OpenAI's Mac Mini Fleet Reveals a New Training Architecture for Computer-Use Agents

OpenAI has quietly assembled tens of thousands of Apple Mac mini and Mac Studio systems for reinforcement learning and computer-use agent training, The Information reported Saturday — an unusual hardware decision that reveals how one specific category of AI workload has moved beyond the GPU-cluster model that has defined frontier AI infrastructure since 2020. Anthropic is pursuing the same approach through a different procurement path, renting Mac minis at scale through Amazon Web Services. Neither company publicly confirmed the arrangement, but the secondary evidence is already visible: Apple's Mac business posted $10.35 billion in revenue last quarter, up 28.7% year over year, the fastest growth rate of any Apple hardware category in a period when iPad revenue actually declined.

The story lands on a historically freighted day. Tim Cook steps down as Apple CEO at midnight, handing the company to hardware engineering chief John Ternus on September 1, per Apple's announcement. The transition arrives as Apple's desktop products have become, by accident rather than design, a meaningful piece of the infrastructure underlying some of the most advanced AI development in the world. Ternus inherits a Mac business that is winning enterprise deals Apple never set out to chase, in a market Apple has no dedicated team to serve.

Why Reinforcement Learning for Computer-Use Agents Doesn't Need a GPU Cluster

The architectural logic behind OpenAI's decision requires understanding what makes computer-use agent training different from the frontier model pretraining that runs on Nvidia's GPU clusters.

Pretraining a large language model is a matrix-multiplication-dominated workload: enormous weight matrices are multiplied against enormous data batches, continuously, for months. Every GPU in a training cluster is doing the same operation in lockstep on partitioned shards of the same model and dataset. The bottleneck is raw compute throughput and the bandwidth between GPUs, which is why Nvidia's NVLink interconnects and HBM3e memory have been so commercially important. The entire economic case for Nvidia's AI infrastructure products rests on this model of training.

Computer-use agent reinforcement learning is structurally different in a way that matters. An agent is placed inside a real operating system environment — macOS, in this case — given a task (organize an inbox, edit a codebase, summarize a folder of documents), observed as it takes actions, and scored on the result. That score becomes the reward signal for updating the agent's policy. The loop then repeats, millions of times, across thousands of parallel instances.

The bottleneck here is not matrix math. It is the overhead of running real operating system environments at scale and efficiently moving observation data — screen states, UI element trees, file system state — between the environment and the training framework. On a system with a discrete GPU, this data must cross the PCIe bus from the CPU's memory domain into the GPU's VRAM and back again with every inference step. When the environment is a full macOS session rather than a synthetic gym, that overhead is not trivial.

Apple's M-series unified memory architecture collapses this boundary. The CPU, GPU, and Neural Engine all operate from a single physical memory pool on the same chip package. The macOS process running the simulated environment and the MLX training framework running the agent update both read from and write to the same memory without any bus transfer. For the specific workload of training an agent that interacts with a real computer, this design is a genuine advantage over a discrete GPU cluster where CPU and GPU are memory-isolated by the PCIe bus.

There is also a straightforward parallelism argument. Running 50,000 independent Mac units simultaneously, each hosting one or a small number of agent training environments, scales horizontally without inter-node communication overhead. GPU clusters are optimized for tightly coupled parallelism on a shared model; agent training on macOS environments benefits more from loosely coupled breadth across many independent machines. A single A100 GPU cluster running tens of thousands of parallel macOS virtual machines would require routing display, input, and file system state across PCIe for every agent step — at latencies that accumulate quickly in a training loop running millions of iterations.

There is a secondary advantage that rarely appears in coverage of this story: energy economics. A Mac mini with M5 Pro draws roughly 25 to 35 watts under sustained AI load. A GPU server running an NVIDIA H100 draws 700 watts per GPU card, with full system power typically exceeding 10 kilowatts for an 8-GPU node. Running 10,000 Mac minis consumes approximately 250 to 350 kilowatts — comparable to one medium GPU cluster — while providing a much larger number of fully independent macOS environment instances. For the specific task of agent training, where each parallel instance is running independently rather than contributing to a shared computation, the per-environment energy efficiency of Apple silicon is substantially better than discrete GPU infrastructure. Exact figures depend on workload configuration and are not publicly reported by OpenAI or Anthropic, but the directional advantage is real and is part of why AI labs have adopted this approach.

The thermal profile of Mac mini and Mac Studio reinforces this fit. Unlike MacBook, which relies on passive or low-power active cooling and throttles under sustained load, Mac mini and Mac Studio were designed for professional workloads that run for extended periods. The Mac mini's fanless-seeming exterior is deceptive: it contains an active thermal system with a dedicated heat spreader and fan designed for sustained professional use, not the constrained thermal envelope of a portable. Reinforcement learning training runs for hours or days; thermal stability is a practical requirement, not a marketing point.

Apple's M6 and M5 Ultra Arrive Six Weeks Before They Were Expected

Apple's decision to announce new Mac hardware on August 25 — six weeks ahead of its usual October or November window — was not accidental. The Information's reporting suggests Apple moved up the launch in part because demand from AI labs had created supply shortages in the prior generation. High-RAM Mac mini and Mac Studio configurations have reportedly been out of stock for months, leaving enterprise buyers waiting.

The new hardware accelerates the performance story for this specific workload. The M6 chip — Apple's first built on TSMC's 2-nanometer N2 process, using gate-all-around nanosheet transistors rather than the FinFET design in M1 through M5 — debuts in the refreshed Mac mini, which starts at $899. Apple claims a nearly 30% increase in peak GPU compute for AI tasks compared to M5, driven by both the denser N2 process and a new Neural Accelerator embedded in each of the chip's 12 GPU cores — the first time this design has appeared at the Mac mini tier. A Dual 16-core Neural Engine handles tasks that benefit from dedicated matrix-multiply hardware, while the Neural Accelerators in the GPU handle operations in the middle of graphics workloads. The combined effect, per Apple's own testing, is approximately 4x faster AI performance than the M4 Mac mini and up to 4.8x faster LLM prompt processing in LM Studio.

The Mac Studio update is more architecturally significant for the AI lab use case. The M5 Ultra is the first quad-die chip in Apple's M-series history: where prior Ultra-tier chips fused two M-series Max dies, the M5 Ultra connects two dual-die M5 Max chips using a next-generation UltraFusion interconnect running at more than 4.4TB/s, with more than six times the connection density of prior UltraFusion implementations. The result is a 36-core CPU, an 80-core GPU (the largest Apple silicon GPU ever produced, with Neural Accelerators in every core), and a 512GB unified memory ceiling at 1.2TB/s of bandwidth. That 512GB configuration — enabling a single machine to hold 200B-parameter models entirely in unified memory — is the specification that matters most for inference-side research workflows, though Apple has confirmed that 512GB units will not be available until late October, with the M5 Ultra Mac Studio starting at $5,499 for configurations currently available.

Read more: NVIDIA AI server prices rise as memory costs hit the AI stack

The Software Stack That Makes Mac-Based Agent Training Viable

Hardware architecture is only part of the explanation. Two software layers have made Mac-scale AI training practically feasible for engineers who want to move beyond single-machine experimentation.

MLX, Apple's open-source array framework for Apple silicon, provides PyTorch-style automatic differentiation and GPU acceleration via Metal Performance Shaders — natively on unified memory, without the memory-copy overhead that PyTorch's MPS backend still carries in some training configurations. For research teams writing custom RL training loops for agent workloads, MLX is increasingly the default framework on Apple hardware.

EXO, the open-source distributed inference project from EXO Labs, allows multiple Apple Silicon Macs to pool their unified memory into a single distributed inference cluster via Thunderbolt 5 and RDMA (Remote Direct Memory Access). A four-machine EXO cluster delivers roughly 3.2x faster inference than a single machine and, with RDMA over Thunderbolt 5, inter-device latency falls to near-zero. The practical result: four M5 Ultra Mac Studios in a Thunderbolt cluster present approximately 2 terabytes of accessible unified memory, enough to hold and run frontier-scale open-weight models that would require an enterprise GPU cluster at cloud prices. The project carries an Apache 2.0 license and has accumulated more than 46,000 GitHub stars.

Apple has also released Core AI, a new developer framework that coordinates LLM and custom model deployment across all compute surfaces of Apple silicon — CPU, GPU, Neural Engine — simultaneously, rather than routing all operations through a single processor. Core AI is the layer that makes clustered Mac Studio inference viable at the production level rather than only in research contexts.

One important caveat about the software stack deserves direct acknowledgment. PyTorch's MPS (Metal Performance Shaders) backend — which allows PyTorch code to run on Apple's GPU via the Metal API — is still maturing. Academic research from late 2024 and early 2025 documented training-side optimization gaps compared to PyTorch's CUDA backend: operations like gradient checkpointing, mixed-precision training, and certain custom CUDA kernels used in frontier model training pipelines have no equivalent in the MPS backend yet. Research groups at several universities found throughput gaps of 1.5x to 3x for certain training configurations on Apple silicon compared to equivalent NVIDIA hardware, partly attributable to MPS optimization immaturity rather than the underlying compute capacity of the M-series chip.

MLX partially addresses this by providing a native Apple Silicon array framework that does not depend on the MPS backend at all, using Metal directly through Apple's own implementation. For custom RL training loops — where the engineer controls the entire stack — MLX can match or exceed what MPS-backed PyTorch delivers. But for labs that use existing GPU-optimized training libraries, porting to the Mac stack requires engineering effort that a straight GPU procurement does not. This is a real adoption barrier for teams that are not starting from scratch. The choice to use Mac hardware at tens of thousands of units suggests OpenAI and Anthropic have either adapted their agent training pipelines to MLX or found that their specific RL workloads are not bottlenecked by the training-side operations where MPS lags CUDA.

Anthropic's use of AWS Mac instances rather than direct hardware purchase reflects a different operational philosophy. AWS EC2 Mac instances provide M-series hardware virtualized through AWS's infrastructure stack, with on-demand pricing that varies by instance generation and region. For Anthropic, cloud rental avoids capital expenditure on hardware that could become obsolete, while maintaining access to the same macOS environments OpenAI is training on. The tradeoff is ongoing per-hour cost versus the fixed amortization OpenAI has accepted by buying outright.

Apple Has No Enterprise Team for the Enterprise Business It Is Running

The most revealing detail in The Information's reporting is not the scale of OpenAI's purchase but what Apple lacks on the supply side of this demand.

Todd Dailey, who served as Apple's AI Product Enterprise Marketing Manager before leaving in April 2026 to work as an independent AI consultant, reportedly described the Mac's rise in enterprise AI as entirely unplanned. Apple has no engineering team dedicated to business customers and no staff focused on developer relations for the AI workload category that is driving the Mac's fastest revenue growth in years. When enterprise customers approached Apple about purchasing access to its Private Cloud Compute server infrastructure — the same Apple Silicon-based servers Apple uses for its own on-device AI processing — Apple declined.

The structural absurdity of this situation is worth stating precisely: Apple is generating hundreds of millions of dollars in unplanned B2B hardware revenue from AI labs while simultaneously turning away those labs when they ask for something more than the hardware Apple already sells to consumers. The company's last server product, the Xserve rack-mount unit, was discontinued in 2011. Its Mac server operating system was end-of-lifed in 2022. Apple is not a server company, does not want to be a server company, and has built its AI infrastructure strategy around Private Cloud Compute as an internal service rather than an external revenue stream.

Apple is instead channeling enterprise AI demand through two third-party partners: Mount Thor, a startup building managed macOS environments for AI workloads on Apple hardware (currently in stealth mode), and WebAI, which provides AI tools built on Apple hardware for enterprise customers. These are not Apple enterprise sales and support operations — they are partnerships that allow Apple to route demand it cannot serve directly to companies that can.

The depth of Apple's enterprise gap becomes more apparent when measured against the infrastructure that typical enterprise AI hardware vendors provide. NVIDIA ships DGX Spark with DGX OS — a pre-configured Linux environment with CUDA, TensorRT, NEMO, and the rest of the enterprise AI software stack pre-installed, validated, and supported under NVIDIA's enterprise service contracts. When a large bank or pharmaceutical company deploys DGX Spark for internal AI research, NVIDIA has field engineers and support structures to handle the deployment. When the same organization deploys Mac Studio clusters for agent training, it is working with hardware designed for a creative professional at a design agency, supported through Apple's consumer support channels. The engineering lift is different.

Apple's Private Cloud Compute servers — which run M-series chips and handle AI processing for Apple's on-device features when iPhone or Mac lacks the compute for a task — represent a proof of concept that Apple can build server-class AI infrastructure at scale. Images of the server hardware surfaced online in late August 2026, showing custom Apple silicon rack-mount systems clearly designed for data center density rather than desktop deployment. Yet Apple has explicitly refused to sell access to these servers, even as enterprise customers have asked. The rationale likely involves Apple's data privacy commitments: Private Cloud Compute was architected specifically around the guarantee that user data is not accessible to Apple, a guarantee that would be complicated by selling server access to third parties who would process other companies' data on the same hardware. But the commercial outcome is the same: Apple has demonstrated it can build what enterprise AI labs want and then declined to sell it.

Read more: Anthropic launches hardware standard to let AI agents control lab and factory equipment

NVIDIA Responded — and Cannot Fully Close the Mac's Architectural Advantage

NVIDIA has observed Apple's accidental infrastructure play closely enough to respond with hardware products specifically designed to compete with it. The DGX Spark — a compact desktop AI computer in a form factor closely resembling the Mac mini, launched in late 2025 — delivers 128GB of unified memory and the full CUDA/TensorRT software stack that Nvidia's data center customers already run. Where high-RAM Mac configurations were reportedly backordered for months, DGX Spark was in stock when Mac mini supply tightened, giving enterprise customers a GPU-native alternative. Nvidia's forthcoming RTX Spark takes the small-form-factor concept further, with first-batch units reportedly pre-sold before public launch.

The competitive dynamic between Nvidia and Apple in this specific workload class is asymmetric in an important way. Nvidia's compact hardware cannot natively run macOS environments. Training a computer-use agent that operates macOS — the macOS that OpenAI and Anthropic are specifically training their agents on — requires actual Mac hardware. A DGX Spark can simulate Linux or Windows agent environments, and it does so with the full CUDA software stack and superior enterprise support. But an AI company that wants its agent to function on the macOS desktops its users actually run will need macOS environments, and that means Apple hardware. Nvidia cannot compete for that portion of the workload regardless of its hardware specifications.

The cleaner competitive battlefield is inference and non-macOS agent training. There, NVIDIA's 128GB unified memory (DGX Spark), full CUDA ecosystem, professional support contracts, and NVLink clustering capability represent a meaningful argument against Mac infrastructure for teams that do not require macOS environments specifically. The PyTorch and CUDA software stacks are more mature for training-side optimizations than Apple's MLX framework in its current state. For teams evaluating open-weight model inference clusters, the decision between a four-Mac-Studio EXO cluster and a DGX Spark installation is a genuine trade-off across price, software ecosystem maturity, and enterprise support availability.

There is a third structural factor that cuts against Apple in the medium term: the global DRAM shortage that has driven Mac mini and Mac Studio out of stock is the same shortage raising the price of high-bandwidth memory chips used in NVIDIA's GPU products. NVIDIA's AI server prices rose more than 15% in recent months as HBM3e costs increased — a dynamic GPTS24 previously reported. Apple and NVIDIA are competing for TSMC capacity and memory supply from the same constrained pool. Apple's advantage is that M-series chips use LPDDR5-class memory rather than the HBM that fills NVIDIA's GPU VRAM, accessing a different supply chain with different constraints. But the global shortage of high-capacity memory in any form has made the high-RAM Apple configurations — 64GB Mac mini M5 Pro, 192GB Mac Studio M5 Max, 512GB Mac Studio M5 Ultra — the exact configurations AI labs need and the exact ones hardest to source. Apple's inability to forecast or buffer this demand has created the supply crunch that gives NVIDIA's DGX Spark its opening.

What Apple's Accidental Infrastructure Revenue Means for Its Next CEO

John Ternus becomes Apple CEO on September 1 inheriting an unusual business problem: a hardware product line that has stumbled into a new commercial category that Apple was not built to serve.

The Mac's 29% revenue growth in the June quarter reflects demand from AI labs and enterprise developers that Apple's consumer product model does not explain. The supply crunch on high-RAM Mac mini and Mac Studio configurations directly resulted from Apple's inability to anticipate B2B demand at scale — a demand pattern that no consumer electronics company can reliably model using consumer sales data. Apple brought its Mac hardware refresh forward by six weeks in August 2026, an extraordinary departure from its usual autumn cycle, reportedly in response to this demand pressure. That reactive posture — changing product launch timing because enterprise customers consumed supply faster than forecasted — is not how Apple has historically operated.

The structural question Ternus inherits is whether Apple will attempt to build the enterprise support infrastructure that this revenue stream requires, or whether it will continue routing B2B demand through partners like Mount Thor and WebAI while remaining a consumer company that happens to sell into enterprise environments. The latter path accepts the risk that Nvidia's DGX Spark and RTX Spark, which do come with professional support contracts and enterprise software stacks, will capture the portion of this workload that does not require macOS environments. That is a meaningful portion, even among the AI labs currently running Mac-based training.

Nvidia's recognition that Apple is its biggest competitor in local AI — reportedly the assessment of its own executives, per The Information — suggests the competition is taken seriously in both directions. What Apple has that Nvidia cannot easily replicate is the macOS environment that any company's computer-use agents must eventually learn to navigate, because macOS is what tens of millions of enterprise users actually run on their desks. What Nvidia has that Apple has not built is the enterprise go-to-market infrastructure to support it. The hardware architecture gave Apple an accidental early lead. Whether it builds the enterprise organization to turn that lead into a durable business is now Ternus's question to answer.