building more of them. OpenAI is also making consumer hardware after buying Jony Ive's io for $6.4 billion. How much of that device's AI runs on the device is worth asking about.

Our take

Altman is right that memory is the constraint of this AI cycle and that shared hardware uses it better than personal hardware. Enterprise leaders should take four things from that, and none of them require handing the hardware to OpenAI.

Apply his utilization math inside your own walls. Putting 128GB in every engineer's laptop at today's memory prices is the inefficiency Altman is describing. A pooled inference server, on-prem or in a private cloud, keeps the memory busy and keeps the data in your control. Altman's argument favors shared hardware. It doesn't require the hardware to be OpenAI's.

Place workloads by what they need, not by a vendor's efficiency argument. Sensitive data and steady, predictable jobs suit hardware you control. Bursty work that needs frontier models suits the cloud. Most companies will end up running both.

Treat scarcity as a pricing risk in your AI contracts. A seller arguing that supply is tight is also telling you where prices are headed. Lock in capacity and pricing where you can, and keep a working open-weight model running locally so you have an alternative at renewal.

Hold off on fleet-wide AI laptop refreshes unless the use case is proven. If OpenAI's memory agreements don't turn into real purchases, an analyst cited by The Deep Dive expects prices to ease, and memory bought at the peak will look expensive. Price the exit in too. ERI CEO John Shegerian told us recently that resale value and material recovery now decide how much of an AI hardware purchase ever comes back.

The argument over local versus cloud AI was never mainly about utilization. It's about who owns the hardware your models run on, and so who sets the price, sees the data and decides what runs. Altman made a good case for sharing that hardware. Whether you share it with your own teams or with OpenAI is still your call.