For most of my 13 years building software, the first question of any product was the same: who is the user? The answer was always a person. Someone tired, busy, holding a coffee, clicking the wrong button.
On November 12, 2026, that answer got a bit stranger.
What happened
Mount Thor, a Bay Area startup, announced it raised $45 million across its Seed and Series A rounds to build cloud infrastructure on Apple hardware. CRV and Conviction co-led the Series A. Chad Byers and Offline Ventures co-led the Seed. Angels from OpenAI, Meta, Cognition, Together AI and Vercel joined in.
The product, as 9to5Mac describes it, is elastic macOS compute: you rent a bare-metal Mac or an isolated macOS virtual machine with admin access the way you'd rent a Linux server, and pay per Mac in 24-hour blocks. The company says it built its first data center in under six months.
The headline pitch is the interesting part. Mount Thor argues that AI agents are becoming macOS users, so demand for Macs will grow with the number of agents, not the number of people.
Why would an AI need a Mac?
If you're not technical, here's the simple version. The newest AI agents don't just answer in a chat box. They operate a computer: they look at the screen, move the cursor, open apps and fill in forms. Researchers call these "computer-use" agents.
To get good at that, an agent needs practice on a real desktop, thousands of times. Picture a driving school where the student never sleeps and every car is a Mac. That's roughly what "training computer-use models in reinforcement learning environments" means.
The other uses are more down to earth:
- Sandboxes. You want an agent to do work on a Mac, but not on your Mac, where one wrong click deletes a folder. You give it a disposable one.
- Long-running helpers. An agent that keeps working across your devices under one Apple ID while your laptop is closed.
- Build farms. Developers who ship iOS and Mac apps need Macs to compile and test them. That part is old news.
There's a hardware argument too. Apple silicon uses unified memory, and the newer Mac Studio models can hold very large open-weight models on one machine while drawing much less power than a GPU server. Mount Thor says part of the money goes into an inference lab for open-source models on Apple silicon.
This didn't come out of nowhere. Back in August, reports said OpenAI had bought tens of thousands of Mac minis and Mac Studios to train agents. Mount Thor is betting that everyone else will want that without buying a warehouse of Macs themselves.
What I like about it
As a founder, I respect a startup that picks a boring-sounding problem everyone needs solved. Running Macs at scale is a pain. They weren't built to live in racks, and macOS licensing and management have their own rules. If an AI lab wants 500 Macs on Monday and 50 on Friday, someone has to deal with power, cooling, imaging and broken machines. That's real work worth paying for.
It also changes how I think about products. If some of your users are agents, your app's interface becomes an API whether you planned it or not. Agents will click your buttons, read your error messages and get stuck on your cookie banner just like people do, only faster and a thousand times in a row.
Where I'm skeptical
Renting Macs in the cloud isn't new. AWS has offered Mac instances for years, and MacStadium has been at it even longer. The real question is whether agent demand is a lasting habit or a training rush that fades once the big models have learned to use a desktop.
I've built products on top of a wave before. Some waves keep rising and some just leave you holding hardware. Data centers full of Macs are a big bet on the first kind.
And one small, petty observation: in this new world, the bot gets a Mac Studio before most junior developers do. I'm not bitter. Mostly.
A quick check if you're building software
If agents might use your product, try this:
- Do your main flows work without hover-only menus or surprise pop-ups?
- Are your error messages clear enough for a machine to act on?
- Can someone do the core task through an API instead of the UI?
If all three are yes, you're friendly to both humans and agents. If not, people were probably struggling too.
Your turn
Would you let an AI agent use your own computer for an hour, or would you only trust it on a rented machine far away? Why?
