OpenAI has reportedly purchased tens of thousands of Apple Mac mini and Mac Studio computers to support reinforcement learning and the development of AI agents capable of operating computers, highlighting an unexpected role for Apple’s desktop hardware in frontier AI development. The purchases, reported by The Information and attributed to people familiar with OpenAI’s efforts, are separate from the large GPU infrastructure the AI company uses for conventional model training.
The reported buying spree comes as demand for Apple’s compact desktops rises among AI developers and companies looking to run models locally or distribute workloads across many machines. OpenAI is said to be seeking even more Macs, while Anthropic is reportedly renting Mac minis through Amazon Web Services for similar work. The development also comes shortly after Apple refreshed both product lines with newer chips, as shortages and strong AI demand put additional pressure on the supply of high-memory desktop systems.
OpenAI Reportedly Buys Tens Of Thousands Of Macs
OpenAI has purchased tens of thousands of Mac mini and Mac Studio systems over the past several months, according to The Information.
The report says the machines are being used for reinforcement learning (RL) and for training computer-use AI agents.
Neither OpenAI nor Apple has publicly confirmed the reported purchase volume.
The “tens of thousands” figure should therefore be treated as a reported estimate rather than an officially disclosed procurement number.
OpenAI Mac Purchase At A Glance
| Particular | Reported Details |
|---|---|
| Buyer | OpenAI |
| Hardware | Mac mini and Mac Studio |
| Reported scale | Tens of thousands |
| Main use | Reinforcement learning |
| Additional use | Computer-use AI agents |
| Purchase period | Past several months |
| Official confirmation | Not publicly disclosed |
| Apple response | No public confirmation of OpenAI purchase |
| Anthropic’s approach | Renting Mac minis through AWS |
The scale is notable because OpenAI is one of the world’s largest users of specialized AI computing infrastructure.
Why Is OpenAI Using Mac Minis And Mac Studios?
The reported use case is different from the massive GPU clusters normally associated with frontier AI training.
OpenAI is reportedly using the Macs for reinforcement learning, where AI systems learn through repeated trial and error.
The machines are also being used to train agents that can operate computers, navigate software interfaces and complete multi-step tasks.
How Reinforcement Learning Works
AI agent
│
▼
Attempts task
│
▼
Takes action
│
▼
Receives feedback
│
▼
Learns from result
│
▼
Adjusts behavior
│
▼
Repeats process
For computer-use agents, this can involve thousands or millions of individual interactions with software environments.
That makes having many independent machines useful for running parallel training and testing environments.
What Are Computer-Use AI Agents?
Computer-use agents are AI systems designed to interact with computers in ways that resemble human operation.
Instead of simply responding to a question, an agent can potentially:
- Open applications
- Navigate websites
- Click interface elements
- Enter information
- Write and test code
- Organize files
- Complete multi-step workflows
The technology is becoming an important part of the next generation of AI assistants.
Computer-Use Agent Workflow
User instruction
│
▼
AI interprets task
│
▼
Plans actions
│
▼
Interacts with computer
│
├── Click
├── Type
├── Navigate
├── Read
└── Execute
│
▼
Checks result
│
▼
Continues or corrects
Training these systems requires large numbers of controlled environments in which agents can practice.
Macs Are Being Used As AI Training Workstations
The reported OpenAI deployment highlights a different way of building AI infrastructure.
Instead of relying entirely on centralized data-center GPU clusters, AI labs can use large numbers of relatively compact computers for specialized workloads.
Apple’s silicon is particularly interesting because its CPU, GPU and other processing components share a unified memory architecture.
That can make certain AI workloads practical on a desktop system without a separate high-end discrete GPU.
Why Apple Silicon Is Attractive
| Feature | Potential AI Benefit |
|---|---|
| Unified memory | CPU/GPU share memory pool |
| Compact design | High compute density |
| Low peripheral requirements | Suitable for headless use |
| Local processing | Less cloud dependency |
| Apple silicon | Strong performance per watt |
| Mac Studio cooling | Designed for sustained workloads |
| Large memory configurations | Useful for local AI models |
Apple’s desktop systems are not replacements for the largest Nvidia GPU clusters in every workload.
Instead, they can complement them for specialized tasks.
Why Mac Minis Instead Of MacBooks?
The choice of Mac mini and Mac Studio is significant.
Both are desktop systems designed without built-in displays, keyboards or batteries.
For a large AI deployment, those missing components can actually be an advantage.
A company running thousands of machines does not need a display or keyboard attached to every system.
Desktop Vs Laptop For AI Infrastructure
| Feature | Mac Mini / Studio | MacBook |
|---|---|---|
| Display | None | Built-in |
| Keyboard | None | Built-in |
| Battery | None | Included |
| Cooling | Desktop-oriented | Portable design |
| Rack / cluster deployment | Easier | Less practical |
| Long-running workloads | Suitable | More constrained |
| Bulk compute use | Efficient form factor | Includes unnecessary hardware |
The desktop form factor allows organizations to focus their spending on processing power and memory rather than components that are irrelevant to automated workloads.
Apple Silicon’s Unified Memory Is A Major Factor
One of the biggest technical advantages of Apple’s silicon for local AI workloads is its unified memory architecture.
Instead of having separate pools of memory for the CPU and GPU, Apple’s architecture allows them to access a shared memory pool.
This can be useful when running AI models that need to keep substantial amounts of data in memory.
Traditional PC
│
┌────┴────┐
CPU GPU
│ │
RAM VRAM
└────┬─────┘
│
Separate memory pools
Apple Silicon
│
┌────┴────┐
CPU GPU
└────┬────┘
│
Unified Memory
The architecture does not make Apple silicon universally faster than Nvidia GPUs.
But for certain local inference, agent and reinforcement-learning workloads, the combination of memory capacity, power efficiency and compact hardware can be attractive.
OpenAI Still Relies On Large GPU Infrastructure
The Mac purchases should not be interpreted as OpenAI abandoning Nvidia GPUs or cloud-based AI infrastructure.
The reported Mac deployment is intended for a specific class of workloads.
Frontier-model pretraining requires enormous amounts of parallel compute and high-speed networking, areas where large GPU clusters remain dominant.
AI Infrastructure By Workload
| Workload | Typical Hardware Approach |
|---|---|
| Frontier model pretraining | Large GPU clusters |
| Large-scale inference | Data-center accelerators |
| Reinforcement learning | Potentially distributed Macs + GPUs |
| Computer-use agents | Local/desktop environments |
| Model experimentation | Local workstations |
| AI development | Macs, PCs and cloud systems |
OpenAI’s reported Mac purchases therefore appear to supplement rather than replace its broader compute infrastructure.
Anthropic Is Also Using Mac Minis
OpenAI is not the only AI company turning to Apple’s desktop hardware.
According to The Information, Anthropic is renting Mac minis through Amazon Web Services for similar workloads.
The difference is that OpenAI is reportedly purchasing the hardware directly, while Anthropic is accessing Mac systems through a cloud provider.
OpenAI Vs Anthropic
| Company | Mac Strategy | Reported Purpose |
|---|---|---|
| OpenAI | Buying tens of thousands | RL + computer-use agents |
| Anthropic | Renting through AWS | Similar AI workloads |
| Apple | Supplying hardware | AI-capable desktops |
| AWS | Providing rented Macs | Cloud Mac infrastructure |
The use of Macs by multiple frontier AI companies suggests the hardware is finding a niche beyond traditional software development.
Apple’s Mac Business Is Benefiting From AI Demand
The reported AI-lab demand comes at a favorable time for Apple’s Mac business.
Apple’s latest quarterly Mac revenue rose nearly 29% year over year to about $10.4 billion, according to The Information, making the Mac its fastest-growing product category in the cited quarter.
The broader increase in AI workloads is creating new demand from developers, businesses and AI companies.
Apple’s Mac Revenue Growth
| Metric | Latest Reported Quarter |
|---|---|
| Mac revenue | ~$10.4 billion |
| Year-over-year growth | ~29% |
| Relative performance | Fastest-growing Apple product segment |
The AI boom is therefore creating an unusual secondary market for computers that were traditionally marketed as consumer and professional workstations.
Apple Recently Refreshed The Mac Mini And Mac Studio
Apple launched updated Mac mini and Mac Studio models on August 25, 2026, emphasizing AI capabilities and professional workloads.
The new Mac mini starts at $899 and is available with Apple’s M6 chip or M5 Pro.
The refreshed Mac Studio uses M5 Max and M5 Ultra chips, with pricing starting at $2,499 for the M5 Max version and $5,499 for the M5 Ultra configuration.
Latest Apple AI-Capable Desktops
| Product | New Chip | Starting Price |
|---|---|---|
| Mac mini | M6 / M5 Pro | $899 |
| Mac Studio | M5 Max | $2,499 |
| Mac Studio | M5 Ultra | $5,499 |
Apple said the M6 Mac mini can deliver up to four times faster AI performance and improved graphics and storage speeds compared with the previous M4 model.
AI Demand May Have Caught Apple Off Guard
The reported surge in enterprise and AI demand appears to have exceeded Apple’s expectations.
According to reports based on The Information, the company lacked some dedicated resources for enterprise AI customers, including a dedicated engineering team focused on business customers and a specialized enterprise AI strategy.
At the same time, higher-memory Mac configurations have faced supply constraints.
Factors Driving Mac Demand
AI model development
+
Local AI inference
+
Computer-use agents
+
Reinforcement learning
+
Developer demand
│
▼
Higher demand for
Mac mini + Mac Studio
│
▼
Inventory pressure
This represents a significant shift in how Apple’s desktop computers are being used.
Mac Shortages Are Becoming More Visible
Reports indicate that some high-memory Mac mini and Mac Studio configurations have experienced extended delivery times.
The supply pressure is occurring against a broader backdrop of tight memory markets driven by the rapid expansion of AI data centers.
Higher-capacity memory is increasingly valuable not only for traditional computing but also for running larger AI models locally.
Why High-Memory Macs Matter For AI
| Component | Importance |
|---|---|
| Unified memory capacity | Determines model size that can fit locally |
| Memory bandwidth | Affects AI workload performance |
| Chip efficiency | Influences operating cost |
| Cooling | Important for sustained workloads |
| Storage | Useful for models and datasets |
For AI companies, the ability to purchase thousands of machines with substantial memory capacity can be more important than conventional desktop specifications.
Nvidia Sees Apple As A Local AI Competitor
The growing popularity of Apple silicon for local AI workloads is also attracting attention from Nvidia.
Reports have described Apple as a major competitor in the local AI-computing market.
Nvidia launched its DGX Spark desktop AI system as part of its effort to bring powerful AI computing closer to developers and businesses.
The competition highlights a broader shift from centralized AI data centers toward smaller systems capable of running models locally.
Local AI Hardware Competition
| Company | Hardware Direction |
|---|---|
| Apple | Mac mini / Mac Studio |
| Nvidia | DGX Spark |
| Traditional PC vendors | AI PCs |
| Cloud providers | Hosted GPU + Mac infrastructure |
| AI labs | Mixed hardware environments |
The competition is no longer limited to giant data-center clusters.
Why Local AI Hardware Is Growing
AI developers increasingly want to run models locally for several reasons.
Local processing can reduce latency, provide greater control over data and lower cloud-computing costs for some workloads.
For repetitive agent tasks, running thousands of local environments can also provide a practical way to create controlled testing infrastructure.
Cloud-only AI
│
├── High scalability
├── Centralized resources
└── Ongoing compute costs
Local AI
│
├── Lower latency
├── Data control
├── Predictable hardware
└── Potentially lower marginal cost
The ideal approach depends on the workload.
Reinforcement Learning Is Different From Model Pretraining
The reported Mac deployment also highlights an important distinction between AI training methods.
In conventional supervised or pretraining workloads, enormous datasets are processed across highly interconnected accelerator clusters.
Reinforcement learning involves agents repeatedly interacting with environments and receiving feedback.
That can create many independent workloads that can be distributed across separate machines.
Pretraining Vs Reinforcement Learning
| Feature | Model Pretraining | Reinforcement Learning |
|---|---|---|
| Main objective | Learn general patterns | Learn from outcomes |
| Data source | Large datasets | Environment interactions |
| Work pattern | Highly parallel compute | Many repeated trials |
| Hardware | Large GPU clusters | Distributed systems can help |
| Feedback | Dataset-based | Reward / evaluation |
| Relevance to computer agents | Foundation model | Agent behavior training |
This helps explain why thousands of relatively independent Macs could have a role in OpenAI’s reported setup.
OpenAI May Be Building A Large Agent Training Environment
If the reported purchases are accurate, tens of thousands of Macs would provide OpenAI with a substantial number of independent computing environments.
That could allow the company to run many computer-use agents simultaneously.
Each machine could potentially host an operating-system environment where an agent learns to navigate software, execute tasks and recover from errors.
Mac #1 → Agent training
Mac #2 → Agent training
Mac #3 → Agent training
Mac #4 → Agent training
...
Mac #10,000+ → Agent training
...
Mac #20,000+ → Agent training
The exact deployment architecture has not been publicly disclosed.
It is therefore not known whether the machines are organized into a single cluster, distributed across facilities or deployed in another configuration.
The Strategy Could Help AI Agents Become More Reliable
Computer-use agents face a unique challenge.
An AI model can generate technically correct text while still failing when confronted with a graphical interface.
Buttons can move, websites can change, applications can display unexpected warnings and tasks can require multiple steps.
Repeated interaction with real software environments can help AI systems learn how to handle such situations.
Computer-Use Training Challenges
| Challenge | Example |
|---|---|
| Interface changes | Button moves |
| Unexpected dialogs | Error message |
| Multi-step tasks | Long workflow |
| Visual interpretation | Understanding screen |
| Recovery | Correcting mistakes |
| Tool interaction | Keyboard and mouse |
| Reliability | Completing task consistently |
This makes computer-use training a potentially important application for distributed hardware.
OpenAI’s Mac Purchases Could Influence Apple’s Product Strategy
If enterprise AI demand continues, Apple could increasingly view the Mac as infrastructure rather than simply a personal computer.
The company’s latest product messaging already emphasizes that the Mac mini can function as an always-on agentic device, while the Mac Studio is positioned for developers and professionals working with large AI models.
This could eventually lead Apple to offer more enterprise-oriented deployment tools, management capabilities or hardware configurations specifically designed for AI workloads.
There Are Still Limitations
Apple silicon has clear advantages for certain workloads, but it is not a universal replacement for Nvidia accelerators.
Large-scale model training depends heavily on high-bandwidth interconnects, specialized software stacks and massive parallel processing.
Mac systems also have different networking and deployment characteristics from purpose-built data-center hardware.
Mac Vs Data-Center GPU
| Factor | Mac Mini / Studio | Data-Center GPU |
|---|---|---|
| Form factor | Compact desktop | Server / rack |
| Unified memory | Major advantage | Typically separate VRAM |
| AI ecosystem | Growing | Mature at scale |
| Large-scale training | Limited | Strong |
| Local inference | Strong use case | Strong |
| Agent environments | Potentially useful | Strong |
| Power efficiency | Attractive | Designed for high throughput |
| Networking | Desktop-oriented | High-speed cluster networking |
The reported OpenAI purchase should therefore be understood as workload-specific optimization rather than a wholesale shift away from GPUs.
The Bigger Picture
OpenAI’s reported purchase of tens of thousands of Mac mini and Mac Studio computers illustrates how the AI hardware market is expanding beyond conventional Nvidia GPU clusters. The machines are reportedly being used for reinforcement learning and for training computer-use agents, where AI systems repeatedly interact with operating systems and software environments to learn how to complete multi-step tasks.
The development is also significant for Apple. Mac revenue rose nearly 29% year over year to about $10.4 billion in the latest reported quarter, while Apple has just refreshed the Mac mini and Mac Studio with newer chips aimed partly at AI workloads. The new Mac mini starts at $899, while Mac Studio pricing starts at $2,499, giving AI developers a range of desktop systems capable of running substantial local workloads.
Looking Ahead
The most important question is whether OpenAI’s reported Mac deployment becomes a lasting component of its AI infrastructure or remains a specialized solution for reinforcement learning and computer-use agents. If the approach proves effective, other AI labs could follow by purchasing or renting large numbers of Apple desktops to create distributed environments for agent training, testing and evaluation. Anthropic’s reported use of rented Mac minis through AWS suggests that interest in this type of hardware is already extending beyond OpenAI.
For Apple, the development could open an unexpected enterprise market for Mac hardware. The company’s unified-memory architecture and compact desktop designs are increasingly attractive for local AI workloads, while Nvidia and other chipmakers are also moving into the desktop AI-computing market. OpenAI’s reported purchases therefore represent more than a large hardware order: they could be an early signal that the next phase of AI infrastructure will combine massive data-center clusters with thousands of smaller machines dedicated to training and operating increasingly autonomous AI agents.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
