Torus Raises $500 Million, Ships AI Inference to Spare Grid Power

Torus has raised $500 million to put GPU compute wherever the grid has power to spare. Magnetar led the round, announced Wednesday, with ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital joining. Bloomberg, which first reported the deal, described it as a Series C; no valuation was disclosed.

The money backs a new product called Foundry: prefabricated AI inference sites, 1 to 20 megawatts each, installed behind the meter at commercial and industrial facilities with unused grid capacity. Torus says contract-to-compute takes 12 to 16 weeks — a fraction of the timeline for a purpose-built data center.

"Most providers are bringing power to the chips. Foundry brings the chips to power," founder and CEO Nate Walkingshaw said.

What exactly is Foundry?

Each unit is prefabricated at GigaOne, Torus's 545,000-square-foot factory in Salt Lake City, and packages power, storage, cooling, networking and compute into one box. The smallest system is roughly the size of a small cargo van.

Units plug into headroom at existing transformers, and each one is paired with Torus Reserve — the company's flywheel-and-battery storage system. Reserve smooths the spiky load of AI inference so more compute fits on an existing grid connection, and it provides backup power during outages.

The deployment math is fast. Site preparation takes about a week. Installation and commissioning take one to two days. Full process from signed contract to running workloads: 12 to 16 weeks, per Torus.

Buyers have three options:

• A powered shell without GPUs
• A turnkey bare-metal unit with GPUs installed
• Dedicated single-tenant inference capacity

Who's buying, and at what scale?

The first customers are GPU cloud company Hydra Host and a second, unnamed AI cloud provider. Torus says thousands of GPUs are already deployed through Foundry, and the company targets more than 47,000 GPUs by the end of 2027.

Capacity ambitions match. Torus says its Foundry production line can build out 500 megawatts of IT load, with plans to expand to 1 gigawatt.

Where does the $500 million go?

The funding will finance expansion of the GigaOne factory and deployments of both Reserve and Foundry across the US, the company said.

Magnetar is no newcomer here. The firm invested $200 million in Torus in September 2025, when the Salt Lake City company was still selling storage systems to utilities, data centers and industrial sites. At that point Torus also held a memorandum with PacifiCorp for up to 500 megawatts of demand response capacity.

The pivot is telling. Torus built its business on flywheel and battery storage; now it's stacking compute on top of that storage stack. The pitch is that inference workloads — unlike training — can tolerate distributed, edge-adjacent siting, and that grid interconnection queues, not chip supply, are the binding constraint on AI capacity growth.

Why the grid angle matters

Every Foundry unit monetizes capacity that already exists: transformer headroom, not new interconnection. That's the commercial hook. If thousands of megawatts of latent grid capacity can host inference at 12-to-16-week timelines, Torus competes not with hyperscale builders but with the multi-year wait for new power contracts.

With 47,000 GPUs targeted by end-2027 and a production path toward 1 gigawatt of IT load, Torus is betting that distributed inference behind the meter becomes a durable slice of AI infrastructure — not a stopgap.