OpenAI buys tens of thousands of Macs for reinforcement training! Apple has snatched the business that originally belonged to NVIDIA.

OpenAI was exposed to be frantically hoarding Macs, purchasing tens of thousands of units at a time.

They even bought out the stock, and are still trying every means to get more units.

They do not buy MacBook laptops, but specifically purchase Mac mini and Mac Studio that come with no screen and no keyboard.

Here comes the question: what kind of AI business cannot be handled by NVIDIA GPUs and Google TPUs, and has to rely on Macs?

Tens of thousands of Macs are used for reinforcement learning, and Anthropic has followed suit

According to The Information, OpenAI has purchased tens of thousands of Mac mini and Mac Studio units, which are specially used for reinforcement learning.

It is not just OpenAI, Anthropic is also renting Mac mini units through Amazon Web Services (AWS) to complete similar tasks.

These Macs are used to train "computer-use agents", which are AI systems that can operate computers independently, and complete multi-step tasks such as editing and testing codes, automatically organizing emails, and summarizing documents.

This boom is directly reflected in Apple's financial reports.

In the most recent quarter, Mac's sales revenue increased by nearly 29% year-on-year to reach 10.3 billion US dollars. Its growth rate outperformed all other Apple product lines including iPhone and iPad, making Mac the fastest-growing business of Apple.

On June 23, Apple held an event named "Business at the Park" at its headquarters Apple Park. This is not a common practice in Apple's history. Apple has always focused on the consumer market and rarely holds events specifically for enterprise customers.

Executives from Disney and Ford attended this event, so did Jared Kaplan, co-founder of Anthropic. Tim Cook, Apple's outgoing CEO, and John Ternus, the incoming CEO, were both present at the event.

According to an attendee, Apple repeatedly emphasized at the event that its hardware is very suitable for processing AI tasks locally, and Mac mini was the absolute focus of the whole event.

AI training has long been dominated by NVIDIA GPUs, but Macs can be purchased in large quantities in the niche scenario of reinforcement learning, thanks to their unified memory architecture.

The video memory of NVIDIA GPUs is separated from the system memory, which creates a bottleneck when data is transmitted between the two. Apple's M-series chips adopt a single shared memory pool, where the CPU and GPU directly access the same block of memory, giving them performance advantages when processing AI workloads.

In addition, different from the lightweight MacBook, Mac mini and Mac Studio are equipped with dedicated heat dissipation systems, so they will not reduce their operating frequency due to overheating when running complex AI tasks for a long time. This is extremely critical for reinforcement learning training that needs to last for hours or even days.

Apple is also promoting the EXO Labs open source software project, which can cluster multiple Mac units to run trillion-parameter level AI models locally.

The newly released Mac Studio from Apple also specially emphasizes the cluster capability. Multiple Mac Studio units can be connected in series to form a more powerful system to run cutting-edge models.

The timing of this new product release is also unusual. Apple usually releases year-end updates for its Mac product line in October or November every year, but this time it was advanced to August.

NVIDIA has set its sights on this market, and Apple is responding hastily

The rise of Mac in the local AI field has already attracted the attention of NVIDIA.

According to an insider who has discussed the competitive situation with NVIDIA executives, NVIDIA regards Apple as its largest competitor in the local AI sector.

At the end of last year, NVIDIA launched DGX Spark, an AI desktop computer with a design style similar to Mac mini, directly targeting this market.

Apple is facing the practical problem that its supply cannot keep up with demand.

The huge demand for memory chips from AI data centers has led to a historic industry-wide shortage, and Apple is not immune to this situation.

The high-spec versions of Mac mini and Mac Studio that are the most attractive to AI developers have been out of stock for several months.

Todd Dailey, former enterprise marketing manager for AI products at Apple, revealed that due to the limited supply of Macs over the past year, some enterprises have begun to look for alternative solutions. NVIDIA DGX Spark is a frequently mentioned option, and it is available in stock now.

Dailey left Apple in April this year and is now an independent AI consultant. He also revealed that the popularity of Mac in the enterprise AI market is completely accidental, rather than an active plan of Apple. Apple does not have a dedicated engineering team for enterprise customers, nor does it have any employees focusing on developer relations.

The last server product Apple sold was Xserve, which was discontinued back in 2011. The Mac-based server operating system also stopped development in 2022.

However, some people have already sensed the business opportunity.

Peter Voell, a former OpenAI computing infrastructure employee, founded Mount Thor, a cloud computing company based on Apple hardware. It is still in stealth mode, and its official website describes its product as "an AI execution environment built on Apple hardware".

Apple is also counting on partners like Mount Thor and webAI to push Macs deeper into the enterprise market.

Apple has finally started to build its own servers with Mac chips recently. However, these servers are only for internal use, supporting the Private Cloud Compute service to process AI tasks that exceed the processing capacity of iPhone or Mac.

Some enterprise customers have asked Apple if they can purchase the access rights of these servers, but Apple has rejected all such requests so far.

References: [1]https://www.theinformation.com/articles/apple-stumbled-ai-hardware-success-mac[2]https://www.apple.com/newsroom/2026/07/apple-reports-third-quarter-results/

This article is from the WeChat official account "QbitAI", the author focuses on cutting-edge technologies, and 36Kr publishes it with authorization.