Power bottleneck
Grid connections are the new scarcity. Large datacenter capacity queues now run into years, not months.
Sovereign AI inference / The Netherlands
Compute, close to home.
Inference nodes in Dutch homes and buildings, running open models on surplus and off-peak power, under European law.
The problem
A single building is the wrong shape for compute.
Europe is trying to fit the next decade of AI into a handful of giant campuses. They take land, water, power, and grid capacity, then leave behind fenced-off industrial boxes at the edge of towns. The grid will not allow endless repetition of that model.
Wasted surplus
Sunny afternoons, windy nights, and negative price hours should become useful work instead of curtailment.
Land and water
Hyperscale campuses are ugly because they are blunt instruments: land-hungry, water-hungry, and hard to place near people.
Data sovereignty
European prompts, documents, and workloads should not depend on infrastructure controlled from abroad.
Political risk
Compute concentrated in a few firms and jurisdictions is a dependency. Europe needs another shape.
Europe's AI should run on European ground. Distributed / close to home
Manifesto
Energy is local. Compute should be too.
The cloud taught us that a data center must be enormous, central, and somewhere else. That shape is becoming a liability.
A node is a flexible load. It ramps up when energy is cheap or in surplus, then stands down when the grid is tight.
Distributed across thousands of buildings, inference can stay inside European borders, close to the people using it.
How it works
Three layers, one quiet box on a wall.
Instead of pouring one more concrete campus, distribute the load across places that already have walls, roofs, meters, batteries, and surplus power. The network does the hard work of routing requests to the nearest eligible node.
Photoreal concept render
01 / The node
A sealed outdoor unit.
Installed against a wall, in a shed, on a farm, or at an SME site. Quiet, weatherproof, tamper-aware, and designed to be serviced without becoming a household project.
- Compute4-8 inference GPUs
- Noise<55 dB target
- SecurityTPM / HSM / attestation
02 / The energy layer
It runs on the swing.
The node tracks solar surplus, day-ahead prices, and local load. It works when the grid can support it and pauses when it should.
- SignalPrice + grid load
- InputP1 smart meter
- BehaviorFlexible load
03 / The network
The nearest eligible node answers.
Requests use an OpenAI-compatible API and are routed by locality, availability, tenant isolation, and jurisdiction.
- RoutingLocality-first
- ModelsOpen weights only
- ResidencyEU enforced
Locality tiers
Distance is the latency.
Because the network is made of where people live and work, the compute can sit nearby instead of a continent away.
<5 ms
Hyperlocal
Your own node, on-premise or at home.
<20 ms
Neighborhood
Within roughly 10 kilometers.
<50 ms
Regional
Same-country clusters for SMEs and developers.
<100 ms
National
Country-level pool for batch and bulk inference.
Host a node
Your building uses power. Let it get paid.
We install and maintain the node. You provide a wall, connectivity, and a grid connection. The model only works when hosting is simple and worth it.
Energy covered
Every watt the node draws is settled transparently. Hosting should not land on your electricity bill.
Paid to host
A monthly payment plus a share of served inference, visible down to energy and request volume.
Backup buffer
The integrated battery can support essential loads during outages.
Full control
Set opt-out windows, cap draw, pause service, or disconnect. It is your building.
Join the waitlist.
The first pilot is for hosts, developers, public-sector teams, and regulated organizations around Haarlem and Kennemerland.
You're on the list. We'll be in touch.