OpenAI Purchases Tens of Thousands of Macs for Agent Training
According to reports from The Information cited by English media, OpenAI has purchased tens of thousands of Mac mini and Mac Studio units for reinforcement learning and computer-using agent training. The procurement focuses on headless units without screens or keyboards, rather than MacBooks.
These machines are used to train agents that can operate computers autonomously, covering multi-step tasks such as coding, testing code, organizing emails, and summarizing documents. Most of the time in reinforcement learning is spent generating rollouts, and unified memory hosts are used for long-term reasoning rather than replacing large model pre-training in data centers.
OpenAI is still urging Apple to expedite delivery, indicating that current inventory is insufficient. Apple has just released the M6 version of the Mac mini and the M5 Ultra version of the Mac Studio, with high-capacity unified memory and multi-machine chaining capabilities directly aimed at this type of workload.
Anthropic has not made equivalent large-scale purchases but instead rents Mac minis through Amazon Web Services (AWS) for similar tasks. AWS itself has provided cloud hosts based on Mac mini and Mac Studio, allowing labs to access macOS environments without having to build their own server rooms.
Apple executives have previously attributed the high demand for Mac mini and Mac Studio to AI and agent demand exceeding expectations, with high memory specifications being out of stock for months. On the developer side, there are also practices of clustering multiple Macs using Thunderbolt interfaces to run larger models locally.
In market mechanisms, the buyers are labs needing massive real computer interface rollouts, while the sellers are consumer-grade hosts with unified memory and cloud-based Macs. This event is driven by agent training demand, with funds flowing from NVIDIA clusters to Apple desktop machines and AWS Mac instances; benefiting are Apple's high-spec Mac shipments and cloud Mac rentals, while pressured are already tight high-capacity memory configurations and paths that tie all training to discrete graphics cards.
At the same time, NVIDIA still views Apple as a major competitor in local AI inference, with consumer-grade AI computers and lab purchases overlapping on the same memory shortage chain.
Source: Public Information
ABAB AI Insight
Apple does not specifically make training cards but relies on the M series unified memory to bring consumer hosts into labs. OpenAI's decision to purchase only headless Mac mini and Mac Studio indicates a need for stackable macOS sandboxes rather than office computers. This follows Tim Cook's public acknowledgment that "agent demand has pierced inventory," with lab orders turning retail shortages into B2B market conditions.
The capital path is divided into two segments: OpenAI uses its own funds to buy out the hosts and locks the environment in its own cabinets; Anthropic opts for AWS rentals, converting capital expenditures into operational expenses. Pre-training still occurs on NVIDIA and cloud accelerators, while Macs handle the highest proportion of inference rollouts in reinforcement learning. The money shifts from "renting a training card" to "replicating ten thousand real computers."
Similar to how Google once stacked ordinary servers for search and Tesla used a mix of consumer-grade and data center GPUs for simulations before Dojo, the current phase is about replicating environments for agent training, not replacing H100 as base models. Local AI computers like NVIDIA DGX Spark are simultaneously being cleared from channels, indicating that the same demand is also seeking non-Apple outlets.
Structurally, this represents a technological substitution. The mechanism is that computer-using agents must trial and error in real operating systems, with unified memory reducing VRAM transport, and the unit price and power consumption of consumer hosts being lower than equivalent VRAM workstation clusters. Whoever can stack "ten thousand desktops" faster can turn clicks, typing, and window switching into training signals more quickly.
ABAB News · Cognitive Laws
To train agents, you first need ten thousand computers
Unified memory turns consumer hosts into training components
What is lacking is not models, but replicable real environments