Networks & InfrastructureData CentersAIEnergy StorageGrid Edge
Torus raises $500m, launches modular AI data centers at grid edge
Torus raised $500m in Series C funding led by Magnetar and launched Foundry, a 1-20MW modular AI data center unit built in Salt Lake City, targeting 47,000 GPUs by end-2027.
Why it matters
- Torus raised $500 million in Series C funding led by Magnetar to scale its Foundry modular data center product.
- Foundry units deliver 1-20MW per site in shipping-container-sized chassis, installed behind the meter on existing transformer headroom.
- Contract to operations takes 12-16 weeks; site prep takes about a week, installation one to two days.
- Torus plans more than 47,000 GPUs deployed in Foundry units by the end of 2027.
- The Salt Lake City factory, 545,000 sq ft, will reach 500MW of IT load capacity, with plans to expand to 1GW.
The story
Utah energy storage developer Torus has raised $500 million in Series C funding and launched Torus Foundry, a prefabricated modular data center product it says delivers 1-20MW per site in shipping-container-sized units. Magnetar led the round. The money will fund scaling of the Foundry line, which Torus positions as an answer to the power bottleneck constraining AI compute deployment.
The design premise is simple. Foundry units install behind the meter, using headroom at existing transformers rather than new grid connections or generation. Each unit ships with Torus Reserve, a flywheel-and-battery storage system that the company claims smooths AI inference loads, squeezing more capacity out of an existing connection while providing backup power during outages. Because the units run off existing grid capacity, Torus says, they avoid new power construction entirely.
Timing is the commercial pitch. Site preparation takes roughly one week. Installation and commissioning take one to two days. The full cycle from signed contract to operations runs 12 to 16 weeks — a fraction of the multi-year timelines typical of new-build data centers and grid interconnections.
Torus manufactures the units at a 545,000 sq ft (50,630 sqm) factory in Salt Lake City. Each unit is vertically integrated, combining power, storage, cooling, networking and compute in one chassis, and uses closed-loop cooling.
What does Foundry change for AI capacity buyers?
Customers have three purchase options: a powered shell without GPUs, a turnkey bare-metal unit with GPUs, or dedicated single-tenant inference capacity. The first announced customers are GPU cloud operator Hydra Host and a second, unnamed AI cloud provider.
Torus founder and CEO Nate Walkingshaw framed the product against an industry building net-new infrastructure to feed AI demand. "Inference doesn't have to wait," he said. "We've spent years learning how to make the grid more affordable, flexible, resilient, and intelligent, and Foundry applies that to compute: capacity in weeks, with power that's already available."
He added: "Most providers are bringing power to the chips. Foundry brings the chips to power."
That is a marketing claim as much as an engineering one. Building net-new generation and interconnection, Walkingshaw argued, "causes costs to potentially rise and takes years to bring compute online." The counter-argument embedded in Foundry is that distributed storage plus existing transformer headroom can host meaningful inference capacity now, at 1-20MW increments, without waiting on utilities.
What is the deployment target?
Torus says thousands of GPUs already run in Foundry units. The company plans more than 47,000 GPUs deployed across Foundry sites by the end of 2027. The Salt Lake City plant will be capable of producing 500MW of IT load, with expansion to 1GW planned in the near future. Torus gave no date for that expansion, and no revenue or backlog figures accompanied the announcement.
The $500m raise and the unnamed second customer leave open questions about demand concentration, but the order of magnitude is clear: a storage vendor with half a billion dollars of fresh capital is now competing in a modular data center market where speed to power, not raw construction, sets the pace. Whether flywheel-buffered transformer headroom can carry large inference workloads at scale will be tested against the 47,000-GPU target over the next two years.
Also reported
Source: datacenterdynamics.com
More from James Calloway
Show full bio
Staff writer covering companies, deals and strategy in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
499 articles