Tens of thousands of Mac minis instead of GPU clusters: The surprising training infrastructure chosen by OpenAI
According to a report by The Information on August 29th, OpenAI has been purchasing tens of thousands of Apple Mac minis and Mac Studios over the past few months. This is for reinforcement learning and training AI agents that interact with computers. What's interesting from an engineer's perspective is the technical reason why consumer-grade desktop computers were chosen for this purpose, rather than NVIDIA GPU clusters.
The unexpected suitability of "unified memory"
Behind this choice lies the characteristics of the "unified memory" architecture used in Apple's chips. Normally, CPUs and GPUs have separate memory spaces, but Apple's silicon is designed so that the CPU and GPU share a single memory pool.
Why is this design suitable for training computer-operating agents? Because the workload of this type of reinforcement learning is significantly different in nature from the pre-training of typical transformer models. Computer-operated agents actually run within the operating system, observing what is displayed on the screen, taking actions, and learning by receiving feedback from the results millions of times. This process requires relatively light parallelism while using a large amount of memory, rather than massive parallel computing. This characteristic perfectly aligns with the strengths of Apple's integrated memory architecture.
Anthropic also makes a similar move via AWS
This trend is not limited to OpenAI. According to the same report, Anthropic, supported by Google and Amazon, is also taking a similar approach, renting the computing power of Mac minis via Amazon Web Services (AWS) and using it for its reinforcement learning tasks. The fact that Frontier AI Labs have reached the same technical conclusion, even if not by prior arrangement, suggests that the suitability of Apple silicon for this application is not mere coincidence.
Demand is so large that it is straining Apple's supply chain
The scale of this demand is even impacting Apple's own product supply plans. For high-performance Mac mini and Mac Studio configurations with large amounts of memory, delivery times are reportedly extending from several weeks to several months. Apple brought forward its product refresh, normally scheduled for the fall, to August 25th, introducing the new M6 chip (Apple's first chip manufactured using a 2-nanometer process) in the Mac mini, and the M5 Pro, M5 Max, and M5 Ultra in the Mac Studio.
CEO Tim Cook has reportedly warned that this shortage will continue for several months after this product refresh. This indicates that large enterprise orders are beginning to directly impact regular consumer demand.
Positioning as a "differential use" rather than a "GPU replacement"
An important caveat here is that this development does not mean OpenAI is breaking away from its reliance on NVIDIA GPU clusters. Large, tightly coupled GPU clusters are still essential for pre-training frontier models, and NVIDIA's dominance in this regard remains unchallenged.
The use of the Mac mini in this instance is positioned as a complementary option for specific types of workloads—specifically, the computational demands of reinforcement learning and agent training, which have different natures. This suggests that a multi-architecture configuration, where data centers use NVIDIA accelerators for training, their own chip designs for inference, and Apple silicon-based systems for specific agent-related tasks, may become more common in the future.
The "Unintended By-product" Aspect for Apple
Interestingly, this application appears to be something Apple did not initially anticipate. Apple's goal in developing its own chips and transitioning from Intel CPUs was primarily to improve the efficiency and performance of personal computers. However, the resulting integrated memory design has unexpectedly proven to be a strength in areas where the weight of AI models with enormous parameter counts needs to be handled efficiently.
Apple's Mac business sales in the most recent quarter recorded a 29% increase year-over-year, and it has been suggested that large purchases from AI labs may be supporting part of this growth. Apple is arguably becoming an unintentional winner in the AI infrastructure investment competition, as it can benefit from this type of demand without requiring massive investments in data centers.
What Engineers Should Note
This news indicates a clear trend towards AI model training and inference, moving away from a single hardware architecture and towards selecting the optimal hardware based on the nature of the workload. We may see further optimization for specific applications, such as GPU clusters for large-scale transformer pre-training, dedicated accelerators for efficiency-focused inference, and Apple silicon for memory-intensive agent training.
When designing your own AI development infrastructure, it will become increasingly important to move beyond a simplistic approach of "simply adding more GPUs" and instead consider what each workload actually demands, flexibly examining hardware options accordingly.