TL;DR — Key Takeaways
- Your computer, your choice. Sam Altman’s argument for centralized AI overlooks the privacy, independence and control that local computing provides.
- Utilization isn’t profitability. Wall Street is questioning massive AI infrastructure investments, even as providers argue that more centralized capacity is the answer.
- Cloud AI must earn its place. Customers should choose where AI runs based on cost, performance and control — not what best serves Big Tech’s business model.
Sam Altman picked an unfortunate moment to explain why the memory in our computers would be better utilized in somebody else’s data center. As his comments about local AI inference circulated on October 30, Wall Street was selling technology stocks amid growing concerns about the cost of the AI buildout. The industry wants more infrastructure, more capital and more time to prove the returns. Apparently, some of us also need a lesson in why owning the hardware to run AI ourselves is an inefficient use of scarce resources.
Coming from the CEO of OpenAI, that sounds an awful lot like “let them eat cake.” Altman did not say those words, and I am not suggesting he did. But there is an entitlement in presenting an arrangement favorable to your business as the sensible allocation of resources for everybody. Before we accept that prescription, we ought to examine whose interests it serves and what it leaves out.
In his October 29 appearance on John Collison’s Cheeky Pint podcast, Altman challenged the idea that open models running on personal computers would solve the infrastructure problem. He argued that scarce DRAM distributed across individual machines receives too little use, citing a 5% figure without supplying a supporting study or explaining how he measured it. His prescription was direct: “You want maximum utilization in the cloud if you’re trying to deal with this very supply constrained environment.”
There is a legitimate argument here. Buying expensive equipment for occasional use can be a lousy investment, and sharing infrastructure across customers can improve its economics. Businesses have been making that calculation since long before anyone could ask a chatbot to write a procurement policy. If Altman were simply advising someone against buying a dedicated machine for a handful of weekly prompts, I would understand the advice. His global-allocation framing makes a much broader claim.
My Computer Has Other Jobs
My computer does not exist to maximize an AI provider’s utilization rate. Its memory supports the things I need to do, whether that involves writing, research, communications or running a model. Memory that is not performing inference is not automatically doing nothing. And a customer using equipment already purchased is making a different economic decision from someone ordering a new machine exclusively for AI. A serious comparison has to account for those differences before declaring one arrangement wasteful.
Ownership also delivers benefits that a utilization percentage cannot measure. A business may want sensitive information to remain on equipment it controls. Someone working without a reliable connection may need a tool that functions offline. Another customer may value the ability to choose a model, manage its deployment and avoid dependence on a provider’s changing prices or service availability. Those choices carry costs and responsibilities, but customers are entitled to decide what those benefits are worth.
These are practical capabilities, not a romantic defense of the desktop. Apple’s developer documentation explicitly describes on-device AI that works offline and keeps data on the device. OpenAI itself promoted gpt-oss-20b as capable of running within 16 GB of memory, supporting local inference and experimentation without expensive infrastructure. Neither example proves local execution is always cheaper. Both demonstrate that the alternative has value beyond the amount of time a memory chip stays busy.
OpenAI’s support for local models does not automatically contradict every argument Altman makes about hardware scarcity. It does, however, complicate the suggestion that distributing memory across personal computers is poor resource allocation. His company has explained why people might reasonably want to do exactly that. The economics depend on the workload, the hardware already available and the customer’s requirements. His commercial interest belongs in that calculation, too: cloud inference is central to the business he leads.
Wall Street Checks the Bill
Investors were conducting their own examination of resource allocation on October 30. Reuters reported that the Nasdaq Composite closed down 2.86% and the S&P 500 fell 1.82%, as disappointing earnings and concerns about AI spending weighed on technology shares. Company-specific problems and guidance contributed to the selling. There is no basis for attributing the decline to Altman’s remarks, but the juxtaposition is revealing. While he was making the case for better utilization of scarce hardware, investors were questioning the financial results of acquiring so much of it.
Amazon’s earnings show why this deserves more thought than a declaration that AI demand is imaginary. AWS reported quarterly revenue of $46 billion, up approximately 39%, and operating income of $17.8 billion. Those are substantial results, although AWS includes much more than AI. Across Amazon, trailing 12-month operating cash flow reached $168.3 billion. This is a company with customers, profitable operations and considerable cash generation.
Yet Amazon’s trailing 12-month free cash flow was negative $42.4 billion, compared with positive $14.8 billion a year earlier. The company attributed the deterioration primarily to increased property and equipment purchases, with that increase principally reflecting AI investment. These are company-wide cash-flow figures, not an AWS operating loss. They show how an enterprise can produce substantial operating cash while spending even more on its expansion.
That spending may ultimately prove worthwhile. It may create capacity that generates attractive returns for years. But the scale of the commitments makes the assumptions about future demand, pricing, operating costs and equipment life consequential. Investors are entitled to scrutinize those assumptions even when revenue is growing quickly. Keeping infrastructure busy is one requirement for a successful investment; earning enough from the work to justify its full cost is another.
Sam Understands the Cycle
One of the most revealing parts of the podcast came before the discussion of local inference. Altman acknowledged that memory manufacturers remember earlier boom-and-bust cycles and have become cautious about expanding capacity. They have experienced the consequences of building too much after hearing persuasive arguments about future demand. That history helps explain why high prices do not immediately produce all the new supply AI companies want.
I would give that caution more credit. A shortage today does not guarantee attractive returns on equipment delivered years from now, and the people financing expansion have to consider both conditions. Memory suppliers and AI investors occupy different parts of the supply chain, but both must decide how much confidence to place in forecasts. Neither owes an infrastructure buyer unquestioning faith that this expansion will pay off as expected.
The same freedom belongs to customers. Cloud inference should win work when it delivers the best combination of capability, cost, reliability and appropriate protections. Local execution should win when it meets the requirement more effectively, and some businesses will use both. Providers can make their case with measured results. Describing customer-owned capacity as a poor global allocation does little to establish what an individual customer should actually buy.
I am bullish on what people can accomplish with AI. I am considerably less receptive to the suggestion that scarce resources naturally belong wherever the largest AI companies want to concentrate them. Investors are asking those companies to justify their spending, and customers should demand comparable discipline before committing their own money. If the cloud offers superior value, sell us on that value. The industry’s enormous appetite for infrastructure does not give it a claim on everybody else’s computing choices.
