Superpower Daily
Google Courts Hollywood for AI Licenses, but Talent Rights Are...John Ternus Takes Apple’s Top Job as AI and China Pressures Con...Trump Posts AI Kharg Strike Videos After U.S. Hits Launchers on...Hack The Box Benchmark: Best Human Team Cleared 36 Challenges;...MIT Maps Julia’s Path to 1 Million Users as JuliaHub Pushes AI...Israeli Government-Backed Hanover Site Published 560,000 Words...Google Courts Hollywood for AI Licenses, but Talent Rights Are...John Ternus Takes Apple’s Top Job as AI and China Pressures Con...Trump Posts AI Kharg Strike Videos After U.S. Hits Launchers on...Hack The Box Benchmark: Best Human Team Cleared 36 Challenges;...MIT Maps Julia’s Path to 1 Million Users as JuliaHub Pushes AI...Israeli Government-Backed Hanover Site Published 560,000 Words...
Productspublished

OpenAI Reportedly Buys Tens of Thousands of Macs for Computer-Use Agent Training

The reported purchase points to dedicated capacity for agent training alongside GPU systems, while its cost and performance tradeoffs remain unclear.

By 2 min read
OpenAI Reportedly Buys Tens of Thousands of Macs for Computer-Use Agent Training
OpenAI Reportedly Buys Tens of Thousands of Macs for Computer-Use Agent Training
Sign in to save and react

Listen to this story

The audio brief

About 1:24
0:001:24
Read transcript
OpenAI has reportedly bought tens of thousands of Mac minis and Mac Studios to train computer-use agents, according to The Information. These are systems designed to handle multi-step tasks on a computer, rather than simply answer a single prompt. The reported fleet also supports reinforcement learning, linking the Apple hardware to a specific part of agent development. The important distinction is that this does not describe a replacement for OpenAI’s wider model-training infrastructure. It points instead to dedicated capacity for workloads where an agent must repeatedly observe a computer, choose an action, and learn from the result. The arrangement also contrasts with Anthropic’s approach. Anthropic is reportedly renting Mac minis through AWS, while OpenAI is said to be buying the machines directly. That raises a practical question about ownership: whether direct control makes sense when the hardware is specialized and demand may change. The Macs also sit alongside, not necessarily instead of, the GPU systems that dominate large-scale AI infrastructure. There is still no like-for-like data on configuration, cost, performance, or scalability versus Nvidia hardware for the same agent workload. And high-end Macs have reportedly been sold out for months amid a memory-chip shortage. The key thing to watch is whether this becomes an economically compelling niche for computer-use training, or simply an expensive way to secure scarce capacity.

Story brief

3 key points

OpenAI is reportedly building a large Apple-based compute fleet—tens of thousands of Mac minis and Mac Studios—for reinforcement learning and agents that operate computers across multi-step workflows. The move signals demand for specialized, non-GPU capacity, but not a broader retreat from Nvidia hardware. Anthropic is taking a more flexible route by renting Mac minis through AWS. The economics remain unclear:...

  1. 01

    The reported fleet is for computer-use and reinforcement-learning workloads, not a replacement for OpenAI’s broader model-training infrastructure.

  2. 02

    Anthropic rents Mac minis via AWS, contrasting with OpenAI’s reported direct purchases.

  3. 03

    High-end Macs have reportedly faced months-long sellouts amid a memory-chip shortage.

OpenAI has reportedly bought tens of thousands of Mac minis and Mac Studios for computer-use-agent training, according to The Information. The machines are meant to train agents that autonomously handle multi-step tasks.

The reported deployment has a defined role rather than a general claim about all of OpenAI’s model training. It also supports reinforcement-learning and computer-use-agent work, tying the Apple hardware to the development of systems that can carry out a sequence of actions.

Ownership versus rented access

Anthropic is using the same class of hardware through a different arrangement: it rents Mac minis through AWS for AI work. The contrast is straightforward: OpenAI is reported to be acquiring machines, while Anthropic is accessing Mac capacity through a cloud provider.

A different compact hardware design

The purchases add Apple hardware to AI-lab infrastructure that has often centered on GPUs. Nvidia’s compact DGX Spark takes a different technical route, using dedicated GPUs, CUDA software and Tensor Cores rather than Apple’s unified-memory architecture.

That architectural contrast does not establish which system is better for agent work. Neither account provides a like-for-like measure of the reported Macs’ configuration, cost, performance or scalability against alternative hardware on the same workload.

Demand limits the high end

The most powerful Mac models have reportedly been sold out for months because of a memory-chip shortage. That availability constraint sits alongside OpenAI’s reported interest in obtaining more systems.

For now, the reported fleet is best understood as specialized capacity for computer-use agents, not evidence that AI labs are broadly replacing GPUs. The missing cost, configuration and comparative-performance details matter to judging how broadly this approach can translate beyond that workload.

Sources

  1. the-decoder.comOpenAI and rival AI labs are buying tens of thousands of Mac minis to train computer-use agents
  2. cryptobriefing.comOpenAI acquires thousands of Mac minis, Mac Studios for AI training: The Information

Related coverage

Continue reading