$45 million across two rounds, announced on 12 November, for a company that rents Macs.
Mount Thor is a San Francisco startup founded this year. It sells elastic macOS instances on Apple silicon at data-centre scale, provisioned through a portal, a command-line tool, an API or Kubernetes, and billed per Mac on a 24-hour basis. CRV and Conviction co-led the Series A. Chad Byers and Offline Ventures co-led the Seed. The company also raised non-dilutive capital for hardware and did not disclose how the equity splits between the rounds.
The round is the easy part. The harder claim is that the next bottleneck in AI infrastructure is an operating system, not another accelerator. That claim is testable, and it prices very differently from the GPU story.
The moat is lifecycle management — provision, isolate, wipe, bill — not access to silicon.
Inference on Apple silicon is the funding thesis; the published evidence is still a roadmap, not a benchmark table.
What the product actually provisions
Mount Thor documents two compute types. Bare metal gives a customer one physical Mac and a permissive non-root account. A virtual machine gives an isolated macOS environment with administrator access, two slots per host, and the VM service is still in preview. Both connect over SSH or a remote desktop through private authenticated tunnels, with guest addresses kept off the public internet.
The control layer matters as much as the hardware. VMs and Macs appear as custom resources inside a customer's Kubernetes namespace, so an infrastructure team can declare a machine and manage its lifecycle the way it manages any other workload. Fleets come in two flavours: Mount Thor-managed, where the company prepares, monitors and wipes machines, and customer-managed, where an organisation runs its own Macs at a Mount Thor site under its own security model.
None of that is glamorous. It is also the part that decides whether the business works. Anyone with a purchase order can buy Mac Studios. Few teams want to track firmware, images, access reviews and machine wipes across hundreds of them.
Equity raised, November 2026
Announced 12 November, plus undisclosed non-dilutive hardware financing. Mount Thor, 2026
Agents need the desktop, not the endpoint
An agent that clicks through software needs somewhere to open applications, inspect their state, move files and resume work when the next instruction arrives. A model endpoint does not provide that. macOS does, and a large share of business software only runs there.
That is the demand Mount Thor is selling into. Research labs building reinforcement-learning environments for computer-use models, providers offering macOS sandboxes, and agent products that keep a machine alive for weeks all need the same primitive: a Mac that can be allocated, isolated and thrown away.
As we wrote in November, the agent stack keeps splitting into new infrastructure categories faster than buyers can name them. Memory became a product; now the desktop is becoming one.
Macs in the cloud are not new. AWS has rented Mac instances for years, and MacStadium did it earlier still. Mount Thor's pitch is narrower and more interesting: the customer is an agent, so demand grows with the number of agents rather than the number of developers. That reframing is the whole investment case, and it is also the first thing a sceptic should test.
Bare metal or virtual machine?
Virtual machine — isolated macOS with administrator access, two slots per host, still in preview. The choice when an agent has to change system settings.
Releasing a bare-metal Mac wipes it. Deleting a VM removes the VM and its disk.
The inference half is still a promise
The second half of the pitch is cheaper inference. The M5 Ultra Mac Studio shipped in August with up to 512GB of unified memory, which puts large open-weight models within reach of a single machine at a fraction of the power draw of a GPU server. Mount Thor says part of the new money goes to building what it calls the best inference lab for open-source models on Apple silicon.
Look for the numbers that are missing. There is no comparison of tokens per second against a GPU node, no price per million tokens, no utilisation figure and no customer benchmark. Without those, the Apple inference case remains an engineering hypothesis with a plausible memory argument attached.
That gap is not fatal. It is simply unpriced. A company can raise $45 million on an environment story and still have to prove a serving story later, and the two are judged by different buyers.
The numbers behind the pitch
Dealroom places the $45 million in the 96th percentile of US Series A rounds in enterprise software, against 5,259 comparable deals. The demand signal around it is real: OpenAI reportedly bought Mac minis and Mac Studios for agent-training work, Anthropic rented Mac capacity through AWS, and Apple's Mac revenue reached $10.4 billion in the June quarter, up almost 29% year over year.
Those data points support the category. They do not support any particular vendor. Enterprise buyers comparing a Mac cloud will ask for a per-agent cost, a wipe guarantee and an audit trail. The interesting question is whether Mount Thor sells those as a service contract or as infrastructure that a platform team operates itself.
What would turn a fleet into a business
The disclosed facts are thin in one specific place. The company did not break out the Seed and Series A amounts, did not give a valuation, and did not name customers. Mount Thor deployed its first data centre in under six months and runs capacity with Tier 1 colocation partners in major US metros, with Europe and Asia listed as coming soon.
Two operational numbers will decide the outcome. The first is utilisation per host, because a Mac rented by the day is a terrible asset if it idles. The second is the share of the fleet sold as VMs, because two slots per host change the economics of every machine that can be shared. Whole-Mac tenancy is easier to isolate and harder to fill.
The team's background helps here. People who built infrastructure at Cerebras, CoreWeave, Fluidstack, Nebius, OpenAI and xAI know what a rack-level deployment costs and how fast a customer can consume it. That experience is worth more at this stage than another benchmark slide.
The colocation bet behind the price
Mount Thor runs its fleets with Tier 1 colocation partners in major US metros and lists Europe and Asia as future markets. That is a deliberate trade. Metro colocation keeps latency low and puts capacity near the customers who care about a persistent desktop, and it also buys the most expensive square footage in the data-centre business.
Power is the friendlier part of the equation. A rack of Mac Studios draws a fraction of what a dense GPU rack pulls, so the usual interconnection queue fight does not apply. The constraints move to floor space, network plumbing, physical security and the labour of preparing and wiping machines. Those costs scale with the number of boxes, not with the number of tokens served.
Billing exposes the same tension. A Mac rented on a 24-hour basis earns nothing during the hours an agent sleeps, and agent workloads are bursty by nature. The operator that packs short jobs into the same machine, or resells the idle window, keeps a margin that a manual fleet cannot match. Without that scheduling layer, a Mac cloud is a leasing business wearing cloud clothes.
Refresh risk sits on top. Apple shipped the M5 Ultra Mac Studio in August, and every generation makes the previous fleet cheaper to rent and faster to write down. A provider that owns thousands of Macs is making a hardware bet with a software company's valuation. The winner in this category will probably be the operator with the best scheduler and the shortest depreciation curve, not the one with the largest fleet.
Utilisation per host, disclosed in any future capacity update.
A published inference benchmark with a price per million tokens.
Named enterprise or lab customers, and whether they renew.
Europe and Asia capacity dates, against the promised expansion.
The $45 million does not prove that agents need Macs. It buys Mount Thor the time to show that a rack of them can be rented at a profit, and that the operating system is the product rather than a side effect of buying Apple hardware.