Every few years construction gets a building type it has never priced before. Right now it is the data center, and the buildout around it is the biggest construction program of the decade. Most of the public conversation is about chips and power purchase agreements. Almost none of it is about the part that actually decides whether the capacity arrives: getting a site found, designed, documented, and built on schedule.
A construction problem, not a compute problem
The binding constraint on AI infrastructure is not silicon. It is delivery. Can a qualified site be found, can it be designed, can the set be produced, and can it be built before the demand it was meant to serve has moved on. That is a feasibility, drawing set, and coordination problem, and it is exactly the problem the estimating and preconstruction world was never re-tooled for. We kept the workflow that priced office parks and pointed it at a campus.
The mismatch is not cosmetic. Three things about a data center break the assumptions underneath a normal preconstruction cycle, and each one breaks it in a different place.
The workload is a module, tiled
A campus is not one hard building. It is one module, repeated. A data hall, an electrical gallery, an admin block, the same pad again and again, phased across a site. That sounds easier than a bespoke tower, and in the field it often is. In preconstruction it is a trap. When the unit of work repeats, so does the unit of error. Price the module a little wrong once and the mistake does not stay small. It tiles. A per-module miss that would be a rounding error on a single building becomes the gap between winning and losing the program once it is multiplied across every pad and every phase.
So the honest way to estimate a campus is not to estimate the campus. It is to measure one module completely, prove that measurement, and then let the repetition work for you instead of against you. That only holds if the first module is measured rather than guessed, and if each repeat is a real derivation instead of a copied number that nobody can trace back to anything.
The schedule is hostile
The team that answers in days loses to the team that answers in hours.
Power windows and interconnection timelines do not wait for a drawing set. Megawatts, utility capacity, and delivery phases move faster than a manual preconstruction cycle can turn, and every week spent in drawing production is a week of capacity that nobody is serving yet. The market rewards speed on exactly the questions that traditionally took the longest to answer: is this site real, what will it cost, and when can it deliver.
That is why feasibility cannot stay the slow first step it usually is. The value of answering the power and site questions early is not tidiness. It is that dead sites die cheap, before land money and design money are committed to a parcel that was never going to interconnect in time.
The owner wants receipts
Hyperscale owners audit everything, and they are right to. When one number moves a quarter of a schedule or a large piece of a site, the owner is going to ask where it came from, and the estimator's judgment is not an answer that survives that room. Every number needs provenance: measured, derived, or assumed, and honest about which. A campus estimate that cannot show its work is not conservative. It is unfalsifiable.
This is the through-line of everything that follows. Not a prettier number, a traceable one. Coverage you can see, confidence you can read line by line, and a path from any figure back to the drawing or the measurement it rests on. And underneath all of it, one project record, so the same data is never re-keyed as work moves from feasibility to design to bid to build. Multi-site programs live or die on that consistency.
What this series does
The rest of the series follows one data center campus through the platform, in the order the work actually happens, and it is honest at every step about what ships today and what is still in preview.
Part two takes feasibility: power availability, interconnection, water, and clearances gathered into one answer before commitment, the design phase that is in preview now. Part three goes from a generative site scheme on the real parcel to a drawing set where every dimension knows where it came from, also in preview. Part four is takeoff at campus scale and part five is the estimate an owner can audit, and those are the parts grounded in what already ships: measured takeoffs and priced estimates running today for the shell, site, and fit-out scopes. Part six looks at bid to build, and part seven at one graph from the first pad to the last handover, the direction the platform is being built toward.
That is the whole idea. One graph underneath the work, measured where it can be measured and labeled where it cannot, from the first site question to the last handover. The buildout is the reason it finally has to be true.