Torus says Foundry goes from contract to token in 12 to 16 weeks. Each unit is a prefabricated 1.2-megawatt block holding 576 Nvidia HGX B300 GPUs in a 577-square-foot footprint.

The comparison number is closer to four years. That is the large-load interconnection wait in the tightest US data-center markets, and it is why hyperscalers keep signing long leases on campuses that will not energize until 2028.

Our read: the 12-week product is a genuine wedge, and it does not repeal the queue. Torus is selling the half of the grid that already exists, at a scale the queue was never built to serve.

The $500M round led by Magnetar is the industrial version of that bet. It pays for GigaOne, the 545,000-square-foot plant in Salt Lake City, and for Foundry deployments across the United States.

The plan is 47,000 GPUs by the end of 2027. Thousands are running now for Hydra Host and a second AI cloud provider that Torus has not named.

Why the queue is the real constraint

America's data-center problem is usually described as a generation problem. It is a delivery problem. A new substation, transformer bank and transmission upgrade take years, and the interconnection queue is where those years are spent.

The grid itself is not empty. It was sized for a handful of peak hours a year, so most of its capacity sits idle between them. The scarce hardware is the transformer, the switchgear and the service limit between that capacity and a rack.

That is the gap Torus is attacking. Foundry installs behind the meter, on headroom at an existing transformer, and pairs with Reserve, the company's flywheel-and-battery storage. Reserve charges when the site is quiet and discharges into GPU load spikes, so the utility sees a smoother draw under the service limit.

The company calls this "bringing chips to power." The engineering is less slogan than load management: a 1.2 MW unit with eight server racks at 125 kW each, four networking racks, and a closed-loop direct-to-chip cooling loop that uses no on-site water.

As we wrote in November, Brazil's 38 GW AI data-center queue is a siting problem before it is a build-out. Torus's answer is to stop sitting in the queue at all.

The load-shaping claim is the part worth testing. AI inference draw is spiky, closer to an EKG than a baseline. A storage buffer that absorbs those spikes can fit more compute on a connection that a utility already approved, which is a real operational gain and not a marketing one.

It is also where the model gets harder. Headroom exists because other customers are not using it at that moment. A foundry-sized load parked behind a transformer changes the site's demand profile, and utilities will want to see the operating record before they bless a fleet of them.

The case for the 1.2 MW block

Inference is the workload that fits this shape. Training needs thousands of GPUs wired into one failure domain. Inference serves one request at a time and tolerates being spread across many small sites.

Torus puts the split at the unit level. One Foundry is 72 nodes, 576 B300 GPUs, 80% liquid cooling and 20% air. It ships pre-tested, installs in one to two days, and scales in 1.2 MW increments.

The commercial logic follows the rack. At 125 kW per server rack, a hall built for 20 kW racks cannot absorb a Blackwell-class deployment without a rebuild. A prefabricated unit is sized for that density from the start, and Torus says the power path is ready for NVL72 and the next Vera Rubin generation.

The floor plan is deliberately boring. 577 square feet, eight server racks, four networking racks, no water connection, no building construction. Boring is the point when the alternative is a multiyear build.

🎯
Inference capacity is becoming a siting race, not a model race.

The scarce asset is transformer headroom, not generation.

Torus has a manufacturing plan. It does not yet have a disclosed backlog.

Torus sells three configurations: a powered shell without GPUs, a turnkey bare-metal unit with GPUs, and single-tenant sovereign capacity for enterprises that want exclusive hardware.

The demand argument comes from McKinsey, which Torus cites for 35% annual growth in inference data-center demand through 2030, against 22% for training. Treat vendor-cited forecasts with care. The direction is consistent with every serious capacity estimate published this year.

The customers matter more than the forecast. Hydra Host is a bare-metal GPU cloud, which means the first buyer is a company that already resells compute and understands utilization. That is a better signal than a logo from a model lab.

576 GPUs per 1.2 MW unit

Foundry unit density

Each prefabricated unit carries 72 nodes and eight 125 kW server racks. Torus, 2026

Two ways to buy a megawatt

The greenfield campus is still the right answer for training. It concentrates power, cooling, security and staff in one place, and it is the only practical way to wire tens of thousands of accelerators into one cluster.

The Foundry model is the right answer for a different buyer: a neocloud, a telecom, an enterprise or a compute aggregator that needs 1 to 20 MW near a specific customer and cannot wait for a campus.

The trade is real in both directions. Distributed sites give up scale economies and some redundancy. They gain speed and siting freedom, which is the binding constraint for the next two years.

This is not a one-company bet. EPRI, Nvidia, Prologis and InfraPartners have been designing a distributed-inference pilot since February, siting 5 to 20 MW micro data centers at utility substations and shifting load between them. The category is moving from slideware to interconnection agreements. Torus is the commercial, productized version of the same argument.

ParameterFoundry unitGreenfield campus
Time to first token ✔ 12–16 weeks ✗ roughly 3–4 years
Unit of capacity 1.2 MW prefabricated 100 MW+ campus
Power source ✔ existing transformer headroom ✗ new interconnection
Cooling closed-loop, waterless often evaporative
Best workload distributed inference ✔ training, dense clusters

Torus product specifications and interconnection timelines, 2026

Where the thesis breaks

Transformer headroom is finite and unevenly distributed. Torus needs sites where the utility has already allocated capacity and the load shape can be smoothed. Those sites cluster in commercial and industrial parks, not necessarily where AI customers are.

The 47,000-GPU figure is a production target, not contracted demand. One customer is named. The second is not. That is the number to watch in 2027, because a second disclosed buyer would prove the product is repeatable rather than a single flagship deal.

GigaOne's Foundry line is rated for 500 MW of IT load, with a stated path to 1 GW. A factory that opens a production line every 21 days is impressive. A factory is not a grid connection, and the bottleneck can move from the assembly floor to the utility desk.

Security is the other open item. Distributing valuable accelerators across many behind-the-meter sites changes the physical threat model. Torus bundles networking and monitoring into its Foundry Protect layer, but the operating record is short and the customer list is thin.

There is also a ratepayer question. Behind-the-meter load that uses existing headroom is cheaper for the operator. Whether it is cheaper for the distribution network depends on how the utility prices the flexibility, and that negotiation will differ state by state.

The other risk is operating skill. Building racks is not the same as running a fleet of them across dozens of leased parcels. Torus says it will handle lifetime operations and maintenance. That service burden grows linearly with every site, and it is the least visible line in the business model.

What would change our mind

Three disclosures would move this from a plausible wedge to a proven model. A named second customer with a dated deployment. An interim GPU count during 2027, not just the year-end target. A cost-per-deployed-megawatt comparison against a greenfield build of similar capacity.

Until then, the honest summary is narrower than the press release. Torus has found a real inefficiency in how US capacity is delivered, and it has raised enough money to test whether the inefficiency is big enough to build a business on.

The 12-week clock is the product. The four-year queue is still the market.

Torus launches Foundry and raises $500M for modular inference capacity
Primary announcement: the 1–20 MW per-site range, the 12–16 week deployment claim, the Magnetar-led round and the 47,000-GPU target for 2027.
The source for every capacity and timing figure in this piece. Read the deployment models closely: only the turnkey tier includes GPUs.
Foundry specifications: 1.2 MW, 576 B300 GPUs, 577 square feet
The engineering sheet behind the pitch: rack loads, node count, cooling ratio, footprint and the NVL72 upgrade path.
Useful because it is specific. A 125 kW rack and a waterless loop are the two details that separate this from a container with GPUs in it.
Torus raises $500M to run AI on spare grid power at existing sites
Independent account of the round, including Bloomberg's earlier reporting, the Series C framing and the absence of a disclosed valuation.
The cleanest third-party summary. It also flags what the company did not say: no valuation, no named second customer.