We build and operate sovereign AI compute inside the cities where intelligence is made — low-latency, sustainably powered, and engineered for the next decade of demand.
AI is built in the cloud, but it lives in the physical world — in power, in silicon, in the buildings that hold them.
BMETAL is a vertically integrated compute company. We secure urban sites and the power to run them, install and operate accelerated hardware, and deliver it as bare-metal, inference, and private post-training capacity to teams building frontier products.
The result is compute that sits closer to users and data, runs cleaner, and is held to a standard of reliability that production AI demands — without the trade-offs of legacy, out-of-town facilities.
One stack. Land, power, silicon, and serving — under a single roof.
Bare-metal GPU
Dedicated accelerated hardware with full isolation and no noisy neighbours — reserved capacity, on-demand bursts, and private pods where teams tune models on data that can't leave their control.
Inference
Production-grade inference served from inside the metro — disaggregated across the silicon that fits each phase of the model, with token endpoints and managed serving that keep models close to the people using them.
Power & heat
Behind-the-meter generation and intelligent thermal design — including reuse of waste heat — so density scales without burdening the grid or the neighbourhood.
performance.
Every millisecond between a model and its user is a tax on the product experience. By siting compute in the metro core, we collapse that distance — and the economics that come with it.
Built to a higher standard, end to end.
Secure the ground
We control urban sites and the power behind them — the scarce inputs everything else depends on.
Engineer for density
Purpose-built thermal and power design lets us run modern accelerators where others can't.
Operate the silicon
We run the hardware ourselves, with an in-house team and a single accountable standard of service.
Deliver as a platform
Capacity reaches customers as clean, reliable bare-metal and inference — abstracted, not exposed.