enzwa lab · conduit · agentic data-centre delivery
Conduit models an 80MW build as eight linked orders with real lead times. Move one of them and the model re-solves the whole schedule. This page runs the plan as signed down the left, and the same build after the transformer slipped down the right. They start on one line and never meet again.
A tool for showing how one late order moves the day a data centre can switch on.
A large data centre is a chain of orders, and some of them, like the high-voltage transformer, take years to arrive. A normal schedule records a delay only once somebody reports it, and by then the room to recover has gone.
Conduit models the whole build as linked orders with real lead times. Move one and it re-plans the entire programme in front of you, including the date the site starts earning.
Both columns below start from the same June 2026 date and the same eight orders. On the right, the transformer OEM pushes its delivery slot by sixteen weeks. That is the only change made to the model.
Delphi North. 80MW, 24,000 GPUs, a two billion dollar campus. Eight orders on the critical path.
Fifty-three weeks to a firm energisation date. It runs alone: nothing in the chain overlaps it.
Eighty-four weeks, single source, ordered in week six. That is the earliest the design releases it.
The order already sits as early as the design allows. There is nothing left to pull forward.
Switchgear, UPS, cooling, white space, GPUs. Each overlaps what precedes it, and the plan already counts that overlap.
Seven weeks of integrated systems test, then energisation.
The campus starts earning.
The same site, the same eight orders, the same start. Nothing has moved yet.
The utility holds its date. The queue position stands.
The OEM moves the slot sixteen weeks. This is the only change on the page.
Ordering on day one returns six of the sixteen weeks. There is no earlier than day one, so ten are carried.
All five finish ten weeks late. Their overlap was spent in the plan, long before the slip arrived.
The same seven weeks of test, ten weeks later.
Same building, same order book, ten weeks of nothing.
both programmes to scale · swipe to read the full 172 weeks
Conduit lets the chain break anywhere. The same sixteen weeks are worth different amounts at different points, because each order carries a different share of overlap against the one in front of it. The outline bar is the delay injected. The solid bar is what arrives at the energisation date.
computed from Conduit’s chain, speed-to-power objective · swipe for the full scale
| disruption | injected | at the end | why |
|---|---|---|---|
| Interconnection study delayed | 20 wk | 20 wk | The utility node runs alone. Nothing overlaps it, so it passes through whole. |
| HV transformer slips | 16 wk | 9.7 wk | Six weeks come back by ordering on day one, and day one is the floor. |
| Cooling vendor defaults | 12 wk | 3.6 wk | Cooling runs largely in parallel, so most of the delay disappears into the overlap. |
| GPU allocation pulled forward | 10 wk early | 5 wk earlier | Good news travels at half speed. Half of it is absorbed by the same overlap. |
Grid interconnection and the high-voltage transformer are the two orders that pass a delay through untouched. They are also the two with the longest queues in the real market.
| the real constraint | today | context |
|---|---|---|
| Large power transformer, 100 MVA and above, US delivery | 128 to 143 weeks | 7 to 14 months before 2020 |
| Medium-voltage switchgear, 15 kV | 52 to 80 weeks | manufacturers sold out into 2028 |
| Interconnection request to commercial operation, median | over 5 years | around 2,060 GW queued across 8,200 projects |
Power infrastructure is 40 to 50% of what a data centre costs to build, and the queue figures come from Lawrence Berkeley National Laboratory’s Queued Up, 2026 edition, covering projects to the end of 2025. A shell with no energisation date is a finished asset earning nothing.
A static schedule records a slip once somebody reports it. Conduit holds every dependency as a node with its lead time, supplier and risk, takes live signals from grid, vendors, transit and site, cascades a change downstream the moment it arrives, and proposes the recovery. Where it is authorised, it executes the re-route. Set what the build should optimise for and every routing decision and lead-time buffer re-weights to match.
| objective | energised | plan confidence | single points of failure |
|---|---|---|---|
| Speed to power | June 2029 | 80% | 3 |
| De-risk supply chain | January 2030 | 93% | 1 |
| Minimise cost | March 2030 | 74% | 4 |
Resilience costs seven months and removes two single points of failure. Cost discipline costs nine months and adds one. The three columns are the same eight orders under three weightings, and the model will run any of them in front of the people who have to choose.
Conduit runs in the browser. Set the objective, fire a disruption, and watch the orchestrator cascade it and re-plan the energisation date. Every date on this page was knowable the day the OEM moved the slot, two and a half years before anyone on site would have felt it.
Open the control room →