OpenAI Splashes Out on Tens of Thousands of Macs for AI Training — Apple Makes a Surprise Incursion into Nvidia's Turf

The hardware landscape for AI lab training is undergoing an unexpected shift. According to The Information, citing sources inside OpenAI, the AI giant has purchased tens of thousands of Mac mini and Mac Studio units specifically for reinforcement learning training and the development of computer-use agents. Anthropic is likewise renting substantial Mac compute capacity through Amazon Web Services (AWS). Apple computers — once viewed as consumer-grade products — are now becoming the compute infrastructure of choice for Silicon Valley's top AI labs.

These Macs are not ordinary office equipment. Sources say OpenAI's shopping list consists exclusively of Mac mini and Mac Studio units without screens or keyboards — not portable MacBooks. They are being used to train AI systems capable of autonomously operating computers. Such agents can handle multi-step tasks like editing test code, automatically organizing inboxes, and summarizing documents, placing extreme demands on sustained compute output and memory bandwidth. Demand is so intense that OpenAI has been repeatedly pressing Apple for faster deliveries.

The core advantage Mac holds in this AI buying spree stems from Apple's custom silicon and its Unified Memory Architecture. In traditional Nvidia GPU setups, VRAM and system memory are separate, and moving data between them creates significant performance bottlenecks. Apple's M-series chips, by contrast, let the CPU, GPU, and Neural Engine share a single memory pool, so model parameters and intermediate data don't need to be repeatedly copied between different memory regions. Take the M5 Ultra as an example: its unified memory capacity reaches 512GB, enough to directly load quantized models with tens of billions — or even some exceeding a hundred billion — parameters.

Thermal design is equally critical. Unlike the thin-and-light MacBook, the Mac mini and Mac Studio are equipped with dedicated fans and thermal modules that sustain stable performance during reinforcement learning training runs lasting hours or even days, without throttling due to overheating. Apple is also promoting the EXO Labs open-source project, which supports clustering multiple Macs to run trillion-parameter AI models locally. The new Mac Studio released in August specifically highlighted multi-machine clustering capabilities.

The explosion in enterprise demand is directly reflected in Apple's financials. In the most recent quarter, Mac revenue grew nearly 29% year-over-year to $10.4 billion, outpacing every other product line including iPhone and iPad, making it Apple's fastest-growing hardware business.

Apple's response to this shift has been somewhat caught off guard. Todd Dailey, Apple's former enterprise marketing manager for AI products who left the company in April, said bluntly that Mac's success in the enterprise AI market was entirely accidental rather than the result of deliberate planning. Apple has no dedicated engineering team for enterprise customers and lacks a mature developer relations apparatus. On June 23, Apple hosted a closed-door enterprise event called "Business at the Park" at Apple Park — an unusual move in the company's history. Executives from Disney and Ford attended, as did Anthropic co-founder Jared Kaplan. Outgoing CEO Tim Cook and John Ternus, widely seen as his successor, both made appearances. According to attendees, the Mac mini was the centerpiece of the event, with Apple repeatedly emphasizing its hardware's advantages in on-device AI workloads.

Nvidia has clearly taken notice of this undercurrent. According to a source who has discussed the competitive landscape with Nvidia executives, Nvidia now regards Apple as its biggest competitor in the local AI space. Late last year, Nvidia launched the DGX Spark desktop AI computer, adopting a boxy form factor similar to the Mac mini, powered by a combination chip featuring a Grace CPU and Blackwell GPU, with support for up to 128GB of unified memory. The product is squarely aimed at the desktop AI inference market.

Product	Chip Solution	Max Unified Memory	US Starting Price	Launch/Availability
Mac mini (M6)	Apple M6	32GB	$899	Pre-orders Aug 25, 2026; on sale Sep 22
Mac Studio (M5 Ultra)	Apple M5 Ultra	512GB (available October)	From $5,499 (512GB version approximately $15,000)	Announced Aug 25, 2026; on sale Sep 22
Nvidia DGX Spark	Grace Blackwell GB10 (20-core CPU + Blackwell GPU)	128GB LPDDR5x	$4,699 (launch price $3,999 in Oct 2025)	Available Oct 2025

▲ All three products embrace the "unified memory" approach, but with clear positioning tiers: the Mac mini targets value, the Mac Studio aims at professional workstation-class local inference, and the DGX Spark is Nvidia's desktop AI machine built for developers. (Data source: Apple and Nvidia official websites, August 2026)

Yet Apple's supply chain is becoming its biggest constraint. The explosive demand for memory chips from AI data centers has triggered an industry-wide DRAM and NAND shortage — and Mac is precisely a product that depends heavily on large memory capacities. The high-end Mac mini and Mac Studio configurations most attractive to AI developers have been out of stock for months. Dailey revealed that some enterprises, unable to wait for inventory, have begun pivoting to Nvidia's DGX Spark, which is readily available.

Supply tightness is also pushing prices higher. Take the China-market Mac mini as an example: the M4 version launched at 4,499 yuan (approximately $669) for the 16GB+256GB configuration, but by June 2026 the official price for the same configuration had risen to 5,999 yuan (approximately $893). The M6 version announced on August 25 starts at 6,999 yuan (approximately $1,041). In two years, the entry price for Apple's most affordable desktop has climbed by more than 50%.

Market hunger is spawning new business models. Peter Voell, a former OpenAI compute infrastructure employee, founded Mount Thor, a cloud computing company built on Apple hardware that is still in stealth mode. Its website describes the product as "an AI execution environment built on Apple hardware." Namespace Labs, which provides cloud development environments for AI coding agents, posted a data center expansion video on X showing staff unboxing Macs in bulk and sliding them into server racks. With Mac mini and Mac Studio severely backordered, the company has even resorted to mounting MacBook Pros — screens and keyboards included — directly into racks.

Apple itself is also repositioning in the server space. The company has begun building servers using its own silicon, but strictly for internal use, serving its Private Cloud Compute infrastructure to handle AI tasks that exceed on-device capabilities. Despite inquiries from enterprise customers about purchasing access to these servers, Apple has so far declined every request. The last time Apple sold a server product was the Xserve, discontinued in 2011.

From consumer electronics to AI compute infrastructure, Mac's role transformation has unfolded in barely a year. For Nvidia, Apple's rise in the local AI space means the competitive map now extends from data centers to the desktop. For Apple, this AI lab-driven demand has delivered both an unexpected growth engine and an exposure of its passive enterprise-market positioning and fragile supply chain.