Torus Raises $500 Million to Build Modular AI Inference Sites on Spare Grid Capacity

Salt Lake City–based Torus announced Wednesday it has raised $500 million in Series C funding and launched Foundry, a line of prefabricated units that package power, storage, cooling, networking and GPUs into 1-to-20-megawatt modules designed to be installed behind the meter at commercial and industrial sites. The round was led by Magnetar, with participation from Iconiq Capital, Origin Ventures and Trinity Capital, according to Bloomberg [1]; Torus's own statement adds Ridgeline Partners to the participant list [3]. No valuation was disclosed [1][4].

What Foundry is

Torus has spent years selling modular battery systems (branded Reserve) to utilities, data centers and industrial customers. Foundry adds AI compute to that platform. Each unit is built at Torus's 545,000-square-foot GigaOne factory in Salt Lake City and shipped to a site where the local transformer has unused headroom. The company says site preparation takes about a week, installation and commissioning one to two days, and the full contract-to-compute cycle 12–16 weeks [3]. Bloomberg cited a "within four months" timeline from CEO Nate Walkingshaw [1].

The smallest unit is roughly the size of a small cargo van. Each Foundry is paired with Reserve, Torus's flywheel-and-battery storage system, which smooths the spiky load of inference workloads so more compute can fit on an existing grid connection and provides backup power during outages [3][4]. Cooling is closed-loop and waterless, using 95 percent less water than evaporative coolers [3].

Torus offers three models: a powered shell without GPUs (Strategic), a turnkey bare-metal unit with GPUs (Partner), and dedicated single-tenant inference capacity (Direct) [3].

Funding and manufacturing scale

The $500 million will fund expansion of GigaOne and deployment of Reserve and Foundry across the United States [3]. Dealroom noted the round sits in the 98th percentile of all-time U.S. Series C energy deals, signaling strong investor conviction [2]. Magnetar previously invested $200 million in Torus in September 2025 [4].

GigaOne's Foundry production line is online with a capacity of 500 MW of IT load; the company says it plans to expand to 1 GW [3][4]. The factory opened its first line in early 2026 and has opened a new production line every 21 days in 2026 [3]. Torus employs nearly 400 people at the facility [3].

Claims and evidence

Torus says "thousands of GPUs" are already deployed in Foundry units and that it will have more than 47,000 GPUs deployed by the end of 2027 [3][4]. Only one customer is named: Hydra Host, a GPU cloud provider. A second AI cloud provider is described as unnamed [3][4]. Dealroom flagged that the company has disclosed only one named customer so far and that modular hardware businesses carry manufacturing and deployment risk at volume [2].

The 12–16-week deployment claim is a company projection; no independent verification was provided. The 47,000-GPU target is a forward-looking statement contingent on manufacturing ramp and customer uptake.

Why now: power constraints and political headwinds

AI inference demand is projected to grow 35 percent annually through 2030, compared with 22 percent for training, according to McKinsey [3]. Developers seeking to build large campuses face longer timelines and greater scrutiny, especially after last week's midterm elections in which Democrats who campaigned on tighter data-center restrictions won gubernatorial races in Texas and Ohio [1][5]. Torus's pitch is that inference doesn't need sprawling campuses—it needs distributed capacity that can use power already available at the grid edge.

"Most providers are bringing power to the chips. Foundry brings the chips to power," Walkingshaw said [3][4].

What remains unknown

• The valuation and terms of the Series C.
• Whether the 12–16-week timeline holds across diverse utility jurisdictions and permitting regimes.
• How many GPUs "thousands" represents today, and what mix of GPU generations.
• The identity and commitment level of the second AI cloud customer.
• Unit economics: pricing per model, margin targets, and payback periods for customers.

Torus's approach—deploying compute to stranded grid capacity rather than building new generation—addresses a real bottleneck. Whether the modular, behind-the-meter model can scale to tens of thousands of GPUs while meeting the reliability and latency requirements of enterprise inference workloads is the test ahead.