On August 30, 2026, The Information reported that AI labs including OpenAI had purchased tens of thousands of Mac minis and Mac Studios over the past few months. According to the report, OpenAI is using them for reinforcement learning of computer-use agents and is seeking even more Macs. Anthropic is also said to be renting Mac minis from Amazon Web Services (AWS) for similar purposes. None of the companies have disclosed unit counts, model breakdowns, or purchase amounts. Still, piecing together publicly available technical documentation reveals how the supply constraints in AI development are expanding from raw GPU compute to the desktop environments in which agents actually operate.
Scaling Desktop Environments, Separately From GPUs
In computer-use agents, model inference and desktop-environment execution are kept separate — and on AWS EC2 Mac, each physical Mac hosts exactly one Mac instance. According to Anthropic's published specifications, the model reads a screenshot and returns an action request such as a click or keystroke. An application on the user's side executes that action in a virtual machine or container and sends the new screen back to the model. This action loop repeats until the task is complete.
In reinforcement learning, a large number of these action trials are collected, and the model is updated based on successes and failures. Even if the model's inference can be handled on separate compute infrastructure, an execution environment with its own browser and files must be provisioned separately. The application and screen state must also be managed for every single trial. There is also the work of resetting failed environments to a known initial state and isolating them so they don't interfere with the next trial. The environment isn't auxiliary infrastructure — it's part of the training pipeline itself.
Renting Macs via the cloud doesn't eliminate this physical constraint either. AWS EC2 Mac offers Mac minis as dedicated bare-metal hosts, with only one Mac instance running per host and a minimum allocation period of 24 hours. If reports that Anthropic uses Macs through AWS are accurate, the company has shifted purchasing and maintenance to AWS — not eliminated its dependence on physical Macs. Meanwhile, the virtualization configuration of the Mac fleet OpenAI owns is unknown, so it's not necessarily the case that one physical machine always maps to one action trial.
High Scores on Ubuntu Don't Guarantee Success on macOS
On OSWorld, a benchmark that operates real operating systems, the computer-use agent OpenAI published in January 2025 achieved a 38.1% success rate — far behind the human benchmark of 72.4%. That figure isn't so much a measure of current model performance as a clue to the nature of the training environment.
At the time, OpenAI used a VMware-based Ubuntu virtual machine for its OSWorld evaluation. Notably, the company recorded that performance improved slightly simply by moving the dock from the left side of the screen to the right. Even when an app's functionality stays the same, changing icon positions or screen layout can throw off the model's behavior. For a model that manipulates a GUI based on coordinates and images, the OS and screen layout aren't interchangeable wrappers.
An unreviewed preprint called "MacArena," published in June 2026, pushed this OS-dependency finding further. The research team set up 421 tasks across 50 apps in virtual machines running on Apple Silicon. Model rankings reversed on Mac-specific tasks, with one leading model scoring more than 26% lower on the Mac-specific portion. Simply adding more Ubuntu environments doesn't teach a model to navigate macOS menus, apps, and window management.
The longer the task, the longer the environment stays occupied. Another unreviewed preprint, "OSWorld 2.0," compiled 108 tasks and reported a median human completion time of about 1.6 hours. Claude Opus 4.7 required an average of 318 tool calls — a sharp jump from roughly 30 in the previous version. Even when evaluations allowed up to 500 steps, the best completion rate was just 20.6%. If individual trials take a long time and often fail, then increasing the amount of training data collected in a given period requires parallelizing the environments themselves.
Unified Memory Is Only Part of the Story
On August 25 — just before The Information's report — Apple refreshed the Mac mini and Mac Studio lineups. The M5 Pro Mac mini offers up to 64GB of unified memory and 307GB/s of memory bandwidth, which Apple says is enough to run large local AI models. The M5 Ultra Mac Studio can be configured with up to 512GB.
However, running local AI models and training environments for computer-use agents aren't the same kind of workload. The Mac Studio's large memory capacity gives it more headroom for running models locally. The Mac mini's compact chassis, meanwhile, suits deployments where many independent environments with their own displays need to be lined up side by side. But which machines OpenAI is using to host models, and which it's using purely as environments, hasn't been reported. Filling in that distinction with speculation risks the mistaken conclusion that Macs have replaced GPU clusters.
Price differences also point to different use cases. The new M6 Mac mini starts at $899 in the US. The Mac Studio starts at $2,499 for the M5 Max and $5,499 for the M5 Ultra. If what's needed is a large number of environments with a display and OS, buying more Mac minis makes sense; if the goal is fitting a large model onto a single machine, the Mac Studio looks like the rational choice. Still, without a breakdown of the tens of thousands of units purchased, this remains only an inference drawn from published specs.
Mac Mini's Entry Price Rose 50% in Under Four Months
The Mac mini's lowest US price rose from $599 on May 1, 2026, to $799, and then to $899 with the M6 generation on August 25 — an increase of roughly 50% in under four months. No causal link to AI-lab procurement has been confirmed. What can be confirmed is that the Mac mini's minimum price rose in the US, and that reports in Japan pointed to long lead times for high-memory configurations.
That 50% figure isn't simply a price hike on an unchanged product. According to Bloomberg, Apple removed the 256GB storage base configuration in May, making the 512GB configuration the new entry point and raising the minimum price from $599 to $799. With the M6 generation in August, the starting price rose to $899. Because the generation and storage capacity differ, what's actually comparable is the change in the minimum amount a consumer must spend to buy a Mac mini at all.
Supply-side anomalies predate the August refresh. WIRED.jp reported that lead times grew longer for higher-memory configurations, with some models facing months-long waits in Japan as well. Bloomberg reported that AI demand and processor supply shortages were straining inventory. OpenAI's purchases alone can't be definitively blamed for the price increase. But if research labs are seeking consumer desktops in batches of thousands or tens of thousands, that creates a demand wave distinct from ordinary PC sales.
Can Apple Become an AI Infrastructure Company?
Apple itself has begun marketing the new Mac mini as a product suited to "always-on agentic processing." This marks a shift for the Mac — from a product placed one at a time on an individual's desk to computing infrastructure that businesses deploy and operate in bulk. But what research labs need doesn't end with the hardware chassis. Environments must be initialized quickly, and failed units must be swappable. Unless OS and app versions are kept consistent and action trials can be tracked, sheer unit count won't translate into faster training.
If the reported figure of tens of thousands of units is accurate, Apple hasn't entered head-on into the same compute market as Nvidia. Rather, it's becoming a supplier of "the place where agents work" — a layer that has emerged outside the GPU. Whether this succeeds will depend on whether OpenAI and Anthropic keep expanding their Mac fleets for next-generation training, whether Apple can reliably supply high-memory configurations, and whether Apple can turn management of Mac-specific environments into an enterprise-grade service. If those conditions align, the Mac mini will straddle two markets at once: a compact personal computer, and a data-center component for raising computer-use agents.