A $6 trillion question hangs over the AI construction boom

The artificial intelligence construction boom has more money to spend and more problems to solve.

Annual AI infrastructure spending could reach about $1.5 trillion by 2031, according to Bain, a Boston-based global management consulting firm. Through 2030, the firm projects a cumulative $5 trillion to $6.5 trillion in data center spending. To justify those outlays, the AI market would need to approach $6 trillion in revenue annually by 2031, according to the company.

The eye-popping price tag is “conceivable,” considering how quickly the world’s largest technology companies have ramped up data center spending, said Peter Hanbury, a partner at the consulting firm. However, he questions whether these AI companies will eventually produce enough revenue to keep construction going at that pace.

For contractors, immediate headwinds still revolve around power constraints and electrical labor, said Hanbury. Lead times on key materials remain long enough that decisions made today determine capacity several years from now, according to Bain.

Increased community opposition is also impacting timelines. Local opposition blocked or delayed at least 75 projects worth $130 billion in the first quarter of 2026, nearly matching the impacts in all of 2025, according to the firm.

Here, Hanbury talks with Construction Dive about the duration of the AI investment wave, construction constraints and viability of the buildout.

Editor’s note: This interview has been edited for brevity and clarity.

PETER HANBURY: I think it’s realistic based on the spending trajectory we are already seeing. The largest hyperscalers, Microsoft, Google, Amazon, Meta and Oracle, could spend roughly $780 billion in capital expenditures in 2026 alone, nearly five times what they were spending three years earlier.

Not all of that is data center capital expenditure, but it shows the scale of capital already being deployed. Against that backdrop, a cumulative $5 trillion to $6.5 trillion of data center investment through 2030 is certainly conceivable.

The bigger question is whether the economics can sustain that level of investment over time. Our AI work estimates that annual AI infrastructure spending could reach about $1.5 trillion by 2031. If capital expenditure represents roughly 25% of industry revenue, supporting that spending would require an AI market approaching $6 trillion in annual revenue.

So the near-term issue is not necessarily whether enough capital exists.

A few things have to happen at the same time.

First, AI has to move beyond use cases focused on efficiency and productivity and start creating entirely new sources of revenue and economic value. We would need substantial growth from new sources of value such as autonomous systems, physical AI, new consumer experiences and entirely new AI-enabled products and industries.

Second, the physical bottlenecks have to ease. Capital doesn’t help if a project can’t get power, chips, skilled labor or permission to build. Power, in particular, needs system-level solutions, more generation, faster interconnection, behind-the-meter capacity, storage and greater coordination among utilities and technology companies.

Third, the capital model will have to become more creative. Hyperscaler balance sheets can carry an enormous amount, but at this scale we are likely to see more risk-sharing across developments, infrastructure investors, utilities, sovereigns and governments, as well as approaches that include jointly funded power infrastructure.

And finally, the industry will have to become much more disciplined about which projects get built. Not every announced gigawatt will make sense.

If those things happen, the investment level is achievable.

It will require the industry to move from managing individual projects to running industrial-scale programs. That means integrated power and site planning, earlier procurement and more prefabrication, modular designs and coordinated portfolios of contractors and suppliers.

At this scale, the slowest constrained input sets the schedule for the entire program.

Electrical labor is likely to be the sharpest constraint, particularly high-voltage, substation and mission-critical electrical talent. Mechanical and pipefitting trades will also come under pressure as liquid cooling scales, along with smaller but highly specialized pools of controls and commissioning talent.

The issue becomes more acute as campuses scale because these projects need large numbers of specialized workers at the same time. That will push more work toward offsite fabrication, prefab assemblies, regional labor strategies and longer-term partnerships with key trades.

Contractors should increasingly underwrite the pipeline.

I would ask four questions: Is the power real? Is the customer and financing commitment real? Is the permission to build real? And is the design stable enough to build?

Permission increasingly means more than a permit. It includes community support and social license around power use, water, noise, emissions and other local impacts.

Power is becoming the gating item. A data center can be built in a few years, but adding major new grid capacity can take four years or more. At the largest campuses, that means power infrastructure is now firmly on the data center projects’ critical path.

Contractors are increasingly working around substations, transmission, interconnection and, in some cases, onsite generation and storage as part of the same delivery program. A gigawatt-scale data center looks more like a power project with a very large computing load.

One of the biggest changes is chip-to-grid codesign. Historically, the facility and IT stack could be designed with a fair amount of separation. The building did not need to change dramatically with every new chip generation. AI is collapsing those boundaries.

Changes in graphic processing units and custom silicon drive rack density. Rack density drives networking and cooling. Cooling and compute density change electrical architecture. And all of those choices affect the building and ultimately the power source.

For contractors, the implication is the server roadmap is increasingly helping design the building. And that silicon roadmap is becoming more varied and faster-moving as Nvidia continues to accelerate its platform cadence while hyperscalers simultaneously push custom silicon.

The players that can coordinate compute, power, cooling and construction will have an advantage over those optimizing each layer independently.