OpenAI is buying Mac Mini's by the tens of thousands and the interesting part is what this says about where the bottleneck has moved. The scarce input for this generation of models isn't only GPUs, it's realistic places for agents to practice. Labs already pay heavily for those places. Anthropic leadership has reportedly discussed spending over $1 billion on RL environments in the next year, and vendors like Surge (about $1.2B in revenue from the big labs) and Mercor (valued at $10B) are racing to build task environments (gyms) for them. Why computer use needs real computers: Both OpenAI and Anthropic are trying to build models that operate software the way a person does. Look at the screen, click, type, check the result. The training method is reinforcement learning at scale roughly following this process: 1. spin up thousands of real desktop sessions 2. hand the agent tasks like editing and testing code or organizing email 3. score the outcome 4. update the model on what worked Practitioners describe single runs sustaining this loop for days or weeks, and most of that time goes to the agents working through tasks, not to updating the model weights. The reason to focus on Mac is that macOS accounts for roughly a fifth of the world's desktops, second only to Windows, and overrepresented among the developers these agents are supposed to help. The only way to get a model that is expert at using a Mac is to have it use a Mac, millions of times. All that practice is showing up on the scoreboard between @OpenAI and @AnthropicAI. Who's ahead right now in the Computer Use race: The main eval is OSWorld-Verified and in this test, an agent gets 369 real desktop tasks and passes only if the machine ends up in the right state. Humans scored about 72% when the benchmark launched. The top models are now past that. - Claude Fable 5 (Anthropic's flagship): 85% (https://t.co/8Yd0qGUozZ) - GPT-5.5 (OpenAI's best published on this test): 78.7% (https://t.co/12ipJY4hDl) NOTE: OpenAI's newest flagship, GPT-5.6 Sol, reports on a different variant: 62.6% on OSWorld 2.0, a newer and harder task set where it's the state of the art (https://t.co/xdjfNwcufk). These are different tests, so the numbers don't compare directly. The honest summary is Anthropic leads the older benchmark, OpenAI leads the newer one. Why Mac minis and not macOS Virtual Machines in the cloud: This part has to do with licensing, not physics. Apple's macOS license only permits macOS to run on Apple hardware, so there is no legal way to spin up a macOS VM on an ordinary server fleet the way you would a thousand Linux or Windows machines. AWS's "Mac instances", which Anthropic uses, are literally racks of physical Mac minis, billed with a 24-hour minimum allocation to comply with the Apple macOS Software License Agreement. So if you want macOS desktops at scale, you buy Macs or you rent real ones. A headless $899 mini drawing 60-120W is the cheapest legal container for a macOS session that exists. The rent vs buy math: Buying an M6 Mac mini lists at $899. Renting the closest AWS host, an M4 mini, runs about $1.23 an hour. So left on 24/7 that's roughly $10,800 a year, or about $6,000 with a 3-year commitment. This means that break even for a single Mini running around the clock, happens in about a month compared to the rental prices. So Anthropic's route costs 7-12x more per box per year. What it buys: no cash up front, no data center ops for an unusual hardware class, and a fleet that scales down when a run ends. My guess is it also signals commitment level. You buy what you know you'll use for years and rent what you're still deciding about. Amazon being Anthropic's biggest cloud partner and investor makes AWS-first the natural posture anyway. What actually happens on these machines: 1. The minis get racked headless and imaged with sandboxed macOS sessions. They are sandboxed because a half trained agent can theoretically click anything, which means practitioners need to keep them off the open web so they don't delete data or buy $5,000 items. 2. Each session runs one agent attempt: screenshot in, action out, for trajectories up to hundreds of steps. 3. An automated verifier checks the end state. Did the code pass its tests, did the email get filed correctly? 4. The trajectories flow back to a GPU cluster where the weights actually update, and the improved model redeploys to the fleet. The Macs mainly generate the training data, and the gradient math almost certainly stays on Nvidia GPUs. Procurement: Compared to acquiring GPUs, acquiring Mac Minis is simple. Data center GPU capacity is allocated years in advance. Mac minis sit in retail channels you can buy from this week. The catch is that consumers share those channels. High memory Mac configs have been out of stock for months and Apple pulled the 512GB Studio from sale at one point, while Mac revenue jumped 29% year over year to $10.3B. Apple even shipped its M6 mini refresh this month instead of the usual fall window. My read: GPUs became the commodity layer for training compute, and Mac minis are starting to become part of a commodity layer for something newer, generating agent experience. A $899 desktop is the cheapest legal unit of real macOS practice, and the labs are treating it as a component, not a computer.