Mac mini Shortage Explained: OpenAI and Anthropic Are Buying Them Up

The starting point of this Mac mini shortage is not ordinary consumers replacing their machines, but bulk purchasing by AI labs. According to The Information, citing internal OpenAI sources, OpenAI has bought tens of thousands of Mac minis and Mac Studios over the past few months for reinforcement learning and computer-use agent development; the same report also says Anthropic is renting large amounts of Mac compute through Amazon AWS. Both pieces of information come from secondhand retellings and are source claims, not official confirmation from Apple or OpenAI.

It is worth noting that OpenAI and Anthropic are not short on GPUs; they own massive Nvidia accelerator clusters and continue to buy data center compute. Therefore, interpreting tens of thousands of Macs as "frontier model main training migrating to Apple devices" does not hold up. A more reasonable inference is: Macs handle parallel interaction, trajectory collection, task verification, local inference and some post-training, while centralized updates of model parameters most likely still happen on data center accelerators.

Why Macs in particular? Computer-use agents need to click buttons, type text, drag files and work continuously across different applications. Training such systems requires a large number of operating system environments that can run in parallel and be reset repeatedly, with the system recording operation trajectories, judging whether results are correct, and then using the feedback for subsequent training. A native macOS environment can avoid deviations between simulated systems and real devices in interfaces, permissions and application compatibility — this is the direct reason Macs were chosen for such tasks.

The Mac mini was originally the cheapest desktop computer in Apple's ecosystem: small, low-power, and able to run long-term without a monitor. As always-on AI agents became popular, these originally ordinary product features suddenly became suitable for data centers and local AI deployments. The Wall Street Journal previously reported that the Mac mini accounted for only about 3% of Apple's U.S. Mac sales the previous year, but after entering 2026 it became a popular host for privatized, round-the-clock agents, with some high-memory versions once seeing delivery delays of weeks or even months.

Apple also seems to have been caught off guard by this demand. According to The Information, citing former Apple enterprise AI product marketing manager Todd Dailey, the Mac's rise in the enterprise AI market was highly accidental, and Apple had not previously built a complete enterprise engineering support and developer relations system around this market. In June this year, Apple held a rare closed-door enterprise event at Apple Park, with Anthropic co-founder Jared Kaplan attending, and the Mac mini became the focus product of the entire event.

Apple subsequently responded in products and pricing. On August 25, Apple released the M6 Mac mini and M5 Ultra Mac Studio, with promotional emphasis clearly shifting to on-device AI, large model inference and multi-machine clusters, and explicitly stating that developers can run and fine-tune large models locally on Macs. Prices rose in tandem: the new M6 Mac mini starts at 6,999 yuan in the Chinese market, while the 2024 M4 Mac mini with 16GB of memory started at 4,499 yuan, an increase of about 56% in the entry price over two years. Chip upgrades are part of the reason, and rising memory and storage costs are also pushing up whole-machine prices.

Demand is already reflected in financial data. Apple's Mac business revenue reached $10.352 billion in the third quarter of fiscal 2026, up about 28.7% year over year, growing faster than other hardware categories in the same period. Demand from enterprises and developers for high-memory Macs has become a variable that cannot be ignored in this round of growth — but it should be noted that this causal link is the report's interpretation, and Apple has not separately disclosed the contribution of AI-related purchases.

The technical reason Macs were chosen centers on memory architecture. Apple's M-series chips use a unified memory architecture, in which the CPU, GPU and neural engine access the same memory pool, reducing repeated movement of data between different memory regions. For local large model inference, this can alleviate the VRAM capacity limits common to consumer-grade discrete graphics cards.

Capacity and bandwidth are key metrics. According to source descriptions, the latest M5 Ultra Mac Studio can be equipped with up to 512GB of unified memory, with memory bandwidth reaching 1.2TB/s; such a large memory pool can fit extremely large quantized models in a single machine; the active cooling design of a desktop also makes it more suitable for high-load tasks lasting hours or even days.

Multi-machine collaboration is another path. According to Apple's official introduction, multiple devices can form a cluster via Thunderbolt 5, and distributed AI inference performance can reach up to 3 times that of a single machine. These figures come from Apple's official statements or source retellings, are manufacturer claims and source claims, and have no independent third-party verification.

Apple's changes have already drawn Nvidia's attention. From DGX Spark to this year's newly announced RTX Spark, Nvidia is compressing data center AI computing power into desktops and laptops. RTX Spark integrates a Grace CPU and Blackwell GPU, provides up to 1 PFLOP of AI computing performance and 128GB of unified memory, and supports CUDA, TensorRT and the RTX software stack; Nvidia directly defines it as a new type of PC for personal agents. According to official plans, the first batch of RTX Spark laptops and compact desktops will go on sale this fall.

The product has not yet officially gone on sale, but supply is already tightening. According to reports, Asus's first batch of RTX Spark models has been fully booked by channel customers, and the first batch of MSI's high-end N1x models is also nearly sold out, with both companies seeking more allocations from Nvidia. These are channel-level source claims and lack official confirmation from Nvidia or the manufacturers.

As a result, Apple and Nvidia stand on the same new track: Apple has large-capacity unified memory, higher energy efficiency and a native macOS environment; Nvidia controls CUDA, TensorRT and a more mature AI development ecosystem, and is deeply integrated with the Windows agent system. The core scenarios both sides are fighting for include local inference, agent execution, reinforcement learning trajectory generation, model evaluation and lightweight post-training. These desktop devices will not replace large training clusters in the short term; what they fill in is a layer that cloud compute cannot fully cover: bringing agents close to user data, operating systems and real applications, and running continuously with lower latency.

Supply problems followed. AI data centers are consuming large amounts of DRAM and NAND, pushing up costs across the entire storage market; at the same time, local AI and agents need larger memory capacity. Upstream data centers and downstream AI terminals are competing for memory at the same time, ultimately forming a chain reaction of price increases, delivery delays and shortages of high-end models.

This chain also explains the nature of the Mac mini shortage: it is not a capacity accident for a single product, but a projection onto end devices of a structural change in memory supply and demand. Delivery delays of weeks or even months for high-memory versions, and rising memory costs pushing up whole-machine prices, are two sides of the same pressure.

The boundaries that need to be preserved are equally clear. The scale of OpenAI's purchase of tens of thousands of Macs and the volume of Anthropic's AWS Mac compute rentals both come from secondhand retellings such as The Information and lack first-party confirmation; Apple has not separately disclosed the contribution of AI purchases to Mac revenue; the actual performance and supply situation of RTX Spark will only be verifiable after it goes on sale in the fall. What can be confirmed is the direction: future AI compute will not be concentrated only in giant data centers; more and more computing tasks will be allocated to local nodes in offices, homes and enterprise server rooms, and this competition has already extended from the cloud to every office desk.