Why AI Labs Are Buying Mac Minis Isn’t the Reason Being Repeated

Four Mac Studios, wired together with an open-source tool, ran two full 671-billion-parameter DeepSeek models at once — and never drew more than 400 watts combined. A comparable Nvidia A100 setup for the same workload needs roughly twenty cards, a dedicated server room, and cooling, at a hardware cost north of 2 million yuan. That gap, demonstrated by Chinese tech outlet Aifaner in hands-on testing, is the concrete version of a story that has otherwise been told almost entirely in the abstract this week.

The abstract version is now everywhere. The Information reported on August 31 that OpenAI has purchased tens of thousands of Mac mini and Mac Studio units over recent months to train reinforcement-learning agents that operate computers — clicking through interfaces, filling forms, editing files. Anthropic is pursuing the same workload through a different route, renting Mac mini capacity via Amazon's EC2 Mac instances rather than buying hardware outright. Neither company has confirmed the numbers publicly, and The Information did not disclose how much capacity Anthropic is using or which specific tasks it supports.

The mechanism behind both approaches is real, and it's not marketing. Training a computer-use agent means running thousands of short, largely independent sessions rather than one enormous synchronized job. Apple's unified memory architecture — CPU, GPU and Neural Engine drawing from a single memory pool instead of shuttling data between separate GPU and system memory — removes a transfer bottleneck that matters more for that kind of workload than for pretraining a frontier language model from scratch. Nvidia's answer, the DGX Spark, borrows the Mac mini's dimensions and unified-memory pitch outright, which is its own signal of how seriously the chip industry now takes the approach.

Where the popular framing overreaches is in implying a wholesale shift in AI infrastructure spending toward Apple hardware. A MacStadium survey of roughly 300 DevOps teams, published this week, found that most of the current surge in Mac demand traces back to AI-assisted coding — more pull requests, more commits, more CI/CD build volume — not reinforcement-learning clusters. Half of those teams are actually planning to move Mac workloads to managed cloud providers rather than expand self-hosted fleets. The frontier-lab buying spree and the broader enterprise Mac boom are two separate demand curves that happen to be showing up in the same earnings line.

Apple's own posture reinforces the accidental nature of the windfall. The company has kept capital spending far below Microsoft, Amazon and Google's, and past reporting from The Information shows Apple executives discussed — but never launched — a public cloud rental business built on Apple silicon, similar in concept to AWS or Azure. Apple's disclosed AI infrastructure investment remains Private Cloud Compute, its privacy-focused system for offloading tasks too heavy for an iPhone, not a bid to rent capacity to outside developers or labs. What OpenAI and Anthropic are doing runs through consumer retail channels and AWS's existing Mac program instead — infrastructure Apple built for app developers, repurposed by AI labs because nothing else fit the job as cheaply.

That repurposing is now visible in delivery times. High-RAM Mac mini and Mac Studio configurations have been quoted at four to sixteen weeks out, and Apple pulled its cheapest Mac mini configuration entirely in May before pushing prices up twice more since — the base model now starts at $899, up from $599 in late 2024, with Apple citing component costs rather than AI demand as the driver. Mac revenue hit $10.4 billion last quarter, up 29% year over year, a number Apple has not attributed to AI labs but one that arrived in the same quarter The Information's report became public.