Altman Says 95% of Personal Computer Chips Sitting Idle Is Wasteful, Author Pushes Back
In an interview, Altman said that the memory chips in personal computers are in use only 5% of the time, which is a huge waste, and that chips should be pooled together to maximize utilization. The author argues that this is essentially taking back ownership of personal hardware: local 64G and 128G unified memory can run open-source models, so users no longer need to pay $20 a month in rent to the cloud, and who owns the compute is the real dividing line in the AI era.
Body · AI translation
In an interview, Altman made a highly controversial remark: everyone is occupying precious memory chips in their personal computers, yet they are only in use 5% of the time—a huge waste. This is terrible; the chips should be pooled together to maximize utilization.
This is like a ride-hailing company advising you to smash all your private cars, on the grounds that leaving cars parked in the garage is a waste of steel. On the surface it is about computing utilization, but what it really wants to take away is your hardware ownership.
Or you could say it is a top technocrat agonizing over global semiconductor capacity. But if we dig one layer deeper into this calculation, what he really wants to reclaim is your ownership of personal hardware.
It is like the CEO of a ride-hailing platform standing on stage preaching to everyone:
Each of you buys a private car and leaves it sitting idle in the garage 90% of the time—this is a tremendous waste of global steel capacity. Everyone should sell their cars and ride in my ride-hailing service.
Efficiency is indeed maxed out, but freedom has also been confiscated.
That 95% of idle time on a personal computer has never been waste. It is the freedom premium humanity pays in exchange for being able to boot up at any time, zero censorship, data that belongs entirely to oneself, and still being able to operate after unplugging the network cable.
If we do the math for the giants, we can see the real calculation behind it: when everyone's Mac or PC is stuffed with 64G or 128G of unified memory and can smoothly run high-quality open-source models locally, the giants will face two nightmares:
First, they cannot grab enough memory chips to expand cloud data centers. Second, you no longer need to pay $20 a month in rent to the cloud.
In AI's development to this day, the real watershed is no longer the score rankings of models, but the ownership of computing power: whether to become a digital tenant where all computing power belongs to the cloud, or to retain a piece of offline silicon of your own.
Once all intelligence must connect to the giants' servers to work, what you own is no longer a personal computer, but merely a cloud remote control that can be cut off at any time.
Never easily hand over your local memory. Even if it sits idle 95% of the time, it is still your last private territory in the digital world. https://t.co/YTTfLlKs2P