The primary obstacle to the expansion of artificial intelligence is energy consumption, as the existing power grid cannot provide the massive amount of electricity required by data centers. The Utah-based startup, Torus, believes it has found a rapid solution to this problem.
Torus raised $500 million in a Series C round to scale its modular Foundry data center system. Magnetar led the round, with participation from ICONIQ, Origin Ventures, Ridgeline Partners, and Trinity Capital. The Salt Lake City company did not disclose its valuation.
Before entering the AI space, Torus was founded on modular battery systems. Its founder and CEO is Nate Wolkinson. These installations help businesses and factories efficiently use energy and reduce electricity bills.
In September 2025, Magnetar invested $200 million in the company. Now, Torus is applying its energy expertise to the AI sector. Each Foundry module provides power, storage, cooling, networking, and computing in one ready-to-install unit.
One Foundry container houses 576 Nvidia HGX B300 GPUs covering approximately 577 square feet and operates at a capacity of 1.2 MW. For companies requiring additional power, they can simply add more Foundry modules. The facility size range varies from 1 to 20 megawatts.
Most data center builders first secure land and then wait years for new grid connections. Torus is changing this model. Wolkinson explained that most providers bring energy to the chips, whereas Foundry brings the chips to the power source.
The Foundry module is located behind a meter. It utilizes reserve power from existing transformers, thus requiring no new grid connection. Torus's Reserve product, which consists of a flywheel and a battery, compensates for demand spikes and maintains operation during outages.
The closed-loop cooling system does not use water. The main advantage is speed: the time from contract signing to workload operation is between 12 and 16 weeks. Site preparation takes about a week, and installation takes one or two days.
A more prudent point is that Foundry is designed for inference, not training. Inference occurs when a user asks a chatbot a question. Training is a one-time event, while AI inference is a continuous process.
According to the company, market demand for data center inference is projected to grow by 35% annually until 2030, while demand for training will increase by 22% over the same period.
Thousands of GPUs are already deployed in Foundry modules. Torus plans to have over 47,000 such processors in operation by the end of 2027. Hydra Host, an AI infrastructure company, is its first named client. A second, unnamed AI cloud service provider has also joined.
The funds raised will be used to expand Torus's 545,000 sq ft GigaOne factory in Salt Lake City, where nearly 400 people work. This factory's Foundry line is rated for 500 MW of IT load, and Torus aims to increase this figure to 1 gigawatt.
Time is also critical. Large data centers face public opposition. Last week, voters in 40 out of 41 cities and towns in the United States voted against building data centers.
It would be much simpler to install smaller, waterless modules on existing sites. The AI race has begun, and it is not just about chip speed; the decisive factor is the ability of the first company to get them running, and Torus is betting precisely on that.

