Compute that lives behind the meter.
Britain can't connect new data centres fast enough. So we stop waiting for the grid — and run AI compute on power that's already there, behind the meter, inside managed estates.
- 01Hosts earn from power they already pay for.
- 02Buyers get UK compute without the queue.
- 03The grid gets a load it can turn down.
The power was always there — now it works.
Power flows the way it always has. Compute fills the gap.
A node uses only the electrical headroom a building isn't using — and steps aside the moment residents need it.
Onboard an estate
A building with verified spare power. One landlord supply, installed at no upfront cost.
Supervise within limits
Energy control and battery keep each node inside the building's own limit. The resident supply is never touched.
Route to headroom
BayesOS sends each job to whichever node has room right now — and moves it the moment a node has to step back.
Every building is wired for its busiest moment and rarely gets there, so there's electrical capacity sitting idle most of the day. A BayesNode draws only that slack — never the supply residents depend on. When the building needs more power, the node quietly backs off; when there's room, it fills it. Nothing changes for the people living there. The spare capacity simply does useful work.
The full technical brief.
Safety framework, security model, rollout maths and references.
The compute that belongs closer to people.
BayesGrid suits workloads that distribute across smaller footprints and gain from being near users and data. It complements hyperscale data centres — it doesn't replace them.
AI inference & RAG
the workload most aligned to modular, distributed capacity
Cloud gaming
proximity directly improves the experience
Rendering & streaming
locality eases backbone congestion
HPC & simulation
distributed capacity for bursty compute
Not for frontier model training — that needs single-site cluster scale and belongs in centralised facilities.
Infrastructure, not improvisation.
An always-on load in a residential building has to satisfy Great Britain's electrical assurance framework — the same layered stack that installers, network operators and insurers already recognise.
Designed for multi-tenant security: encrypted communications, workload isolation and tamper-evident hardware, with credentials revocable the moment a node is removed or compromised.
Safe operating limits, controlled export, certified storage and conversion — assured before a single node is energised.
Prove it small. Scale it nationwide.
We begin with a 50-node proof of concept across a small number of UK estates, then scale node by node and estate by estate toward an annual capacity of over 100 MW of AI inference compute — roughly 1.5% of the UK's stated 2030 AI ambition, achieved without breaking ground on a single new hyperscale site.
- Stage 0150-node POC
Across a small number of UK estates
- Stage 02Multi-estate fleet
Node by node, estate by estate
- Stage 03100+ MW
Annual AI inference capacity
A quiet box in the plant room.
Each BayesNode is a matte, liquid-cooled cabinet that sits beside the building's existing electrical kit — 8× B200 SXM compute, a coolant distribution unit, battery ride-through and a circuit-level energy gateway, with an external dry cooler for heat rejection. Engineered to be safe, quiet and unremarkable.
- 8× B200 SXM · NVLink
- ≈13 kW operating (power-capped)
- direct-to-chip liquid cooling
- battery ride-through
- circuit-level energy control
Compute that lives behind the meter.
Get in touch — we're rolling out the first 50 nodes now.