Torus launches modular AI data centers for sites with spare grid power

Torus launched Foundry on November 11, a prefabricated AI data center designed to draw power from spare capacity at existing industrial and commercial sites. It is built for inference, the work of running an already trained model to answer requests. Torus puts the size of a site at 1 to 20 megawatts and says the full deployment takes 12 to 16 weeks from contract to running workloads.

The launch gives cloud operators another way to add inference servers where a property already has enough electrical headroom. Companies can also contract for dedicated capacity operated by Torus. The Utah-based company announced $500 million in financing led by Magnetar on the same day. It said the money would expand manufacturing and pay for more deployments of the compute and storage equipment in the United States.

Storage covers peaks on the existing connection

Foundry sits on the customer side of the electricity meter. A building may have a connection sized for its busiest hours, with capacity left over when its equipment is running below that level. Torus looks for sites with room to add a new load. As the building's own demand rises, the amount of electricity the servers can draw directly through the same connection falls.

Torus pairs the servers with an energy storage system called Reserve. According to its power design notes, batteries charge when site demand is low and discharge when the building and computing equipment need more power at the same time. The storage system supplies the difference to keep the draw from the grid within the site's connection limit. The energy in storage and the time available to recharge it set limits on how long the arrangement can cover an elevated load.

For example, if a site needs half a megawatt above its connection limit for two hours, storage has to supply one megawatt-hour before accounting for losses. Megawatts describe the rate of power draw, while megawatt-hours measure the energy delivered over time. That calculation illustrates the units. It uses a hypothetical site.

Reserve combines lithium iron phosphate batteries with a flywheel. The flywheel stores energy in a spinning rotor. Torus says it is designed to deliver power when demand from the GPUs jumps and taper its output as the grid or on-site generation responds. The US Department of Energy describes flywheels as a technology suited to supplying power over short intervals.

A module and a site have different power ratings

One configuration in Torus's specifications has an IT load of 1.2 megawatts across eight server racks and four networking racks. The modular design lets buyers add equipment in increments of that size. The larger figure of 1 to 20 megawatts describes a site. How much equipment can run there also depends on the existing electrical connection and on other demands for power, including cooling and charging the storage system.

The equipment is assembled and tested at the company's GigaOne factory in Salt Lake City. Torus says preparing a site takes about a week, and installation and commissioning take another one to two days once the equipment is on site. Those are individual stages of the 12-to-16-week schedule the company gives for a full deployment. The product specifications describe a closed circuit that circulates coolant directly to the chips. Torus says the cooling system can operate without a water hookup at the site.

Torus is offering three ways to buy capacity. A customer that supplies its own GPUs can get an enclosure and power equipment. A second option comes with the GPUs installed. Under the third, an enterprise contracts for inference on hardware dedicated to its own workloads. The modules can be added at one location or spread across existing sites.

Torus named GPU cloud provider Hydra Host and an undisclosed cloud provider as early customers. The company says the Foundry production line is operating in Salt Lake City and that thousands of GPUs are already deployed. Its target is to have more than 47,000 GPUs deployed in Foundry units by the end of 2027.