Thousands of Macs Sold for AI, Here's Why OpenAI and Anthropic Are Buying

Competition in AI infrastructure it turns out that it doesn't only take place in data centers that are filled with Nvidia GPU. 

Recent reports suggest OpenAI has purchased tens of thousands of Mac minis and Mac Studios to support its work.reinforcement learning and development of computer-use agents, while Anthropic is said to be using a Mac mini rented through Amazon Web Services.

This step is interesting becauseAI trainingover the past few years has been synonymous with large-scale GPU clusters.

However, the computational requirements for training agents that interact with computer interfaces are different from those for pretraining giant language models.

In other words, the newsOpenAI buys Macdoesn't mean the company is abandoning GPUs.

Macinstead, it fills a specific type of workload that requires multiple independent machines, large memory, a macOS environment, and the ability to run agentic processes continuously.

Key Takeaways

• OpenAI reportedly purchased Mac minis and Mac Studio for reinforcement learning and computer-based agents. Anthropic is using Mac minis through AWS.

• Apple Silicon offers unified memory, high efficiency, and desktop configurations suitable for running multiple AI workloads in parallel.

• Macs aren't a complete replacement for datacenter GPUs. Their role is more relevant for agentic workloads, local inference, experiments, and certain reinforcement learning scenarios.

OpenAI Buys Macs for Tens of Thousands

Mac Mini | Source: apple.com

The Information reports that OpenAI has acquired tens of thousands ofOpenAI Mac mini And OpenAI Mac Studioin the last few months.

Sources that know the needsThe company's computing arm says the device is used for reinforcement learning and AI development that can use computers directly.

The exact number has not been officially announced by OpenAI nor Apple. 

Therefore, the figure of “tens of thousands” should be understood as a report based on sources, not a procurement figure that has been confirmed through corporate statements.

Anthropic is on the same trend, but its procurement model is different.

The company reportedly rentedAnthropic Macmini via AWS for similar workloads,so the headline about AI companies “buying up Macs” doesn’t mean they’re both making physical purchases in identical patterns.

Why Does AI Training Need a Mac?

The answer lies inda typetraining AIwhat is being done.

Training foundation modelfrom scratch typically requires thousands of GPUs connected via a very high-speed network because a single large training process must be split across many accelerators.

Reinforcement learning for AI agents can take different forms. An AI lab might run thousands of computer sessions in parallel, then have agents attempt a task, receive feedback on the results, and refine their behavior through an iterative process.

OpenAI itself previously explained that Computer-Using Agent uses a combination of supervised learning and reinforcement learning.

The model is trained to read screen displays and use the cursor and keyboard, then reinforcement learning helps improve reasoning, error correction, and adaptation when conditions change.

Read Also:AI Boom Pushes Micron, SanDisk, and SK Hynix

Computer-Use Agents Need Many Real Environments

Computer-use agents different from chatbots which only produce text.

This type of agent must understand screenshots, determine the next action, click buttons, type, scroll, and complete a series of steps until the task is complete.

OpenAI explains that Computer-Using Agents can interact with graphical user interfaces without requiring specific APIs for each site or operating system.

These capabilities form the foundation for agents who can handle forms, web navigation, and multi-step digital workflows.

To improve such capabilities, researchers need multiple execution environments where agents can try actions repeatedly.

This is one of the reasonsAI computing it doesn't always have to be one giant GPU cluster;thousands of desktops that can run standalone environments also have strategic value.

Read Also:Will Nvidia Stock Rally Again or Correct?

Why Mac Mini and Mac Studio, Not MacBook?

The Mac mini and Mac Studio are desktops without a built-in display, battery, keyboard, or trackpad.

For large-scale deployments, this design is more practical because companies only purchase the computing components they absolutely need.

Desktop devices also have cooling systems that are better suited to prolonged workloads than thin laptops.

The Information highlights that the Mac mini and Mac Studio can sustain complex AI work for long periods without the same thermal throttling characteristics as portable devices.

Apple itself is now openly positioning the Mac mini as a device for agentic computing which can run continuously.

Apple claims that the M6 ​​generation Mac mini has AI performance up to four times faster than the previous generation.

Read Also:Top 10 AI Crypto Coins to Watch in 2026

Unified Memory Becomes Apple's Weapon for AI Computing

A key advantage of Apple Silicon is unified memory. The CPU and GPU share the same memory pool, so data doesn't have to be constantly moved between separate system RAM and VRAM, as in traditional computer architectures.

This configuration is interesting when the model or context requires large memory.

The latest Mac Studio M5 Ultra can even be configured with up to 512 GB of unified memory and bandwidth of around 1.2 TB per second.

Apple also supports clustering multiple Mac Studios via Thunderbolt 5 and RDMA for distributed AI inference.

This expands the Mac's position from being just a developer's computer to a potential part of the AI hardware on both team and enterprise scale.

If you are interested in following the development of AI technology which also often overlaps with blockchain and digital asset narratives, you can register on Bittime to get to know the crypto market.

Can a Mac Replace an Nvidia GPU?

Not for all needs. Data center GPUs like Nvidia's platforms are still better suited for training foundations for very large-scale models that require extreme parallel throughput and interconnectivity.

Macs are more attractive when workloads require multiple standalone nodes, large unified memory, the macOS ecosystem, or agentic experimentation.

That is, recent developments indicate diversificationAI infrastructure, not total shift from Nvidia to Apple.

This phenomenon also coincides with the high demand for desktop AI devices.

Wccftech reports early batches of Nvidia RTX Spark-based PCs from a number of manufacturers as well experienced strong demand, demonstrating that local and desktop needsAI computing is expanding beyond traditional data centers.

Read Also:NVIDIA Releases Nemotron 3 Ultra, America's Most Powerful Open-Weight AI Model Yet

Reinforcement Learning Changes AI Hardware Requirements

News OpenAI Mac Studio and the Mac mini show that the AI ​​hardware battle is entering a new stage.

When the company started buildingAI agentswhich is not only answer but also take action, the infrastructure also needs to change.

Machines that can run multiple application environments in parallel are becoming increasingly important.

Because of this, consumer computing devices and workstations can take on roles previously occupied by dedicated servers.

Read Also:What Are AI Cryptocurrencies? 9 Popular AI Coins to Watch in 2026

Conclusion

News OpenAI buys MacIt's not just a story of a tech company buying computers in bulk.

The move represents an architectural change.AI training, especially when reinforcement learning and computer-based agents require thousands of computing environments that can run in parallel.

Apple Silicon offers a combination of unified memory, local performance, efficiency, and a desktop design that is well-suited for these workloads.

On the other hand, Anthropic uses a cloud-based approach by renting Mac minis through AWS.

Data center GPUs remain an important foundation for training large models.

However, the era of AI agents creates a definitionAI hardware increasingly wider: starting from super clusters to the Mac mini working as a node in an agentic computing network.

Bittime is a licensed and regulated Digital Financial Asset Trader (PAKD) supervised by Indonesia’s Financial Services Authority (OJK) — where you can buy Bitcoin in Indonesia and hundreds of other crypto assets starting from just Rp10,000. The registration process is fast, secure, and you can get started today.

Track USDT to IDR conversions and monitor your favorite crypto assets in real time. Everything is available in one crypto investment app that you can download for free on the Play Store

Ready to start? Register now on Bittime and execute your investment strategy with a platform trusted by millions of users in Indonesia.

FAQ

Is it true that OpenAI bought tens of thousands of Macs?

According to a report by The Information, OpenAI has purchased tens of thousands of Mac minis and Mac Studios for reinforcement learning and computer-based agents. OpenAI has not yet released official procurement figures.

Did Anthropic also buy thousands of Macs?

Available reports indicate that Anthropic leased the Mac mini through Amazon Web Services. Therefore, Anthropic's hardware usage model differs from the outright purchases reported by OpenAI.

Why is OpenAI using a Mac mini?

The Mac mini offers Apple Silicon with unified memory, a compact design, desktop cooling, and a more efficient device cost because it doesn't include a display or battery. These characteristics are ideal for high-performance, agentic workloads.

What is the relationship between reinforcement learning and computer-use agents?

Reinforcement learning helps agents learn from successes and failures while performing tasks. In computer-based agents, this process can improve reasoning, error-correcting abilities, and adaptation when dealing with computer interfaces.

Will Mac replace Nvidia for AI training?

Not entirely. Nvidia remains strong for pretraining and large-scale datacenter workloads, while Macs have the potential to play a larger role in local AI, agentic computing, inference, and certain types of reinforcement learning.