If you haven't seen it, here's the clip. John Collison (of Stripe) asks Sam Altman (of OpenAI) on the Cheeky Pint podcast why the current high prices for memory aren't resulting in new supply. Shortly afterward Altman says, "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." He says it's used "5% of the time." Elon Musk chimes in to say "The people deserve to have their RAM too!" Gizmodo's headline was "All the RAM Should Belong to Me, So I Can Charge People to Use It." Reddit discussed it in terms of food. It's been covered, and I have nothing to add. (But I did enjoy the memes: How to trap Sam Altman, Collectivization of RAM, 64GB for Roller Coaster Tycoon)
To be clear, Sam wasn't saying that you can't own your own RAM. He was trying to explain why running open models at home isn't going to solve the memory shortage. I did wonder who was using the phrase "global allocation," though. In October 2025 OpenAI signed letters of intent with Samsung and SK Hynix for as many as 900,000 DRAM wafers a month. That's potentially 40% of global output by one estimate. OpenAI may never buy that much, but the biggest potential buyer in the market is telling everyone else where their memory should go. Such is life in 2026.
Projected global DRAM output: 2.25 million wafers a month
Who gets the memory
John Collison's question was how the solutions to high prices were coming along. "Bad," said Sam Altman. "I think the invisible hand is failing right now." He's right that the memory suppliers are reluctant to invest in new capacity - they've been through too many boom and bust cycles.
Supply will be slow to respond to demand, but it is a market. If you bought your RAM you can leave it idle all weekend if you want. If a studio decides to overspec the workstations it buys for its artists, that's its money. I don't want anyone making a decision about whose idle memory is OK, least of all the biggest buyer in the market.
I sell the other side of this argument
I agree with Sam Altman that idle hardware is a waste of money. We work really hard to make sure our customers aren't paying for idle hardware. I can't really complain about renting hardware, since most of the stuff we build runs on AWS or Google Cloud.
Here's an example. Recently on a major film production we provided the render capacity, the storage, and the workstations for the artists. For work like that cloud is the default these days (and it is the majority of what we sell) so it wasn't really a choice. We got the call when the cameras were already rolling. Four weeks later the environment was ready for the artists. Now you could try and do that with purchased hardware. Get the quotes, pull the trigger, then get the equipment delivered, racked, stacked, powered, cooled, and tested for a team distributed around the world. Good luck. There's no way. And remember, any film production is selling time.
You can run up a spectacular bill in the cloud. If you take what you have on-prem and recreate it in the cloud you've failed. And it will cost you an arm and a leg. Sam Altman didn't cite a source for his 5% figure, but a workstation that nobody is using is a line on somebody's bill whatever the number is. A cloud workstation is billed by the hour whether you are sitting at it or not. So on that film each workstation was shut down at the end of the artist's day. Fast storage is a lot more expensive than cold storage, so at weekends, when most of the team wasn't working, we dropped to the minimum, and the data just sat on cheaper tiers of storage until Monday.
We learned that lesson the hard way. In the early days of Gunpowder when we were helping customers move from on-prem to cloud we were mostly focused on getting the work to run in cloud at all. We weren't as focused on cost as we are now. We were spending $80,000 a month on storage for a small company. That wasn't feasible. We started to use vendors who could scale down as well as up and still give you a POSIX file system (Weka and Qumulo for example). On House of Parliament's Super Bowl work we used Weka to reduce the storage spend by 80% compared to more traditional cloud file and block storage.
Cloud is not the answer to everything (and increasingly may not be the answer in the future). There was a brief time when there was a genuine movement back to on-prem. That has slowed (buying hardware is so expensive!) but may start up again. A render node that is working flat out 365 days a year may well be cheaper on hardware you own (on-prem or in colo). You make that decision workload by workload, with the numbers in front of you.
When the GPUs aren't there
Sometimes scarcity forces you to be creative. When a client's team needed more render capacity than their own hardware could provide, the GPUs that we would have used were in short supply. And I'd been waiting for a real production to test rendering on CPUs. So that's what we did. We built a render farm out of cloud machines with no GPUs. We ran over 1,000 instances and gave the team 3x the throughput on peak multi-project loads.
This won't work for all workloads. If you have a GPU renderer and the scene fits in card memory you'll usually get a frame rendered faster than on a CPU. And a CPU render node still needs a lot of DRAM! So we got around the GPU shortage and ended up in the RAM shortage.
What we're responsible for is much more limited
Nobody's asked me for a view on the global allocation of RAM. Somebody picks up the phone and says "I want to film a movie" and it's our job (at Gunpowder) to make the infrastructure for that happen. That production has a release date. The artists have shots they need to deliver. So the compute needs to be available when they sit down at their desk. It needs to be appropriate for the workload. And the bill needs to make sense. In this market that means booking capacity well in advance (like you would crew or stage time), having a backup plan (because the part you wanted is already spoken for) and rechecking the numbers after you launch.
So if you've got a bid going out this quarter: list each workload, how many hours a day/week it will actually be running, what it needs to run, and who else is trying to buy that hardware. Then compare the total cost of owning that hardware against cloud rates (on-demand, committed, and spot - if you can take an interruption). Send me the list if you want a second opinion.
