Musk quipped that people deserve their RAM too after Altman suggested underused PC memory could serve AI data centers.
Altman suggested underused memory in PCs could be better used in AI data centers amid a global supply crunch. Musk mocked the idea, saying, “The people deserve to have their RAM too!”
3.4M
7 Sources, first seen ago
@YL9104Sam Altman just told laptop owners their RAM is wasted, and he has a point. Just not the whole one. On the Cheeky Pint podcast on October 29, Altman argued that keeping the world's scarce DRAM spread across personal computers is, from a global allocation perspective, "quite terrible" — memory sitting in a machine most of the day, used about 5% of the time, while data centers could pool it and run it much harder. His answer is centralization: put the memory where the utilization is, in cloud infrastructure like Stargate. The arithmetic is hard to argue with. A chip bought for one person's busiest hour, idle the other twenty-three, is capacity spent on insurance, not work. Pooling memory the way power grids pool electricity lets the same chip serve more people. If you only care about total throughput per wafer, Altman wins the argument in one paragraph. But total throughput was never the only thing people were buying. Local models spread through October for reasons a utilization number doesn't capture: prompts that never leave your machine, work that continues with no connection, tools nobody can revoke or reprice on you. People weren't buying idle DRAM. They were buying the right not to ask permission. A cloud pool can be more efficient and still be the wrong trade for the person who wanted that right. Then there's the question Altman doesn't answer: who keeps the savings. Centralization concentrates bargaining power as well as memory. When the scarce input lives in a handful of data centers, the people who own those centers set the terms — and the memory price pressure already squeezing PC demand doesn't disappear, it just moves from a buyer's invoice to a provider's cost sheet, and eventually back to the buyer as a subscription. Efficiency gains don't automatically flow downhill. So the test isn't whether cloud memory is more efficient on paper. It clearly is. The test is whether the savings ever reach the person who paid for the laptop, and whether that person still has a real choice when they don't. Until both answers are yes, "your RAM belongs in a data center" is a supplier's argument dressed up as arithmetic.
@0xSigilRT @karankendre: Me buying a MacBook with 128GB of RAM just to run AI models locally Sam Altman :
@destructoidThe AI bubble is showing its first signs of cracking, Sam Altman wants your RAM, and Destructoid is going to ExileCon. Check in with what's going down by subscribing to our weekly newsletter (and to access our special dailies with exclusive content). https://newsletter.destructoid.com/p/destructoid-checkpoint-the-ai-bubble-shows-the-first-cracks@YL9104Sam Altman just told laptop owners their RAM is wasted, and he has a point. Just not the whole one. On the Cheeky Pint podcast on October 29, Altman argued that keeping the world's scarce DRAM spread across personal computers is, from a global allocation perspective, "quite terrible" — memory sitting in a machine most of the day, used about 5% of the time, while data centers could pool it and run it much harder. His answer is centralization: put the memory where the utilization is, in cloud infrastructure like Stargate. The arithmetic is hard to argue with. A chip bought for one person's busiest hour, idle the other twenty-three, is capacity spent on insurance, not work. Pooling memory the way power grids pool electricity lets the same chip serve more people. If you only care about total throughput per wafer, Altman wins the argument in one paragraph. But total throughput was never the only thing people were buying. Local models spread through October for reasons a utilization number doesn't capture: prompts that never leave your machine, work that continues with no connection, tools nobody can revoke or reprice on you. People weren't buying idle DRAM. They were buying the right not to ask permission. A cloud pool can be more efficient and still be the wrong trade for the person who wanted that right. Then there's the question Altman doesn't answer: who keeps the savings. Centralization concentrates bargaining power as well as memory. When the scarce input lives in a handful of data centers, the people who own those centers set the terms — and the memory price pressure already squeezing PC demand doesn't disappear, it just moves from a buyer's invoice to a provider's cost sheet, and eventually back to the buyer as a subscription. Efficiency gains don't automatically flow downhill. So the test isn't whether cloud memory is more efficient on paper. It clearly is. The test is whether the savings ever reach the person who paid for the laptop, and whether that person still has a real choice when they don't. Until both answers are yes, "your RAM belongs in a data center" is a supplier's argument dressed up as arithmetic.@0xSigilRT @karankendre: Me buying a MacBook with 128GB of RAM just to run AI models locally Sam Altman :@destructoidThe AI bubble is showing its first signs of cracking, Sam Altman wants your RAM, and Destructoid is going to ExileCon. Check in with what's going down by subscribing to our weekly newsletter (and to access our special dailies with exclusive content). https://newsletter.destructoid.com/p/destructoid-checkpoint-the-ai-bubble-shows-the-first-cracks@YL9104Sam Altman just told laptop owners their RAM is wasted, and he has a point. Just not the whole one. On the Cheeky Pint podcast on October 29, Altman argued that keeping the world's scarce DRAM spread across personal computers is, from a global allocation perspective, "quite terrible" — memory sitting in a machine most of the day, used about 5% of the time, while data centers could pool it and run it much harder. His answer is centralization: put the memory where the utilization is, in cloud infrastructure like Stargate. The arithmetic is hard to argue with. A chip bought for one person's busiest hour, idle the other twenty-three, is capacity spent on insurance, not work. Pooling memory the way power grids pool electricity lets the same chip serve more people. If you only care about total throughput per wafer, Altman wins the argument in one paragraph. But total throughput was never the only thing people were buying. Local models spread through October for reasons a utilization number doesn't capture: prompts that never leave your machine, work that continues with no connection, tools nobody can revoke or reprice on you. People weren't buying idle DRAM. They were buying the right not to ask permission. A cloud pool can be more efficient and still be the wrong trade for the person who wanted that right. Then there's the question Altman doesn't answer: who keeps the savings. Centralization concentrates bargaining power as well as memory. When the scarce input lives in a handful of data centers, the people who own those centers set the terms — and the memory price pressure already squeezing PC demand doesn't disappear, it just moves from a buyer's invoice to a provider's cost sheet, and eventually back to the buyer as a subscription. Efficiency gains don't automatically flow downhill. So the test isn't whether cloud memory is more efficient on paper. It clearly is. The test is whether the savings ever reach the person who paid for the laptop, and whether that person still has a real choice when they don't. Until both answers are yes, "your RAM belongs in a data center" is a supplier's argument dressed up as arithmetic.@0xSigilRT @karankendre: Me buying a MacBook with 128GB of RAM just to run AI models locally Sam Altman :@destructoidThe AI bubble is showing its first signs of cracking, Sam Altman wants your RAM, and Destructoid is going to ExileCon. Check in with what's going down by subscribing to our weekly newsletter (and to access our special dailies with exclusive content). https://newsletter.destructoid.com/p/destructoid-checkpoint-the-ai-bubble-shows-the-first-cracks@YL9104Sam Altman just told laptop owners their RAM is wasted, and he has a point. Just not the whole one. On the Cheeky Pint podcast on October 29, Altman argued that keeping the world's scarce DRAM spread across personal computers is, from a global allocation perspective, "quite terrible" — memory sitting in a machine most of the day, used about 5% of the time, while data centers could pool it and run it much harder. His answer is centralization: put the memory where the utilization is, in cloud infrastructure like Stargate. The arithmetic is hard to argue with. A chip bought for one person's busiest hour, idle the other twenty-three, is capacity spent on insurance, not work. Pooling memory the way power grids pool electricity lets the same chip serve more people. If you only care about total throughput per wafer, Altman wins the argument in one paragraph. But total throughput was never the only thing people were buying. Local models spread through October for reasons a utilization number doesn't capture: prompts that never leave your machine, work that continues with no connection, tools nobody can revoke or reprice on you. People weren't buying idle DRAM. They were buying the right not to ask permission. A cloud pool can be more efficient and still be the wrong trade for the person who wanted that right. Then there's the question Altman doesn't answer: who keeps the savings. Centralization concentrates bargaining power as well as memory. When the scarce input lives in a handful of data centers, the people who own those centers set the terms — and the memory price pressure already squeezing PC demand doesn't disappear, it just moves from a buyer's invoice to a provider's cost sheet, and eventually back to the buyer as a subscription. Efficiency gains don't automatically flow downhill. So the test isn't whether cloud memory is more efficient on paper. It clearly is. The test is whether the savings ever reach the person who paid for the laptop, and whether that person still has a real choice when they don't. Until both answers are yes, "your RAM belongs in a data center" is a supplier's argument dressed up as arithmetic.@0xSigilRT @karankendre: Me buying a MacBook with 128GB of RAM just to run AI models locally Sam Altman :@destructoidThe AI bubble is showing its first signs of cracking, Sam Altman wants your RAM, and Destructoid is going to ExileCon. Check in with what's going down by subscribing to our weekly newsletter (and to access our special dailies with exclusive content). https://newsletter.destructoid.com/p/destructoid-checkpoint-the-ai-bubble-shows-the-first-cracks#4
#4
