Tom's Hardware reports that Altman says data centers offer better utilization of RAM.
Tom's Hardware reports that OpenAI CEO Sam Altman called it 'quite terrible' for a PC to have RAM, saying data centers offer better utilization.
397.9K
9 Sources, first seen ago
#1
#1
@1littlecoderAfter listening to this, I disassembled my computer, detached my RAM and donated it to OpenAI! I can't believe, I was so stupid and naive to think the DRAM inside my computer mattered. It didn't matter because OpenAI is working for the entire world and the compute concentrated on the data centers matter. Thanks @sama and cheeky pint for opening my eyes and contributing to global supply chain instead of gating it!
@fengwan217Local AI looks virtuous. Altman calls it a DRAM tax. I used to nod along whenever someone said the endgame was open-source models on your own machine — no megacampuses, no GPU landlords, just a model humming in the living room. Then Sam Altman sat down with John Collison on Cheeky Pint and cut that story open. Plenty of people, he said, want local inference. They assume open-source on personal computers means we will not need data-center hardware. His pushback was not taste. It was scarcity. From a global allocation perspective, it is quite terrible to have everybody hold a chunk of the world's limited DRAM supply in their own computer — sitting there, used maybe 5% of the time. That 5% is the part that stuck. In a supply-constrained world you want maximum utilization in the cloud. Most guts pull the other way: keep the silicon at home, feel sovereign, call it freedom. Altman's frame is colder. Idle memory on a million desks is not liberty; it is a tax on the same chips everyone else needs to train and serve models at scale. Then he turned to the people who actually pour the wafers. For a long time, memory makers would sink something like $20 billion of CapEx to earn about $19 billion — boom, bust, almost never a great business across a full cycle. Until right now. These firms are each set to make something like $100 billion of free cash flow in a year. And CapEx still stays conservative. They remember nearly going bankrupt from overbuilding fabs. Altman's line landed flat: the invisible hand is failing right now. Cash floods in; the fabs do not race ahead as fast as demand wants. That scarcity is why OpenAI's prep list reads the way it does — its own chip aimed at intelligence per watt, learning to build data centers, robotics pointed at constructing those halls, even the phrase "self-replicating data centers." The cost frame he used was blunt: roughly $70 billion per gigawatt over a five-year TCO, and they would be happier closer to $20 billion. Local inference does not shrink that bill if it also strands DRAM at a 5% duty cycle. So here is the open question I cannot shake. If the scarce resource is memory, not vibes about where the model lives — who still wants silicon idle on their desk once that tax shows up in the price of everything else? #AI #DRAM #OpenAI
@NikTekYou're not able to afford GPU, RAM and Storage because Sam Altman thinks that its quite terrible to have DRAM sitting in your computer@1littlecoderAfter listening to this, I disassembled my computer, detached my RAM and donated it to OpenAI! I can't believe, I was so stupid and naive to think the DRAM inside my computer mattered. It didn't matter because OpenAI is working for the entire world and the compute concentrated on the data centers matter. Thanks @sama and cheeky pint for opening my eyes and contributing to global supply chain instead of gating it!@fengwan217Local AI looks virtuous. Altman calls it a DRAM tax. I used to nod along whenever someone said the endgame was open-source models on your own machine — no megacampuses, no GPU landlords, just a model humming in the living room. Then Sam Altman sat down with John Collison on Cheeky Pint and cut that story open. Plenty of people, he said, want local inference. They assume open-source on personal computers means we will not need data-center hardware. His pushback was not taste. It was scarcity. From a global allocation perspective, it is quite terrible to have everybody hold a chunk of the world's limited DRAM supply in their own computer — sitting there, used maybe 5% of the time. That 5% is the part that stuck. In a supply-constrained world you want maximum utilization in the cloud. Most guts pull the other way: keep the silicon at home, feel sovereign, call it freedom. Altman's frame is colder. Idle memory on a million desks is not liberty; it is a tax on the same chips everyone else needs to train and serve models at scale. Then he turned to the people who actually pour the wafers. For a long time, memory makers would sink something like $20 billion of CapEx to earn about $19 billion — boom, bust, almost never a great business across a full cycle. Until right now. These firms are each set to make something like $100 billion of free cash flow in a year. And CapEx still stays conservative. They remember nearly going bankrupt from overbuilding fabs. Altman's line landed flat: the invisible hand is failing right now. Cash floods in; the fabs do not race ahead as fast as demand wants. That scarcity is why OpenAI's prep list reads the way it does — its own chip aimed at intelligence per watt, learning to build data centers, robotics pointed at constructing those halls, even the phrase "self-replicating data centers." The cost frame he used was blunt: roughly $70 billion per gigawatt over a five-year TCO, and they would be happier closer to $20 billion. Local inference does not shrink that bill if it also strands DRAM at a 5% duty cycle. So here is the open question I cannot shake. If the scarce resource is memory, not vibes about where the model lives — who still wants silicon idle on their desk once that tax shows up in the price of everything else? #AI #DRAM #OpenAI@NikTekYou're not able to afford GPU, RAM and Storage because Sam Altman thinks that its quite terrible to have DRAM sitting in your computer@1littlecoderAfter listening to this, I disassembled my computer, detached my RAM and donated it to OpenAI! I can't believe, I was so stupid and naive to think the DRAM inside my computer mattered. It didn't matter because OpenAI is working for the entire world and the compute concentrated on the data centers matter. Thanks @sama and cheeky pint for opening my eyes and contributing to global supply chain instead of gating it!@fengwan217Local AI looks virtuous. Altman calls it a DRAM tax. I used to nod along whenever someone said the endgame was open-source models on your own machine — no megacampuses, no GPU landlords, just a model humming in the living room. Then Sam Altman sat down with John Collison on Cheeky Pint and cut that story open. Plenty of people, he said, want local inference. They assume open-source on personal computers means we will not need data-center hardware. His pushback was not taste. It was scarcity. From a global allocation perspective, it is quite terrible to have everybody hold a chunk of the world's limited DRAM supply in their own computer — sitting there, used maybe 5% of the time. That 5% is the part that stuck. In a supply-constrained world you want maximum utilization in the cloud. Most guts pull the other way: keep the silicon at home, feel sovereign, call it freedom. Altman's frame is colder. Idle memory on a million desks is not liberty; it is a tax on the same chips everyone else needs to train and serve models at scale. Then he turned to the people who actually pour the wafers. For a long time, memory makers would sink something like $20 billion of CapEx to earn about $19 billion — boom, bust, almost never a great business across a full cycle. Until right now. These firms are each set to make something like $100 billion of free cash flow in a year. And CapEx still stays conservative. They remember nearly going bankrupt from overbuilding fabs. Altman's line landed flat: the invisible hand is failing right now. Cash floods in; the fabs do not race ahead as fast as demand wants. That scarcity is why OpenAI's prep list reads the way it does — its own chip aimed at intelligence per watt, learning to build data centers, robotics pointed at constructing those halls, even the phrase "self-replicating data centers." The cost frame he used was blunt: roughly $70 billion per gigawatt over a five-year TCO, and they would be happier closer to $20 billion. Local inference does not shrink that bill if it also strands DRAM at a 5% duty cycle. So here is the open question I cannot shake. If the scarce resource is memory, not vibes about where the model lives — who still wants silicon idle on their desk once that tax shows up in the price of everything else? #AI #DRAM #OpenAI@NikTekYou're not able to afford GPU, RAM and Storage because Sam Altman thinks that its quite terrible to have DRAM sitting in your computer@1littlecoderAfter listening to this, I disassembled my computer, detached my RAM and donated it to OpenAI! I can't believe, I was so stupid and naive to think the DRAM inside my computer mattered. It didn't matter because OpenAI is working for the entire world and the compute concentrated on the data centers matter. Thanks @sama and cheeky pint for opening my eyes and contributing to global supply chain instead of gating it!@fengwan217Local AI looks virtuous. Altman calls it a DRAM tax. I used to nod along whenever someone said the endgame was open-source models on your own machine — no megacampuses, no GPU landlords, just a model humming in the living room. Then Sam Altman sat down with John Collison on Cheeky Pint and cut that story open. Plenty of people, he said, want local inference. They assume open-source on personal computers means we will not need data-center hardware. His pushback was not taste. It was scarcity. From a global allocation perspective, it is quite terrible to have everybody hold a chunk of the world's limited DRAM supply in their own computer — sitting there, used maybe 5% of the time. That 5% is the part that stuck. In a supply-constrained world you want maximum utilization in the cloud. Most guts pull the other way: keep the silicon at home, feel sovereign, call it freedom. Altman's frame is colder. Idle memory on a million desks is not liberty; it is a tax on the same chips everyone else needs to train and serve models at scale. Then he turned to the people who actually pour the wafers. For a long time, memory makers would sink something like $20 billion of CapEx to earn about $19 billion — boom, bust, almost never a great business across a full cycle. Until right now. These firms are each set to make something like $100 billion of free cash flow in a year. And CapEx still stays conservative. They remember nearly going bankrupt from overbuilding fabs. Altman's line landed flat: the invisible hand is failing right now. Cash floods in; the fabs do not race ahead as fast as demand wants. That scarcity is why OpenAI's prep list reads the way it does — its own chip aimed at intelligence per watt, learning to build data centers, robotics pointed at constructing those halls, even the phrase "self-replicating data centers." The cost frame he used was blunt: roughly $70 billion per gigawatt over a five-year TCO, and they would be happier closer to $20 billion. Local inference does not shrink that bill if it also strands DRAM at a 5% duty cycle. So here is the open question I cannot shake. If the scarce resource is memory, not vibes about where the model lives — who still wants silicon idle on their desk once that tax shows up in the price of everything else? #AI #DRAM #OpenAI@NikTekYou're not able to afford GPU, RAM and Storage because Sam Altman thinks that its quite terrible to have DRAM sitting in your computer
