SOLYSTOPIAInfrastructure 001
The Ward AI Centre
Build Britain's first experimental ward-scale AI commons — a locally housed computing centre giving residents access to powerful AI without sending their conversations, documents, or community knowledge to an external cloud provider.
A community-owned local computing facility capable of providing private AI, local knowledge, and digital services to authenticated residents without depending on the public internet.
A tiny local model gives the individual baseline capability. The Ward Compute Centre provides the next tier up — capability too expensive or power-hungry for every household to own independently, but that still does not need to be surrendered to a national cloud provider. Internally, we call it Infrastructure 001: the brain. The intranet that eventually carries it to homes is Infrastructure 001B, the nervous system — a separate, larger project.
Layers of computational sovereignty
Tiny offline AI on your laptop or phone
Private, slow-ish, resilient, yours. This is the capability built in Tool 01.
A more capable home server or knowledge store
Family documents, local automation, household systems — still owned by one household.
The Ward Compute Centre
Powerful shared models, a large local knowledge base, specialist tools, and shared GPU compute — too costly for one household, never surrendered to a national cloud.
Optional external capability
Available when useful, but never the foundation the ward depends on.
What the ward centre actually is
A physical local digital territory, not a website that happens to run locally.
One secure room, one rack, one serious AI node, local storage, resilient power, cooling, and a physically separated ward network. Residents connect through Wi-Fi, fibre, or local network access points, authenticate with a ward identity, and use applications served internally. Services keep working even if the external internet fails. We don't need a warehouse full of servers to prove the concept.
HOUSE / PHONE / LAPTOP
│
│ encrypted local connection
▼
┌───────────────────────┐
│ WARD INTRANET GATE │
│ Identity + access │
└──────────┬────────────┘
│
▼
┌───────────────────────┐
│ WARD COMPUTE CENTRE │
│ │
│ AI compute │
│ Local knowledge │
│ Resident services │
│ Storage │
│ Authentication │
│ Monitoring │
└───────────────────────┘
X
NO INTERNET ROUTE
PUBLIC INTERNETThe room is smaller than you'd think
Roughly the size of a modest bedroom.
The first site targets approximately 3m × 4m — around 12m². A standard 42U enterprise rack is only 750mm wide, 1,070mm deep, and 1,991mm tall. The rack itself isn't what needs the space — maintenance access, airflow, electrical distribution, and cooling do. We would actively look for a room inside an existing community centre, church or institute building, library, workshop, or cooperative premises, rather than buying or constructing a standalone datacentre.
┌──────────────────────────────┐
│ │
│ COOLING UNIT │
│ │
│ ┌──────────┐ │
│ │ │ │
│ │ 42U │ │
│ │ RACK │ │
│ │ │ │
│ └──────────┘ │
│ │
│ WORK / SERVICE AREA │
│ │
│ Electrical Secure Door │
└──────────────────────────────┘
~3m x 4m42U RACK
────────────────────
AI compute node
AI compute / future
────────────────────
Fast NVMe storage
────────────────────
Bulk knowledge storage
────────────────────
Authentication server
────────────────────
Application server
────────────────────
10/25Gb networking
────────────────────
Local network gateway
────────────────────
Monitoring
────────────────────
UPS
────────────────────
Spare expansion space
────────────────────The first compute node could start life as a workstation beside or inside the rack enclosure while software is developed, then migrate to proper 2U/4U server hardware. Air-cooled dual-GPU 2U configurations now exist for this class of compute, so it no longer intrinsically demands liquid cooling.
Offline by architecture, not by promise
Offline by architecture. Updateable by deliberate human action.
Saying "not connected to the internet" is a design principle, not a permanent air-gap. A genuinely air-gapped system makes software updates, model updates, security patches, and importing public datasets much harder. A stronger architecture separates two things:
Three zones
Private
Resident data, private AI sessions, personal files. Never leaves the operational network.
Ward
Local datasets, community knowledge, maps, planning information, public documents.
Import
A staging area for software, models, and datasets brought in from outside and inspected before entering the network.
Identity-gated, not identity-tracked
Verify entitlement. Minimise surveillance.
"Identity-gated" does not mean everything you do is attached to your legal identity. The centre needs to know whether a person is authorised to use ward resources. It often does not need to know exactly which authorised resident asked which question.
A privacy-preserving architecture authenticates residency or membership at the gateway, then issues credentials or tokens allowing services to be used without maintaining a permanent, personally identifiable conversation history. Some activities will still need identity — booking scarce GPU time, accessing private storage, administrative functions, safeguarding restricted systems — but ordinary inference should not automatically become a behavioural dossier.
Not just AI model weights
The Ward Library: a curated set of open models, sized to the room.
The first machine targets 2 × 96GB GPUs — 192GB of model-visible memory, enough to run serious open-weight models locally rather than the small 1B–8B edge models Tool 01 uses. The exact model mix will change quickly, so the promise is a maintained library appropriate to different local tasks, not a fixed model name.
Storage is provisioned well beyond what the text library initially needs — the spare capacity becomes a community archive, resident encrypted storage, local maps, offline software repositories, and backups.
001A: the brain. 001B: the nervous system.
How residents physically reach the intranet.
The building
Ward AI available inside the centre itself over secure Wi-Fi/LAN. This is the whole scope of Infrastructure 001A.
Community nodes
Extend it to the library, church, school, community centre, workshops, and participating businesses using point-to-point wireless or existing/private fibre.
Homes
A municipal/community network reaching individual houses — a local ISP without an internet connection. This is Infrastructure 001B, and it is a larger civil-engineering project than the AI centre itself.
Power, heat, and the running bill
The fundraiser can't only fund the hardware.
The dual-GPU pilot workstation alone runs around 2kW under load; storage, networking, UPS losses, and servers add the rest.
Every watt consumed becomes heat in a 12m² room. A duty/standby pair of dedicated server-room cooling units, so one can fail without taking the room down.
At current UK non-domestic average pricing, a continuous 2kW IT load alone is roughly £4,200/year before cooling and ancillary loads.
The budget, at planning stage
~£100,000 builds one working room — not a warehouse.
Using current UK prices as references and adding planning allowances, not pretending we already have contractor quotations:
Not £500k. Not £1m. And not £10k pretending we're going to build something we can't. That £100,000 buys one room, one rack, 192GB of GPU memory, ~50–100TB local storage, a private ward knowledge library, local authentication, local AI services, no public internet dependency, and enough spare rack/electrical architecture to expand — plus an operating prototype another ward can inspect, audit, copy, and improve.
Can a few thousand people collectively own enough compute, knowledge, and networking infrastructure to provide their community with a useful private AI capability?
If it works once, the next question becomes Infrastructure 001B: how do we take the intranet out of that one room and physically spread it across the ward — and then package the architecture so another British ward can replicate it?