There is something wonderfully reassuring about owning far more computer than you need. The aluminium is cool. The specifications are heroic. Somewhere inside, enough memory to support a respectable research project waits patiently while you decide whether an email should begin with “Hi” or “Hello.” This is sovereignty. It also appears to involve a considerable amount of sitting around.
That, essentially, was Sam Altman’s point on Cheeky Pint, speaking with Stripe cofounder John Collison in the episode published on October 29. Discussing local AI, he called it “quite terrible” from a global allocation perspective to scatter scarce DRAM across personal computers used perhaps “5% of the time.” He wanted “maximum utilization” in the cloud. The percentage was illustrative, not a measured verdict on every desktop.
I agree with the underlying argument. If a scarce, expensive resource can serve many people instead of waiting for one person to finish lunch, sharing it deserves consideration. We accept this principle with trains, libraries and hotel rooms. Introduce a graphics processor, however, and suddenly everyone discovers an ancestral obligation to own the means of autocomplete.
Naturally, Altman has an interest in the conclusion. A cloud AI executive recommending cloud AI is roughly as surprising as a vineyard recommending lunch with wine. His incentives deserve scrutiny. They do not, by themselves, invalidate the arithmetic. Sometimes the man selling umbrellas has correctly noticed the rain.
My favourite small illustration comes from Two Minute Papers, a channel I admire for making research feel like a succession of delightful breaches of the laws of nature. Its promotions for Lambda point viewers towards rented GPU computing. Lambda offers infrastructure for training and running models, including machines with multiple accelerators. Even the excitement of discovery has a practical question attached: whose hardware will do the work?
There is nothing embarrassing about renting the answer. You can love astronomy without installing an observatory above the bathroom. You can appreciate aviation without keeping an Airbus in the garden. Computing has somehow acquired a peculiar moral expectation that serious enthusiasm should arrive in a box, preferably accompanied by a frightening invoice. There is a difference between having access to extraordinary machinery and becoming its unpaid facilities manager.
Consider the Mac Studio with 512GB of unified memory. It is an extraordinary desktop, and we should be precise about what that means. Such a machine can run powerful local models. Apple explicitly markets its large memory configurations for that purpose. Claiming otherwise would turn a reasonable economic argument into a technical own goal, complete with slow motion replay.
But memory capacity alone does not buy equivalence with the best hosted AI systems. The model must be available to download. Its weights must fit alongside working memory. Speed depends on computation, memory bandwidth, software and workload. Long conversations and simultaneous requests add demands. Getting the orchestra into your living room does not establish that it can perform there comfortably.
Even the simplest arithmetic is sobering. A hypothetical trillion parameters stored at four bits apiece require about 500GB for the weights alone, before additional overhead. That is an illustration, not a claim that parameter count determines intelligence. Nevertheless, the operating system may reasonably object to being asked to live in the hallway.
The cloud’s attraction is therefore partly organisational. Demand from different users can be pooled, and hardware can serve someone else after my task finishes. That creates an opportunity for higher utilisation, although no provider achieves perfection by putting the word “cloud” on an invoice. An idle rented GPU remains idle. It has merely acquired a billing department.
Nor does centralisation automatically establish environmental virtue. Actual efficiency depends on equipment, electricity, cooling, demand and what useful work gets done. Consumer memory and accelerator memory are also different products; a stick removed from a gaming PC cannot simply be posted to a data centre and promoted. The allocation argument concerns production and investment, not a national RAM collection day.
Local computing still has excellent reasons to exist. Privacy, offline access, predictable costs and control can justify owning capacity that sometimes rests. For sustained workloads, ownership may be economical. Experimentation has value too. A society in which every computational experiment requires a remote company’s permission would have achieved impressive utilisation and a rather depressing definition of freedom.
But those arguments become stronger when stated honestly. “I want control over my tools” is persuasive. “My hardware earns its keep every day” is testable. “I enjoy this machine” is perfectly respectable. The claim that buying vast amounts of personal memory necessarily represents superior resource allocation needs considerably more evidence than a photograph of a very tidy desk.
My preference is pragmatic: keep useful capability nearby and rent exceptional capacity when the task warrants it. That can mean local models for routine or private work, hosted services for demanding jobs, and rented GPUs for experiments. Open models fit this arrangement perfectly well. Freedom to choose software does not impose an obligation to house every transistor personally.
The emotional obstacle is that ownership feels like competence. A subscription feels like asking permission; a heavy computer feels like having prepared for civilisation’s next chapter. I understand the appeal. Still, I would rather judge my tools by the work they help me finish than by their potential to impress another person reading a specification sheet. Let the expensive machinery have a busy day. I have emails to overthink, and they require surprisingly little memory.
No comments yet