OpenAI Bought Tens of Thousands of Mac Minis to Train AI Agents

OpenAI has purchased tens of thousands of Apple Mac mini and Mac Studio systems to train AI agents, according to a report from The Information. Anthropic is pursuing the same hardware but through a different route, renting Mac minis via Amazon Web Services. The surge in enterprise demand has pushed Apple Mac revenue up nearly 29 percent year on year to $10.3 billion in the latest quarter and reportedly prompted Apple to refresh the Mac mini and Mac Studio earlier than its usual autumn release window.

What happened

Data point	Detail
OpenAI hardware purchase	Tens of thousands of Mac mini and Mac Studio systems
Anthropic approach	Renting Mac minis through Amazon Web Services
Apple Mac quarterly revenue	$10.3 billion, up nearly 29% year on year
New Mac mini chip	M6
New Mac Studio chips	M5 Max or M5 Ultra
Refresh timing	Late August, ahead of the usual October/November window

According to a report from The Information published on August 31, 2026, OpenAI is using Mac hardware specifically for reinforcement learning (RL), a training method where an AI model learns by trying actions and receiving feedback, and for training computer-use agents. These agents are designed to navigate software interfaces, run multi-step tasks, and interact with a desktop the way a person would. OpenAI is said to be actively trying to acquire even more units beyond the tens of thousands it has already bought.

Anthropic is taking a slightly different path. Rather than purchasing the hardware outright, it is renting Mac minis through AWS and using them for comparable agent training work.

Why are AI labs choosing Mac hardware over Nvidia GPU clusters?

Apple’s Unified Memory Architecture is the key reason. In a typical PC or server setup, the CPU and GPU each have their own memory pools, and data has to be copied between them. Apple silicon uses a single shared memory pool accessible by the CPU, GPU, and Neural Engine at once. That removes a significant bottleneck when running large models or many simultaneous sessions.

The thermal design of the Mac mini and Mac Studio also matters. Both machines can sustain compute-heavy workloads without the power draw and cooling overhead of a full GPU cluster rack. For reinforcement learning workloads that involve running many parallel agent sessions, that adds up to real cost savings.

Nvidia took notice. According to the report, Nvidia now considers Apple its biggest competitor in local AI compute and launched the DGX Spark (a compact, high-performance AI workstation) in direct response.

What this means for Apple

Apple was reportedly caught off guard. The company had no dedicated engineering team for business customers, no developer-relations staff focused on this segment, and no formal enterprise AI strategy when demand spiked. High-end Mac mini and Mac Studio configurations have been out of stock for months amid a global memory shortage.

The unexpected demand is credited with pulling the Mac mini and Mac Studio refresh forward by roughly two months. The M6 Mac mini and the M5 Max and Ultra Mac Studio were announced in late August rather than the usual October or November slot.

Mac revenue is now Apple’s fastest-growing hardware line, rising nearly 29 percent year on year to $10.3 billion in the most recent quarter. That growth is being driven at least partly by bulk enterprise orders from AI labs, not just consumer upgrades.

Our take

This is a genuine shift, not a one-quarter blip. AI labs are hitting real constraints with GPU cluster capacity and cost, and Apple silicon offers a credible workaround for specific workloads: many short, parallel sessions rather than one massive training run. Reinforcement learning and agent training fit that pattern well.

The more interesting story is what it reveals about AI infrastructure strategy. OpenAI is buying tens of thousands of Mac minis because Nvidia GPU time is expensive and scarce, not because Apple silicon is secretly better at everything. These are complementary tools for different parts of the training pipeline.

For businesses exploring AI integration, the practical lesson is that you do not always need the most expensive hardware. Matching the workload to the right compute is the skill. The labs are figuring that out at scale; smaller teams should be thinking about it too.

We covered a related shift when shared-memory AI cut lab setup times from months to days. The Mac mini story is the consumer-grade version of the same underlying principle: unified memory reduces friction for certain AI tasks significantly.

What to do about it

1. If you run local AI workloads (inference, small fine-tunes, agent testing), benchmark a Mac mini or Mac Studio against your current setup before your next hardware purchase.
2. If you use AWS, check whether Mac instance types (AWS already offers EC2 Mac instances) fit your workload before committing to GPU instances.
3. Watch Mac Studio availability. Stock shortages mean lead times are long. Order early if you need the hardware for a specific project timeline.
4. Do not over-index on one platform. GPU clusters still win for large pre-training runs. Apple silicon is strong for inference and RL workloads, not a universal replacement.

The practical takeaway: Apple silicon is now a legitimate line item in an AI infrastructure budget, and the labs are already voting with purchase orders.

Source: Bing News · OpenAI