OpenAI has purchased tens of thousands of Apple Mac mini and Mac Studio units over the past year, and Anthropic is renting comparable Apple hardware capacity through Amazon Web Services, according to The Information, whose report on the buying spree began circulating widely Sunday. [1] The AI labs are not running Apple's models. They are using the machines to train "computer-use" agents — AI systems designed to operate a desktop the way a person would, clicking, typing and navigating software — work that Apple's unified-memory architecture turns out to suit unusually well. [1]

The demand has been strong enough to create a real supply problem for ordinary buyers. Apple's most powerful Mac Studio configurations have been effectively sold out for months, a shortage the report ties directly to a broader memory-chip crunch rippling through the industry. [1] German outlet The Decoder, summarizing the same reporting, noted that OpenAI wants more Mac minis than it can currently get, and that the labs are not alone: the Mac mini has been picking up momentum as a local AI machine more broadly, a trend partly fueled by hype around AI coding tools that run well on Apple silicon. [2]

The mechanism is specific to how Apple builds its chips. Unlike traditional systems that separate a processor's memory from a graphics card's memory, Apple's chips share one pool of fast memory across the entire machine — a design choice made originally for laptops and desktops, not data centers. [2] For AI workloads that need to hold large models in memory without shuttling data back and forth, that architecture happens to work well, and it works in a small, quiet, relatively cheap box that is far easier to rack in bulk than a server built around Nvidia's data-center GPUs. Nvidia's own answer to that use case, the DGX Spark, takes a different approach, built around dedicated GPU power using Nvidia's CUDA software and Tensor cores rather than Apple's shared-memory design. [2]

The buying pattern extends beyond the two labs' internal training runs. An open-source project called Exo lets users link multiple Macs together into a cluster capable of running large AI models locally, turning consumer-grade hardware into makeshift server farms. [2] And Peter Voell, a former member of OpenAI's own computing infrastructure team, has gone further still, building an Apple-hardware-based cloud service called Mount Thor aimed at renting out exactly this kind of capacity. [2] What began as opportunistic bulk-buying by two AI labs is starting to look like the seed of a parallel, Apple-hardware compute market operating outside Apple's own cloud ambitions entirely.

For Apple, the effect has shown up in a place the company did not engineer for: its own earnings. Mac revenue jumped nearly 29 percent year-over-year to $10.4 billion in the June quarter, a number large enough to draw attention on an earnings call built mostly around iPhone and services growth. [1] [2] That is the frame most coverage of the story has used — an AI-adjacent windfall landing on a legacy product line almost by accident, with Apple as a passive beneficiary of demand it did not anticipate and, per the shortage reports, cannot yet fully supply.

Hardware-focused commentary reads the same facts differently. If OpenAI and Anthropic — two of the companies best positioned to know what infrastructure the next generation of AI agents will actually require — are choosing Apple's shared-memory chips over Nvidia's dedicated GPU stack for a meaningful category of AI workloads, that is not simply a supply-and-demand accident. It is early evidence that Apple's chip architecture, built for laptops, has an unplanned second life as AI infrastructure Nvidia does not currently offer an equivalent for. Nvidia has spent years positioning itself as the default hardware layer beneath every AI lab's ambitions; a computer-use training pipeline that runs better on a Mac mini than a Nvidia rack is a small but genuine crack in that positioning, even if it applies to only one category of AI work for now.

Neither OpenAI nor Anthropic has said publicly how large the Mac-based portion of their infrastructure is relative to their overall GPU spending, and nothing in the reporting suggests Apple hardware is displacing Nvidia chips for the large-scale model training that remains the industry's dominant cost center. What the purchases do establish is narrower and still notable: for the specific task of teaching an AI agent to operate a computer, two of the field's best-funded labs decided the machine built to sit on someone's desk was the better tool for the job — and bought enough of them to strain Apple's own supply chain in the process.

-- THEO KAPLAN, San Francisco