Skip to main content

OpenAI Bought Tens of Thousands of Macs for RL: Apple Just Cut Into Nvidia's Turf

In late August 2026, a headline that seemed to defy common sense swept the tech world: OpenAI is hoarding Macs instead of GPUs. Citing internal OpenAI sources, The Information reported that over the past few months, OpenAI has purchased tens of thousands of Mac mini and Mac Studio units specifically for reinforcement learning and computer-using agent development. At the same time, Anthropic has been renting large amounts of Mac compute capacity through Amazon AWS.

On one side, memory chip prices are soaring and ordinary consumers can barely afford a new computer; on the other, AI-friendly Macs are selling out, with high-end configurations in short supply. Apple's devices are quietly transforming from "personal computers" into "AI infrastructure," quietly taking business from Nvidia on the desktop. In this article, we break down how this "Apple cuts into Nvidia" story happened and what it means for the AI compute landscape.

What Happened: Why OpenAI Bought Tens of Thousands of Macs​

Let's start with the facts. According to The Information, OpenAI has concentrated its purchases over recent months, buying tens of thousands of Apple Mac minis and Mac Studios, deployed in clusters to train and test agents that can operate computers autonomously — the so-called computer-using agents.

These agents need to click buttons, type text, drag files, and work continuously across different applications. OpenAI's previously released Computer-Using Agent learns through reinforcement learning to operate graphical interfaces directly. Training such systems requires a large number of operating-system environments that can run in parallel and be reset repeatedly: the agent performs tasks, the system records the action trajectory, judges whether the result is correct, and feeds that feedback into subsequent training.

The key insight: to train an agent that "operates a computer," it's best to run it on real computers. Macs provide a native macOS environment, avoiding interface, permission, and app-compatibility deviations between a simulator and real devices. For OpenAI, tens of thousands of Macs form a massive "agent training ground" — a scenario Nvidia GPU clusters can't directly replace.

Did Nvidia Really Lose the Job? Let's Clarify What the Macs Actually Do​

The phrase "Nvidia got cut into by Apple" can easily mislead people into thinking OpenAI abandoned GPUs. In reality, we need a more precise picture of the division of labor.

Buying tens of thousands of Macs does not mean OpenAI's frontier model pretraining has moved from Nvidia GPU clusters to Apple devices. The Macs are far more likely to handle parallel interaction, trajectory collection, task validation, local inference, and some post-training work. Centralized parameter updates probably still run on data-center accelerators.

In other words, the Macs are filling a layer that cloud compute struggles to cover: keeping agents close to user data, operating systems, and real applications, running continuously at low latency. Nvidia hasn't lost its "job" wholesale — but in new scenarios like RL trajectory generation, agent execution, and model evaluation, Apple has clearly become a competitor Nvidia can't ignore. Nvidia itself now views Apple as its top rival in the local AI space.

Why Apple Works: Unified Memory + Multi-Machine Clusters + Native Environment​

Apple's machines hold their ground in this fight thanks to three real strengths.

First, unified memory architecture. Apple's M-series chips let the CPU, GPU, and neural engine access the same memory pool, reducing data shuttling between different memory regions. For local LLM inference, this relieves the VRAM limits common with consumer discrete GPUs. The latest M5 Ultra Mac Studio can be configured with up to 512GB of unified memory at 1.2TB/s bandwidth — a memory pool large enough to load extremely large quantized models on a single machine.

Second, multi-machine coordination. Apple has strengthened multi-machine collaboration in the new Mac Studio: multiple devices can form a cluster over Thunderbolt 5, with distributed AI inference performance up to 3x a single machine. For tasks that must run under sustained high load for hours or days, the desktop's active cooling is also more reliable than a laptop's.

Third, the native macOS environment. For training agents that "operate a computer," the native environment means real interfaces, real permissions, and real compatibility — something no simulator can provide.

Apple's Accidental Success: From a "3% Niche" to an "AI Hit"​

What's dramatic is that Apple itself wasn't fully prepared for this demand.

The Mac mini was once the cheapest desktop in Apple's ecosystem: small, low-power, able to run headless for long periods. According to The Wall Street Journal, the Mac mini accounted for only about 3% of Apple's US Mac sales last year. But entering 2026, with the rise of resident AI agents, it suddenly became the go-to host for private, always-on agents, and some high-memory configurations saw delivery delays of weeks or even months.

The Information quoted Todd Dailey, Apple's former enterprise AI product marketing manager: the Mac's rise in the enterprise AI market was largely accidental, and Apple had not built a complete enterprise engineering support and developer-relations system around it.

Apple has since started taking the business seriously. In June, Apple held a rare closed-door enterprise event at Apple Park, attended by Anthropic co-founder Jared Kaplan, with the Mac mini as the centerpiece. On August 25, Apple released the M6 Mac mini and M5 Ultra Mac Studio, with marketing clearly shifted toward on-device AI, large-model inference, and multi-machine clusters, explicitly stating that developers can run and fine-tune large models locally on Mac.

The demand shows up in the financials. In Apple's fiscal Q3 2026, Mac revenue reached $10.352 billion, up about 28.7% year over year — the fastest growth of any hardware category. Enterprise and developer demand for high-memory Macs has become a variable you can no longer ignore in this growth.

Nvidia's Counterattack: RTX Spark Packs the Data Center Into the Desktop​

Facing Apple's advance, Nvidia isn't standing still. From the DGX Spark to the newly announced RTX Spark, Nvidia is compressing data-center AI compute into desktops and laptops.

The RTX Spark integrates a Grace CPU and Blackwell GPU, offering up to 1 PFLOP of AI compute and 128GB of unified memory, while supporting CUDA, TensorRT, and the RTX software stack. Nvidia directly positions it as a new PC for personal agents. The first RTX Spark laptops and compact desktops are planned for release this fall.

Supply is already tightening before the product even hits the market. Reports say Asus's first batch of RTX Spark models has been fully pre-booked by channel customers, and MSI's high-end N1x machines are nearly sold out; both companies are pressing Nvidia for more allocation.

Apple and Nvidia have thus landed on the same new track: Apple brings large unified memory, better energy efficiency, and a native macOS environment; Nvidia brings CUDA, TensorRT, a more mature AI development ecosystem, and deep ties to the Windows agent stack. The core battlegrounds are local inference, agent execution, RL trajectory generation, model evaluation, and lightweight post-training.

The Ripple Effect: The Supply-and-Demand Truth Behind the Memory Shortage​

This "desktop AI war" also exposes a bigger issue — memory supply.

AI data centers are consuming massive amounts of DRAM and NAND, driving up storage costs; meanwhile, local AI and agents need even larger memory capacities. Upstream data centers and downstream AI endpoints are fighting over the same memory, creating a chain reaction of rising prices, delivery delays, and high-end stockouts. Surging memory chip prices, the Mac mini's high-end shortage, and the RTX Spark's first-batch sellout all trace back to the same strained supply chain.

Outlook: The Next Stop for AI Compute Is Every Desk​

The Mac mini's rise reveals a new hardware path for the agent era: future AI compute will not concentrate only in giant data centers — more and more tasks will be distributed to local nodes in offices, homes, and enterprise server rooms.

This also means the next battle in AI hardware has already extended from the cloud to every desk. The "desktop war" between Apple and Nvidia is just beginning.

Sources​

  • Jiqizhixin (机器之心): The real reason Mac minis are out of stock — OpenAI and Anthropic are grabbing them (2026-08-31)
  • Qbitai (量子位): OpenAI bought tens of thousands of Macs for RL training — Nvidia's job got taken by Apple (2026-08-31)
  • The Information: Apple Stumbled Into AI Hardware Success
  • ZDx (智东西): Qualcomm engineering SVP on HBC vs HBM