Sam Altman:
A lot of people want local inference, and somehow they assume that the problem is solved now because we 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, like, have a bunch of the world’s limited supply of DRAM sitting in their own computer and being used five percent of the time.
Like you want maximum utilisation in the cloud if you’re trying to deal with this very supply-constrained environment,
And somehow many people’s intuition is the opposite of that
TL;DR Guy selling product: “Here’s why owning your own product instead of renting it from me is a terrible idea”
No… he actually is right. A local machine will sit idle most if the time.
A data center can balance loads globally.
In terms of resources… data centers ARE cheaper.
However … there are other reasons to use local interference that might be more important than added cost
Totally agree. It’s really about wealth distribution and who controls the resources. Altman is just trying to steer the conversation toward efficiency instead.
Well yes, or data security or whatever.
But on a purely costs level such big data centers are cheaper… especially since they certainly getting better discounts than anyone else.
You can aggregate spare LLM capacity in a collaborative p2p fashion. But my reasons would be to use spare PV capacity to run an uncensored model during daytime with full privacy.
Maybe but nothing will beat a dedicated data center in terms of costs.
Starting with purchasing prices etc etc.
And what happens if a rogue Ai hits a p2p network might be a funny read a few centuries later…
If it was properly built.
That doesn’t include the garbage data centers being built currently, wasting local municipal water for cooling and running huge temporary generators for power.
https://newjerseymonitor.com/2026/09/22/nj-vineland-data-center-fined/
Which costs? It’s a PC. The power is free, since spare PV capacity.
The rogue AI will be the P2P network.
No, his analogy is poor for many reasons. One is that people using local models are likely using that computer for other things as well, it is a resource that does many things unlike their slop farms. Also, DRAM wasn’t a “scarce resource” until they devoured the industry with their AI ponzi scheme.
Only very few people have enough processing power to run a proper model.
Going cheap almost always ends in a more expensive repair bill down the line.
Likewise, doing the most convenient thing to satisfy one’s own desires is probably the most ill-advised course of action. Mainly, because it makes other people take the most convenient course of action to remove the other person from being a problem in their lives altogether.
Sure… locking your customer in is always a good business option.
But the point remains… for any given demand a big data center always is cheaper.
Cheapest doesn’t mean best for all uses cases.
But even then I’d rather rent capacity from a hyperscaler than pay him.