AI inference from telco sites

Many small sites, run as one inference provider.

Shoal turns the spare power in an operator's regional sites into one AI data centre. Grid connections that are already in service are run as a single estate and sold as contracted GPUs and inference under the operator's own name.

The problem

AI inference is growing faster than new power can be connected.

The build-out of AI data centres is slowed by access to power and planning permission as much as by access to the latest GPUs. A new connection can take years. In three years it will be a different market.

Aggregate your sites into one AI data centre.

Requests arriveManchester · power curtailedheld at its floor · 48 kW spareDublin 7.3 ms143 GPUs idle · 1.3 MW spareFrankfurt 10.7 ms113 GPUs idle · 1.9 MW spareMunich 14.8 ms83 GPUs idle · 1.0 MW spareLisbon 23.8 ms83 GPUs idle · 1.2 MW spareAthens 36 ms · batch46 GPUs idle · 70 kW spare10 ms · priority25 msstandard
From the console's map: requests arriving at Slough stay within the latency each tier promises, then go to the site with headroom and the cheapest power. Batch can go anywhere. Curtailed Manchester takes no new work. Idle GPUs and spare power as the demo showed them at 10:00 CEST on 9 October.

What Shoal does

Three jobs, one platform.

01

Aggregate

Run many under-used grid connections as one AI data centre, starting now.

€29Mof margin lost for every year one megawatt of B200s waits for power.

B200s, 60% busy, today's median token prices. GPUs paid off over four years.

  • Describe each site once.Grid connection, connection terms, battery, building, latency, power price and models. The terms compile into rules the platform enforces.
  • Every request goes to the right site.The model is loaded there, the site has room within the power it holds this hour, the customer's residency allows the country, and the tier's service level holds. Among those, the cheapest power wins.
  • Rules decide; AI explains.Each placement comes with its reason in plain words.
02

Upgrade

Grow each power connection on variable terms, and still sell a predictable service.

  • Take the upgrade now.Curtailable, timed and load-managed capacity is on offer today in the UK, California, the Netherlands and Germany, without waiting for the network to reinforce.
  • Inference moves to where the power is.When a site is curtailed, live requests move to sites with headroom and batch work waits for plentiful hours.
  • A stable service level on variable power.Customers get one predictable inference SLA. The estate absorbs each site's ups and downs.
03

Sovereign AI

Every country must build its own sovereign AI.

  • Capacity inside the border.Residency is a placement rule: a request never leaves the countries its customer allows.
  • Built from what a country already has.Connected, permitted buildings across the country, sold by local operators under their own brand.
  • Buyers already pay more for it.Sovereign cloud lists at 10–30% above public cloud (BCG, 2025).

Two products on one platform.

A console for the operator to run the estate, and a white-label service its customers buy from. Customers see one endpoint and one contract. They never see the sites, or the power contract behind each one.

Operator console overview with live figures and site map

The console, for the operator

Sites, power contracts, placement, utilisation and customers.

White-label customer portal listing open models run in Europe, with context length, input and output prices per million tokens, and the countries each runs in

The service, for its customers

A white-label portal: models, prices, keys and usage. OpenAI-compatible, so customers change the base URL and keep their code.

Placement screen: a live log of requests with customer, tier, model and the site each went to, and a panel stepping through the rules that chose Munich

Screens from the working prototype, running on live power prices. Veyra Telecom is a demo operator.

How it works

Uncertain power in, predictable service out.

Firm power carries the firm promises. Flexible power runs work that can wait or move. When a site is curtailed, the customer never notices.

Handover console, Manchester site during a curtailment: held to 1.06 MW with the battery covering 400 kW, tokens served by tier over two days, and the power drawn from the grid and from the battery
Manchester, mid-curtailment in the live demo: held to 1.06 MW, the battery covers 400 kW and tokens keep flowing.

1 · Seconds

Batteries cover the gap

2 · Minutes

Work moves to sites with headroom

3 · Estate full

Overflow to partner capacity

Next

Start with a pilot on a handful of your sites.

We work the case through with you from your site list, connections and power contracts, and the upgrade terms your network operator would offer, before a single GPU is bought.

From you

  • The site list, with each grid connection and equipment load
  • The power contracts, and any upgrade enquiries in progress
  • GPUs installed or planned
  • First customers for GPUs or inference, or a broker

From Shoal

  • The platform, run for you
  • The console and the white-label portal
  • Which sites to upgrade, and on what terms
  • The business case, site by site

Get in touch

Your connections, run as one AI service.

Move bits, not watts.

contact@shoalai.net Email us