[{"data":1,"prerenderedAt":134},["ShallowReactive",2],{"\u002Fblog\u002Fall-the-ram-should-belong-to-him":3},{"id":4,"title":5,"author":6,"body":7,"date":122,"description":123,"extension":124,"meta":125,"navigation":126,"path":127,"readingTime":128,"seo":129,"stem":130,"tag":131,"updated":132,"__hash__":133},"blog\u002Fblog\u002Fall-the-ram-should-belong-to-him.md","Sam Altman wants all the RAM in the cloud","Madhumita S",{"type":8,"value":9,"toc":113},"minimark",[10,14,25,30,33,39,42,46,49,52,55,59,62,65,68,72,75,84,87,91,94,102,105],[11,12,13],"p",{},"OpenAI's Sam Altman thinks the memory in your computer is being wasted. On the Cheeky Pint podcast with Stripe cofounder John Collison, he called it \"quite terrible\" that people have \"a bunch of the world's limited supply of DRAM sitting in their own computer and being used 5% of the time\". The better arrangement, in his telling, is maximum utilization in a data center.",[11,15,16,17,24],{},"Gizmodo ",[18,19,23],"a",{"href":20,"rel":21},"https:\u002F\u002Fgizmodo.com\u002Fsam-altman-all-the-ram-should-belong-to-me-so-i-can-charge-people-to-use-it-2000833412",[22],"nofollow","wrote it up"," under the headline that all the RAM should belong to Sam Altman so he can charge people to use it. He did not say that sentence. He said the reasonable version, which is the one worth arguing with.",[26,27,29],"h2",{"id":28},"what-he-actually-said","What he actually said",[11,31,32],{},"Here is the full thought, so you can judge it yourself:",[34,35,36],"blockquote",{},[11,37,38],{},"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.",[11,40,41],{},"There are two claims stacked on top of each other. The first is about utilization: hardware that idles is waste, and waste is costly when the thing being wasted is scarce. The second is about who should hold the supply. The first is true enough to be boring. The second is where the argument stops working, and it is the one that produced the headline.",[26,43,45],{"id":44},"a-laptop-that-idles-95-of-the-time-has-already-been-paid-for","A laptop that idles 95% of the time has already been paid for",[11,47,48],{},"Utilization is a real problem when you own the machine. A server bought for peak load and left at 5% is burning capital every day it sits there, and the company that bought it feels that in its margins. That is an operations problem, and it is why cloud capacity gets consolidated.",[11,50,51],{},"A person's laptop is a different arrangement. It was bought once, at retail, by the person using it. The 95% of the time it spends close to idle is not a loss being carried by anyone else. There is no second buyer waiting for those free gigabytes, because there is no way to hand them over. Memory sits either soldered to the board or in a slot, wired to one memory controller, with a latency budget measured in nanoseconds. It is not a bucket you can dip a request into from another city.",[11,53,54],{},"So \"global allocation\" is doing more work in that sentence than it can carry. Moving work to the cloud improves utilization because the machines are bigger and shared, not because the idle time disappears. It moves the idle time somewhere else.",[26,56,58],{"id":57},"the-shortage-is-real-but-the-conclusion-does-not-follow","The shortage is real, but the conclusion does not follow",[11,60,61],{},"Gizmodo is right that DRAM has become brutally expensive, and right that data center demand is why. There is also a defensible version of his point: DRAM comes off the same wafers whether it ends up in a rack or a laptop, so capacity that goes into consumer modules is capacity that does not go into servers. That part is a genuine trade-off, and it is the strongest thing you can say for his position.",[11,63,64],{},"Now look at what that trade-off actually implies. It is an argument about how memory makers should spend their wafer starts, and that decision is already being made every quarter by whoever places the largest orders. It is not an argument for collecting the modules out of everybody's computer, which would not work anyway. The SO-DIMM in a laptop and the registered ECC modules in a server are different products, built to different specifications and validated for different platforms. The memory in a phone is soldered down. The allocation happens at the factory, months before you take the laptop home, and it is driven by purchase orders rather than by a planner deciding what would be efficient for the world.",[11,66,67],{},"That is the trick in the quote. It states a real constraint and then implies a conclusion that follows from OpenAI's balance sheet rather than from the constraint.",[26,69,71],{"id":70},"maximum-utilization-is-the-sellers-metric","Maximum utilization is the seller's metric",[11,73,74],{},"Every cloud provider cares about utilization. It is how capital turns into margin, and there is nothing dishonest about it. The problem is that \"maximum utilization\" always points toward more of the resource being in the speaker's racks, and the speaker is not a neutral party.",[11,76,77,78,83],{},"The direction of the ask gives it away. If the world is short of memory, the cheapest fix is to use less of it. What he proposes is routing more of it to the largest single buyer, which is also the buyer whose orders created the shortage in the first place. Gizmodo puts OpenAI's infrastructure rollout at ",[18,79,82],{"href":80,"rel":81},"https:\u002F\u002Ffinance.yahoo.com\u002Ftechnology\u002Fai\u002Farticles\u002Fopenai-raises-planned-ai-infrastructure-134830365.html",[22],"$750 billion through 2030"," and points out that the cost of that buildout is what is pushing up the price of DRAM, including in phones. The waste he is describing is real. The invoice for fixing it goes to everyone who is not OpenAI.",[11,85,86],{},"None of that makes him wrong on the facts. It means the sentence should be read as a statement of interest, in the plain sense of the word.",[26,88,90],{"id":89},"what-it-changes-on-an-ordinary-project","What it changes on an ordinary project",[11,92,93],{},"Memory is no longer a rounding error in a hardware budget. Gizmodo does not see 32GB of DDR5 falling below $450, so the dull optimizations are back in the conversation. Do not hold a whole result set in memory when a cursor will do. Paginate. Cache less, and measure what the cache was actually saving. Size the machine from a week of real traffic instead of a guess.",[11,95,96,97,101],{},"For AI features the choice is sharper. Running a small model locally is cheap, private by default and has no network round trip. A data center model is the only option once the model does not fit on the hardware. Clients ask us which way to go, and the answer depends on the data, the latency the workflow can tolerate, and what a call costs at the volume they expect. It is a normal architecture question with a different answer for a document search tool and a support bot. If you want the longer version of the cost side, our post on ",[18,98,100],{"href":99},"\u002Fblog\u002Fsoftware-development","how AI agents changed software delivery"," walks through where a feature's cost actually accumulates.",[11,103,104],{},"The practical version of his point is worth keeping. Right-size the machine, rent the capacity you only need at peak, and stop paying for hardware that sits idle. That advice is sound. It just lands differently when the person giving it is also selling the capacity.",[11,106,107,108,112],{},"Before you decide where an AI feature should run, spend an hour with your own numbers: the size of the model, the sensitivity of the data, the latency the workflow needs and the cost per call at your real volume. If you would rather have someone do that alongside you, ",[18,109,111],{"href":110},"\u002Fservices#devops","cloud architecture and infrastructure"," is part of our day job.",{"title":114,"searchDepth":115,"depth":115,"links":116},"",2,[117,118,119,120,121],{"id":28,"depth":115,"text":29},{"id":44,"depth":115,"text":45},{"id":57,"depth":115,"text":58},{"id":70,"depth":115,"text":71},{"id":89,"depth":115,"text":90},"2026-11-01","Altman says memory sitting idle in your computer is a waste of the world's scarce supply. The logic only holds if the data center belongs to him.","md",{},true,"\u002Fblog\u002Fall-the-ram-should-belong-to-him",6,{"title":5,"description":123},"blog\u002Fall-the-ram-should-belong-to-him","AI",null,"EP8MamTtDP69G5lSjBa_v_uSak4_n5481n8-rXjgC8M",1793520834971]