Torus Foundry raises $500M for modular AI inference sites

Torus, the Salt Lake City energy company, said on Wednesday that it raised $500 million in growth capital led by Magnetar and launched Foundry, prefabricated modular units that put 1 to 20 megawatts of AI inference capacity at existing commercial and industrial sites in 12 to 16 weeks, using spare grid power already on site.

The company did not give a valuation. Bloomberg, which reported the raise first, also gave none, and described the round as a Series C. Accounts of the company statement list ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital as additional participants. Bloomberg's visible report named Magnetar, Iconiq Capital, Origin Ventures and Trinity Capital among the investors.

What Foundry is

Each Foundry site, Torus says, packages power, storage, cooling, networking and compute into one prefabricated system, built at the company's GigaOne factory in Salt Lake City. Sites range from 1 to 20 megawatts. The smallest unit is about the size of a small cargo van. The company says a site can go from contract to running AI workloads in 12 to 16 weeks, with about a week of site preparation and one to two days for installation and commissioning.

The units sit behind the meter and are meant to use spare capacity at existing connections, so they do not need a new grid hookup, according to the company. Each site is paired with Torus Reserve, the firm's flywheel and battery storage product, which Torus says covers demand peaks and supplies backup power during outages. Founder and chief executive Nate Walkingshaw said in the statement, "Most providers are bringing power to the chips. Foundry brings the chips to power."

The company cites McKinsey figures that data-centre demand for inference, the work of running already trained models, will grow 35% a year through 2030, compared with 22% for training. Those are figures Torus attributes to McKinsey, not an independent Codema check.

Customers and the 47,000-GPU plan

Torus said thousands of GPUs already run in Foundry units, and that it plans to have more than 47,000 GPUs deployed in Foundry units by the end of 2027. That is a company plan, not a completed rollout. First customers are Hydra Host, an AI infrastructure firm, and a second, unnamed AI cloud provider. Companies can buy a powered shell without GPUs, a turnkey unit with GPUs, or capacity reserved for a single tenant.

Codema's hardware coverage and AI chips and memory guide follow the compute side of this market, and the deals and IPOs hub tracks private raises. What to watch is whether spare transformer headroom can actually support 1 to 20 megawatts of high-load-factor inference at many existing sites, and whether the 47,000-GPU plan lands by the end of 2027. Because no valuation was given, the $500 million figure cannot be sized against other infrastructure raises this week.