$ today_in_10_seconds
GPUs aren't the only infrastructure bottleneck for agents. The operating system is one too. Mount Thor raised $45M to rent Macs at data-center scale to AI agents, not human developers. Their bet is that computer-use agents are becoming macOS users, and that demand will grow with the number of agents rather than the number of people.
$ lead_story
Mount Thor Raises $45M to Build Mac Data Centers for AI Agents
Mount Thor announced $45M across Seed and Series A rounds, with CRV and Conviction co-leading the Series A and Chad Byers and Offline Ventures co-leading the Seed. Angels came from OpenAI, Meta, Cognition, Together AI, and Vercel, and the company also raised non-dilutive financing for hardware. The product runs macOS on Apple silicon as elastic compute instances, the way a cloud provider rents Linux VMs, and the customers are agent builders rather than iOS shops. AWS and MacStadium already rent Macs, so the difference here is the buyer: an AI agent that needs a Mac to do its job. The team comes from Cerebras, CoreWeave, Fluidstack, Nebius, OpenAI, and xAI, and deployed its first data center in under six months.
// why_it_matters: Computer-use agents need real desktops, and someone has to rack them.
$ also_worth_attention
The use case: macOS as an RL gym and agent sandbox
The workloads listed are agent-native: training computer-use models in RL environments, secure macOS sandboxes for agents, long-running instances that let an agent work across a user's devices under one Apple ID, and build/test farms. Provisioning works through a web portal, CLI, Kubernetes, or API, which signals they're competing for platform teams, not IT departments.
link →The second bet: open-weight inference on Apple silicon
The M5 Ultra Mac Studio shipped in August with up to 512GB of unified memory, putting large open-weight models within reach of a single machine at a fraction of a GPU server's power draw. Mount Thor plans a lab for testing open-source models adapted to Apple processors. If that works, agents could run their reasoning loops and their GUI actions on the same box.
link →The catch: Apple's license shapes the product
Customers get bare-metal Macs or two-slot macOS VMs, billed per Mac per 24 hours. Those limits follow Apple's macOS licensing terms, not anything Mount Thor chose. That's fine for long-running agents and expensive for bursty, seconds-long tool calls. Skeptics also note that inference is the funding thesis, but the benchmarks are still missing.
link →
$ bigger_picture
The trend behind the headlines
Most agent infrastructure spending has gone to tokens: GPUs, inference, context. Computer-use agents add a second requirement, which is a real environment to act in. That means GUIs, accessibility trees, and native apps that can't be faked with an API wrapper. Linux covers most of that today. macOS is the gap, because Apple's licensing and hardware make it hard to scale. If agent count, not headcount, drives OS demand, then 'environment compute' becomes its own infrastructure category, with its own economics and bottlenecks. Expect similar plays for Windows desktops and mobile emulators next.
Tart
Tart (Cirrus Labs) is a CLI for creating and running macOS VMs on Apple silicon using Apple's native Virtualization.framework. If you're prototyping a computer-use agent and want a disposable macOS sandbox on your own Mac before paying for a fleet, start here. It also has OCI-registry support for sharing VM images.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime