AI's Next Bottleneck Is Construction
AI infrastructure is being sold as a chips-and-software story, but the harder constraint may be permits, power, switchgear, labour and project controls.
Executive summary
AI infrastructure is being sold as a chips-and-software story. The harder constraint may be permits, power, switchgear, labour and project controls.
~40%
US data centre projects due in 2026 that may miss planned dates
57%
JLL-reported projects delayed three months or more in 2025
42 weeks
Average 2026 US data centre equipment lead time reported by JLL
1 GW
Roughly 1000 MW of electrical capacity
Save briefing
Checking your account save status...
The AI boom is not only testing semiconductor supply. It is testing the construction industry's ability to deliver complex, power-heavy capital projects without pretending constraints are someone else's problem.
Opening note
Morning builders,
The AI industry is racing to build digital capacity. The limiting factor may be concrete, steel, power connections, permits and specialist trades.
That is the contradiction behind the current data centre boom. The public story is about models, chips and cloud capacity. The delivery story is about substations, transformers, planning approvals, grid queues, commissioning dates and whether the project schedule is credible.
For construction, this is not a side issue. It is a live demonstration of what happens when capital demand outruns physical delivery capacity.
Best number of the week
Almost 40%.
Recent reporting based on satellite analysis has suggested that close to 40% of US data centre projects scheduled for completion in 2026 may miss their planned dates. Other estimates put the 2026 delay or cancellation risk across US data centre capacity in a wider 30% to 50% range.
Treat those numbers with care. They are estimates from external trackers, not a final claims register. But the direction is hard to ignore: AI infrastructure is running into old capital project constraints.
The bottleneck is no longer only compute
What happened
Demand for AI compute has pulled forward a huge wave of data centre development. Some campuses are now discussed in hundreds of megawatts, and the largest AI infrastructure programmes are moving into gigawatt-scale territory.
That changes the construction problem. A 50 MW facility is already a serious delivery challenge. A campus with hundreds of megawatts, or a power requirement approaching 1 GW, behaves less like a building programme and more like an industrial infrastructure programme with a data centre attached.
Why it matters
If these projects slip, more than a building completion date moves. Cloud capacity, AI product rollouts, lease revenues, investor assumptions, utility plans and regional power politics all move with it.
The AI boom has made construction schedule risk financially visible to people who normally talk about GPUs and software margins. That is good. It may force a more serious conversation about delivery reality.
By the numbers
- Sightline Climate has estimated that 30% to 50% of data centre projects due in 2026 could face delay or cancellation risk.
- JLL reported that 57% of data centre projects experienced a construction delay of three months or more in 2025.
- JLL also reported average 2026 US data centre equipment lead times of 42 weeks, materially above 2019 levels.
- One gigawatt is not a metaphor. It is roughly 1,000 MW of electrical capacity, and Axios notes that 1 GW can power about 1 million US homes.
Hardhat take
The AI boom is not only testing semiconductor supply. It is testing the construction industry's ability to deliver complex, power-heavy capital projects without pretending constraints are someone else's problem.
Where the construction bottlenecks sit
What happened
The reported causes are familiar to anyone who has sat through a difficult project controls review: permitting friction, power availability, grid connection risk, specialist labour shortages, long-lead equipment and weak visibility of actual progress.
Data centres concentrate all of those risks. They need land and approvals. They need utility interfaces. They need electrical and mechanical systems that are already in short supply. They need contractors and commissioning teams with specialist experience. They also need owners, designers, utilities, suppliers and contractors to make decisions on the same clock.
Why it matters
The risk is not one big heroic delay event. It is the interface stack.
A permit slips, so enabling works move. A substation date moves, so commissioning floats disappear. Switchgear arrives late, so rooms are built but cannot be energised. A utility decision waits for design maturity, while procurement needs a frozen load profile. The monthly report says the project is amber until it is suddenly red.
That is not an AI problem. It is a controls problem.
Question to ask on your project
Which item would stop beneficial use even if the building looked complete: permit, grid connection, transformer, switchgear, cooling, controls integration or commissioning resource?
The project controls lesson
The lazy answer is "build faster". The useful answer is "make constraints visible early enough that leadership can act".
These projects need credible schedules, not optimistic milestone posters. They need constraint registers that treat power, permitting and long-lead equipment as first-class drivers. They need field reporting that confirms what happened at the workface, not what a dashboard hoped would happen. They need commercial exposure tracked in dollars, not only days.
Most importantly, they need tighter owner, EPCM, contractor, utility and supplier decision loops. The speed of decision-making is now part of the critical path.
What good looks like
Try this today
- Map power, permitting and equipment as first-class schedule drivers, not notes under "external dependencies".
- Treat grid connection as a critical path risk until the energisation path is proven.
- Use field reporting to validate actual progress against planned progress every week.
- Run schedule diagnostics for weak logic, open ends, excessive float, missing progress and unrealistic durations.
- Track delay exposure in dollars as well as days, with the cost basis clearly stated.
- Connect design, procurement, construction and commissioning data before handover pressure arrives.
- Escalate blockers when they first become visible, not after the monthly report has made them politically safe.
The sceptic's corner
AI will not magically fix poor project controls. If the schedule has broken logic, the field record is late, procurement dates are optimistic and the utility interface is managed in email, a clever model will mostly produce a better-written version of the same uncertainty.
The opportunity is more practical. AI can help read messy records, surface repeated blockers, compare progress signals and explain risk earlier. It still needs a delivery system that owns the response.
The whole-project intelligence shift
The construction technology trend is moving away from isolated AI tools and toward whole-project intelligence. The useful future is not one chatbot per department. It is schedule, progress, field reporting, risk, cost and commercial exposure connected into a single delivery picture.
That is the reason our Toolbox is being built around practical controls workflows rather than generic AI demos.
The Schedule Intelligence Analyzer is aimed at schedule quality, weak logic and risk exposure. The Project Controls Diagnostic looks at reporting maturity and governance. Field Report and Shift Capture tools turn live site signals into a cleaner evidence trail.
None of that replaces planners, project managers or superintendents. It gives them a better early-warning system.
Closing note
AI demand is exposing weaknesses construction already knew about: brittle schedules, hidden constraints, fragile supply chains, late reporting and decision loops that move slower than the project requires.
The winners will not be the teams with the glossiest dashboard. They will be the teams that can see the constraint, quantify the exposure and move the decision while there is still time to protect the project.
The AI boom will not be won only in model labs. It will also be won or lost in permits, substations, switchgear, workfaces and weekly schedule reviews.