Sam, 5% Utilization Is Not 95% Waste.
In the October 29 Cheeky Pint interview, Sam Altman described DRAM sitting in personal computers used just 5% of the time as "quite terrible" from a global allocation perspective.
Let's examine that allocation from both sides.
A personal computer is not a badly managed data-center rack. It belongs to its owner. The freedom to use it once a week, run a private model, build software, or simply keep it available is not waste.
Meanwhile, AI companies are competing for enormous volumes of chips, memory and power. For individual buyers, capable GPUs and enough RAM have become painfully expensive.
A corporation reserving capacity is called strategic. An individual buying capacity is called inefficient. That's a convenient definition of efficiency.
And we have not even asked whether every AI request needs a giant model thousands of kilometers away.
A capable small local model can handle substantial parts of routine work: translation, drafting, local search and basic automation. For harder tasks, I use powerful external models myself. Frontier research and tightly coupled training need data centers. That is precisely why a sensible system uses both.
Local where sufficient. Cloud where justified. The owner decides.
I have also published work on an Economy of Experience within my World Intelligence research. It goes beyond renting spare GPU hours. Personal AI could preserve useful lessons from real work and share narrowly permitted, verified results without exporting private history.
That is not a finished market. It is a more ambitious economic direction than treating everyone as a paying endpoint.
If OpenAI wants idle GPU capacity, offer a voluntary exchange: validated public jobs for meaningful credits, with transparent electricity and hardware costs. BOINC and Salad show distributed computing is not new. But don't confuse using spare GPU cycles with manufacturing scarce DRAM.
The real question is who gets to own AI capability.
Make capable local models practical. Make personal hardware affordable. Let cloud services win because they solve harder problems, not because buying your own computing power has become out of reach.
Measure useful work, including energy, capital costs and human outcomes. Server utilization alone is not social efficiency.
My PC does not need to be busy all day to justify its existence.
Neither do I.
Source: Cheeky Pint, 29 October 2026