Compute where the power already exists
Why the constraint on AI capacity is electricity rather than silicon, what a DAIT node is, and how it reaches a building that already exists.
The bottleneck is electricity, not chips
Demand for AI compute is not limited by how many processors can be manufactured. It is limited by how quickly a building can be connected to enough power — and that is a planning problem, not a technology one.
2,061 GW is waiting
That much generation and storage capacity sits in United States interconnection queues. The median project waits more than five years to reach commercial operation.
Most of it never gets built
Fourteen times more capacity has been withdrawn from those queues than has actually been completed. Getting in line is not the same as getting connected.
Capital sits idle meanwhile
A centralised facility needs land, permits, substations and years of construction before it serves a single request. The money is committed long before anything runs.
Source: Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition.
Put the compute where the power already is
Most commercial buildings already have an energised electrical service with headroom in it. A stadium, a hotel roof, a campus plant room, an industrial pad. DAIT installs a sealed compute node on that existing service instead of waiting for a new one.
- Existing grid connectionThe node draws from a secondary power drop that is already in place. No new substation, no interconnection application.
- No facility buildThe enclosure is the facility. There is no data hall to construct, no land to acquire, no rezoning to apply for.
- Distributed by designCapacity lands near the people and systems using it, rather than in one remote campus.
Why distributed rather than simply bigger
The obvious objection is that one large facility is more efficient than many small ones. For training that is true. For serving, three things work against it.
Latency is physics, not engineering
Every kilometre between a request and the processor answering it adds delay that no amount of optimisation removes. For interactive inference that shows up directly in how the product feels.
Some data cannot travel
Where regulation requires processing inside a jurisdiction, a distant campus is not a cheaper option — it is not an option. Placement becomes a compliance control.
Small sites clear faster
A large build needs land, permits, a substation and years. Attaching to service that already exists needs none of them. The constraint is procedural, and distribution routes around it.
What changes when compute moves into the building
Putting a node on an occupied site rather than in a dedicated facility changes what the infrastructure has to survive, and who benefits from it.
- The enclosure is the facilityThere is no building around it doing the work. Weather, dust, corrosion, heat and fire all have to be handled inside a sealed unit, outdoors, for years.
- The hardware sits on someone else’s propertyWhich is why trust cannot be assumed from network position. Every unit has to prove its own identity, and physical tampering has to be survivable.
- The host gains something tooStorage carried for the node’s own resilience is available to the building during an outage. Infrastructure that would be a pure cost in a data centre becomes a shared benefit here.
- Failure means something differentOne cabinet is a fraction of a distributed fleet rather than a single point of failure. The unit of loss is smaller, and so is the blast radius.
The enclosure itself, the tiers and the mesh software are covered on Technology. Which sites qualify is on Deployments.
See it explained
Coming soon
A short walkthrough of the node, how it mounts and how it connects is in production. It will appear here.
Coming soon
Go deeper
Read the version written for your side of the table, or check the questions people ask most.
Open the glossary