Keraunos operates GPU clusters on energy that is already generated, licensed and connected. One team from the substation to the endpoint. Capacity arrives on the date we give you.
Reserve a cluster and run it yourself, or hand us the workload and consume it as a service. The hardware is current-generation NVIDIA across both, and the contract length is yours to choose.
Isolated multi-node clusters for pre-training and large-scale fine-tuning. Bare metal, Kubernetes or virtual machines, on high-bandwidth fabric with parallel storage attached.
Dedicated endpoints for open-weight and custom models, deployed and operated by our engineers, with capacity held for your peak rather than borrowed from a pool.
Managed training workflows to adapt open models to your data, then promote them straight onto your own endpoints without moving the weights off our infrastructure.
Near-term capacity comes from existing Tier III halls we retrofit. Long-term capacity comes from campuses we develop on power we have optioned. Both pipelines are running now.
Capacities are contracted or optioned positions at various stages of development. Site names and counterparties are disclosed under NDA.
Full footprint →Grid connections in the crowded markets run two to eight years. Brazil's northeast has renewable generation already built and already connected, waiting for load to arrive.
How we develop infrastructure →Contracted directly with generators, fixed or indexed, sized to a 24/7 AI load rather than an office profile.
Positions in Ceará and Rio Grande do Norte, where curtailed renewable output is looking for exactly this kind of load.
Transmission-line options paired with on-site generation, so a campus is not one feeder away from an outage.
Existing Tier III halls with energy and licences in place carry the near-term fleet while the campuses are built behind them.
A preliminary framework over a multi-gigawatt Argentine renewable portfolio extends the same model beyond Brazil.
Origination, energy, construction and cloud operations sit on one balance sheet. Decisions that usually cross four counterparties get made in one room, and the delivery date reflects it.
Performance, uptime and cost are answered by the team that owns the racks, not escalated to a landlord.
South American jurisdiction, data residency in country, and commercial terms that hold across decades rather than election cycles.
One retrofit spec, one operating standard, one commercial template. Each site is faster and cheaper to bring online than the last.
AI labs, application companies at production volume, enterprises with residency requirements, and public research programmes.
We will come back with a cluster specification, a region and a delivery date. Site-level detail and counterparties are shared under NDA.
Reserve capacity →