Enzwa lab
Conduit
Agentic Data-Centre Delivery
Live Demo
After AIPCon 10

A $40M transformercan idle a $2B campus.

The bottleneck for AI compute has moved from the GPU rack to the substation. Conduit puts a mesh of autonomous agents over the entire build — grid, power, cooling, hardware — so a slip in one long-lead item re-plans the whole schedule before a human picks up the phone.

24–48moHV transformer lead
<10%of capex gates 90%+
The constraint

The chain is the product

A gigawatt campus is no longer an IT problem. It is a global supply-chain problem with a power-train on its critical path.
From rack to substation

90% of the spend can't switch on without the other 10%

Batteries, transformers, switchgear and breakers are under 10% of a data-centre's construction cost — yet without them, the shells, cooling, racks and GPUs that make up the other 90%+ cannot be energised. High-voltage transformer lead times have stretched from 12–18 months to as long as 24–48, so procurement has moved from a purchasing detail to the single biggest schedule risk in the build.

A $2B campus can sit dark, waiting on a $40M order — and a static Gantt chart finds out too late.

The energisation problemGrid → Substation → Compute

Conduit models every physical dependency as a software-defined workflow. When reality moves, the model re-solves — proposing alternate logistics, suppliers or sequencing, weighed against what you're trying to achieve.

The control room

Set the objective. Break the chain. Watch it re-plan.

An interactive model of an 80MW build. Everything below recomputes live.
Conduit Orchestrator

Delphi North · 80MW · 24,000 GPU target

A mesh of agents over grid interconnection, the power train, cooling and hardware. Pick what the build should optimise for — the orchestrator re-weights every routing decision and lead-time buffer to match.
? What are you trying to achieve?
Critical path · grid to compute online
On track Watch Critical

Inject a disruption

Reality moves. Fire an event and the orchestrator cascades it through the chain, then proposes a recovery shaped by your objective.
Agent reasoning Idle
Under the hood

What the agents watch

Physical operations turned into live, software-defined signals — the raw material for autonomous decisions.
Grid & power

Interconnection queue position, substation energisation dates, transformer & switchgear OEM slots and expedite windows.

Logistics & inventory

In-transit shipments, port and customs dwell, on-site staging capacity, and buffer stock on long-lead items.

Construction sequence

Shell readiness, mechanical/electrical fit-out, commissioning gates and crew availability against the energisation date.

Compute demand

GPU allocation, model-training timelines and the revenue clock — so the plan optimises for value, not just delivery.

The physical stack

Software-defined, steel-bound

Every tile below is a node the orchestrator reasons over.
The white spaceRacks · GPUs · network
Power trainUPS · switchgear
The gridHV interconnect
Liquid coolingDirect-to-chip
LogisticsGlobal delivery
The loop

How the orchestration works

01
Model

Every dependency — physical and contractual — becomes a node with lead times, suppliers and risk in a single ontology.

02
Sense

Live signals from grid, vendors, transit and site continuously update the model. The plan is never stale.

03
Simulate

When something slips, agents cascade it downstream and generate recovery options against your objective.

04
Act

The orchestrator proposes — and where authorised, executes — re-routes, expedites and re-sequencing in real time.

The idea behind Conduit

Inspired by Nscale's "Agentic Supply Chain for Data Center Delivery," presented on Palantir AIP at AIPCon 10 (June 2026). This is a concept piece — figures are illustrative of the 2026 data-centre delivery landscape, not a specific project. Watch the original demo →

Conduit Concept · Illustrative model