Torus Raises $500 Million to Put AI Inference Next to Spare Grid Capacity With Foundry
Torus, a Salt Lake City maker of distributed power systems, said on November 11 2026 that it has raised $500 million and launched Foundry, a prefabricated unit that packs power, storage, cooling, networking and GPUs into one box. Magnetar led the round, with ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital.
The pitch fits in one line from CEO and founder Nate Walkingshaw: “Foundry brings the chips to power.” Most AI builders hunt for a site with enough grid connection, then wait years for it. Torus wants to place inference compute behind the meter at individual transformers that already have headroom, so no new transmission or generation has to be built.
How fast this is moving
Fourteen months ago Torus was a flywheel-and-battery storage company. In September 2025 it raised $200 million from Magnetar to deploy modular power plants for utilities, data centers and industry. The new round is two and a half times that size and comes with a new product line aimed squarely at GPU capacity.
The deployment claims are just as quick. Foundry sites run from 1 MW to 20 MW. Torus says it takes 12 to 16 weeks from contract to running AI workloads, with about a week of site preparation and one to two days for installation. The smallest unit is about the size of a small cargo van. Each one carries Torus Reserve, its inertial flywheel and battery storage, to cover demand peaks and outages. Cooling is closed-loop and waterless, which Torus says uses 95% less water than evaporative coolers.
The factory has kept pace. GigaOne, Torus’s 545,000 square foot plant in Salt Lake City, opened its first production line in early 2026 and has added a new line every 21 days since. The Foundry line can build 500 MW of IT load a year, with plans for 1 GW. The company has nearly 400 employees.
What holds up and what doesn’t yet
Some of it is checkable. Torus has paying customers: Hydra Host and a second, unnamed AI cloud provider are the launch customers. It says it has thousands of GPUs deployed today. The earlier storage business has real deployments behind it, and its products have collected a TIME Best Invention listing and a 2025 R&D 100 Award.
Other parts are plans. The target of more than 47,000 GPUs in Foundry units by the end of 2027 is a goal. The 500 MW figure is factory capacity, not installed load. The release names no valuation and no round type. The Magnetar partner quoted is its head of alternative credit and fixed income, and Trinity Capital is a venture lender. That suggests at least part of the $500 million is structured capital or debt rather than straight equity. Torus hasn’t said.
The three ways it sells Foundry tell you who it expects to buy. Strategic customers get a powered shell with no GPUs. Partners get turnkey bare metal with GPUs installed. Direct customers get single-tenant, sovereign inference capacity for their own use. That last option is aimed at enterprises and governments that want inference on their own hardware and inside their own borders.
What it changes
The bet rides on a split in AI demand. Torus cites McKinsey projections that data center demand for inference will grow 35% a year through 2030, against 22% for training. Training wants giant campuses. Inference can live in many smaller places close to users, if power is there.
If Foundry works at the claimed speed, the scarce asset in AI infrastructure stops being the gigawatt campus and becomes the overlooked transformer with spare capacity. Utilities, factories and airports with headroom would turn into compute sites. The weak point is the same one every distributed model faces: hundreds of small sites are harder to run, secure and keep busy than one big one. Watch whether the unnamed second customer is named, and how many megawatts are actually running by mid-2027.
For the wider media read on AI buildout, see mediapresser.com.