For NVIDIA GTC Berlin, 20 to 22 October 2026

The GPUs will arrive on time. Will the building be ready for them?

$811 billion is going into data centers in 2026. The chips are ordered, the land is secured, the power contracts are being fought over. Analysis of 750,000 past projects says most of these builds will still be late, and that the cause is not chips, power or permits. It is the decisions made during delivery. nPlan's executive briefing sets out what the data shows, why it happens, and what owners can do about it.

Data center pipeline

$811bn

being spent on data centers in 2026

Early setbacks

88%

of projects hit setbacks before the foundations are laid

Late by a quarter or more

60%

of projects slip by more than three months

Cost of a three-month slip

$27m to $65m

revenue deferred on a 100MW build

Get the executive briefing

Enter your email and we will send the briefing. At GTC? Find us at STATION-Berlin on 21 and 22 October.

nPlan Executive Briefing: The Data Center Delivery Problem (PDF)
What a slip costs, where delay really starts, and two builds where it was called early

We will only use your address to send the briefing and, if you opt in, occasional nPlan updates.

Why it matters at GTC

Compute is perishable. Buildings are late.

Everyone at GTC is racing to get GPUs into racks. The chips ship on a date the vendor controls. The hall, the power and the cooling finish on a date nobody controls, and that is the date revenue starts. Every week of delay is stranded capital, depreciating chips and deferred revenue. The briefing puts numbers on that, and shows where the time actually goes.

What a three-month slip actually costs

The maths on a 100MW build: $27 million to $65 million of revenue deferred, before liquidated damages or a broken SLA. That is the number to hold next to any promise about a ready-for-service date.

Why the delay starts before the foundations

88% of data center projects hit setbacks before concrete is poured. The briefing shows where delay originates on these builds, and why it is not the chips, the power or the permits you should be watching.

How bias gets baked into the plan

Optimism, salience and availability bias quietly stretch durations and shrink risk registers. A forecast built from 750,000 past projects does not share those biases, which is why it disagrees with the plan early.

Two builds where the slip was called in advance

A Virginia hyperscaler build was forecast to within days of its outcome. A $40 billion LNG facility used the same approach to pull start-up forward by two months. Both are in the briefing.

nPlan for data centres

Meet us at GTC

What the briefing shows

Four numbers from the briefing

$1.3bn

what a $1bn facility routinely becomes

The overrun is the norm, not the exception

A $1 billion facility routinely becomes a $1.3 billion one, and 60% of data center projects slip by more than three months. Every week of that is stranded capital, depreciating chips and deferred revenue.

88%

hit setbacks before the foundations are laid

The problem is execution, not supply

Most projects are behind before the first pour. The bottleneck is not semiconductors and it is not permitting. It is the decisions made during delivery, at the moments that decide whether a milestone holds or slips.

$27m to $65m

deferred by a three-month slip on a 100MW build

A quarter of delay has a price

Three months late on a 100MW build defers $27 million to $65 million of revenue before liquidated damages or a broken SLA. The briefing shows the working so you can run it for your own capacity.

2 months

pulled forward on a $40bn LNG start-up

The slip can be called in advance

A Virginia hyperscaler build was forecast to within days of its outcome. A $40 billion LNG facility used the same forecast to bring start-up forward by two months. The briefing sets out what those owners did differently.

Your campus

Run this on the campus you are building

With your own plan, in Primavera P6 or Microsoft Project, the forecast updates every time you report progress. You see the most likely date, the reasons it is slipping, and which interventions buy back the most time. nPlan has worked with hyperscalers since 2020. If you are at GTC Berlin, find us at STATION-Berlin on 21 and 22 October. Otherwise, book a call.

Talk to us

Want to see the method on a real campus first? Our Chapelcross study rebuilds a proposed 1GW AI campus from public filings alone and forecasts it: most likely finish 44 days behind plan, 86 days lost before the first building starts, and seven decisions ranked by when their window closes. Read the Chapelcross study.

The Chapelcross study was produced without any involvement from, data from, or endorsement by CX Tech Limited or its partners. Every input is public and the dates are nPlan's own estimates. NVIDIA and GTC are trademarks of NVIDIA Corporation. nPlan attends GTC Berlin as a visitor and is not affiliated with or endorsed by NVIDIA.