Torus Raises $500M to Put AI Chips Where the Power Already Is

Torus has raised $500 million in growth capital and launched Foundry, a prefabricated modular system for AI inference that the company says can go from contract to running workloads in 12 to 16 weeks. Magnetar led the financing, announced November 11, 2026. ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital also participated.

The money is earmarked for two things: expanding Torus's manufacturing operation in Salt Lake City and deploying its power and compute systems across the United States. A 12-16 week timeline compares with deployment schedules that can stretch to years for a conventional hyperscale campus — though the company supplies its own figures, and no independent verification accompanied the announcement.

What is Foundry, exactly?

Foundry integrates power, storage, cooling, networking and compute in one prefabricated unit. The smallest system is roughly the size of a small cargo van, requires no building construction, and needs about one week of site preparation plus one to two days of installation and commissioning. Each site can supply between 1 and 20 megawatts of inference capacity.

The design starts with unused electrical capacity between substations and existing commercial or industrial buildings. Every system sits behind the meter — on the customer's side of the utility connection — and draws on available transformer headroom rather than requiring new grid infrastructure. In plain terms, Torus is not asking utilities to build anything; usable capacity on existing connections is the binding constraint on every deal.

Founder and CEO Nate Walkingshaw framed the reversal plainly: "Most providers are bringing power to the chips. Foundry brings the chips to power." The line is memorable. Whether enough sites with genuine headroom exist to hit Torus's targets is the open question the company did not address.

How does the power management work?

Every Foundry ships paired with Torus Reserve, the company's energy-storage and management product. Battery and inertial flywheel storage, governed by predefined dispatch controls, shape demand, cover peaks, support ramping and provide backup power during grid outages. Torus claims this lets more compute run on the same interconnection.

The company also handles ongoing operations and maintenance, with Torus Protect adding integrated networking and security. Torus Reserve is the technical premise underpinning the whole pitch — that storage can stretch a constrained grid connection further — and its performance in the field remains unproven at scale.

Who is buying it?

Foundry launches with two customers: Hydra Host and a second AI cloud provider that Torus declined to name. Three deployment models separate how much of the stack a customer takes:

• Strategic — a powered enclosure without GPUs.
• Partner — a turnkey bare-metal unit with GPUs.
• Direct — sovereign inference capacity on single-tenant hardware allocated exclusively to one enterprise's workloads.

Torus reports thousands of GPUs already deployed in Foundry units and targets more than 47,000 by the end of 2027. That figure — roughly a tenfold-or-more increase from the installed base described at launch — gives the manufacturing expansion a concrete yardstick and the financing a measurable obligation. Takeaway: the round buys a factory ramp, and the factory ramp is now on the clock.

Can the Utah factory deliver?

Foundry is designed, engineered and manufactured at GigaOne, Torus's 545,000-square-foot Salt Lake City facility, where nearly 400 employees work and key components are built in-house. The production line is already online. Torus says the line will be capable of producing 500 megawatts of IT load — the power consumed by computing equipment — with plans to expand to 1 gigawatt, double the initial capacity.

Resource use is part of the site proposition. Torus says Foundry's closed-loop cooling system is waterless and uses 95 percent less water than evaporative coolers. Each system also donates GPU time to medical research through Folding@home.

The financing ties manufacturing and deployment together. The expansion now depends on finding enough sites with usable transformer headroom and converting GigaOne's output into operating systems at a pace that supports the 47,000-GPU target for 2027. Execution at the factory and access to existing electrical connections — not demand for inference — are the constraints that will decide whether the $500 million pays off.