AI's $6 Trillion Question: Who's Paying?

Bain & Company’s Technology Report 2026, published September 29, 2026, projects approximately $5 trillion to $6.5 trillion in data center spending through 2030. In Bain’s base case, that adds nearly 150 gigawatts (GW) of new compute capacity and almost triples global capacity in five years.

Two reports in the release frame the challenge from opposite ends. One finds that power, chips, skilled labor, and permitting are all constrained at once. The other asks whether AI will earn enough money to pay for the buildout.

Five years ago, a facility of 50 megawatts (MW) was considered large. Hyperscalers are now planning campuses of 5 GW or more, at $150 billion to $200 billion apiece.

Bain says US data center investment, measured as a share of GDP, now rivals what the country spent building its railroads, telecom networks, and electrical grid. Most of the projected growth will land in the US.

The checkbooks are open. Bain says capex at Microsoft, Google, Amazon, Meta, and Oracle could reach $780 billion in 2026, nearly five times the level of three years earlier.

Power is the first wall. Adding grid capacity takes four years or more, and critical equipment like transformers is in short supply.

Generating power onsite, behind the meter, is one of the few ways to add large blocks of power on data center timelines. Today that mostly means natural gas, which brings regulatory exposure, longer equipment lead times, and more project complexity.

Tech companies are also turning to renewables and storage. Google is funding 1.6 GW of renewables, mostly wind, plus 300 MW of storage alongside a new data center in Minnesota.

Nuclear is on the table, but new nuclear capacity at large scale isn’t expected until the middle of the 2030s. Small modular reactors are drawing bets too, with Google backing Kairos Power and Amazon backing X-energy.

Bain calls SMRs a credible piece of the puzzle. They also say SMRs likely can’t move fast enough or scale on their own.

Chips are spoken for well before the building is. GPUs, memory, and optical networking are being locked in years ahead.

Bain says operators without supply agreements that run for years are effectively out of the queue.

Bain describes skilled labor as the human version of the same bottleneck. There aren’t enough electrical workers, mechanical tradespeople, or cooling specialists to go around.

Closing that gap means recruiting, training, and keeping trades at rates far above historical levels.

Bain says that even on a site with existing power, a 1 GW data center typically takes one to four years to build, drawing on Epoch AI data. When new generation and transmission are needed, it can take longer.

Some projects are finding shortcuts. Colossus, from xAI (now part of SpaceXAI), went up in under a year after the team repurposed a factory and leased dozens of gas turbines without federal clean air permits.

That approach drew a lawsuit. Bain’s broader point is that workarounds like this can’t be repeated at scale.

Permitting and public opinion are now a schedule risk of their own. Local opposition blocked or delayed at least 75 projects worth $130 billion in the first quarter of 2026.

That nearly matches the $156 billion in projects disrupted across all of 2025. Residents and politicians are pushing back on water use, energy prices, and noise.

In July 2026, New York became the first US state to pause construction of the largest data centers, according to Bain. Fourteen other states have introduced similar legislation.

Overseas, Ireland, the Netherlands, and Denmark have restricted or paused new grid connections, at least for now.

Some communities are writing their conditions into contracts. Lancaster, Pennsylvania signed one with Chirisa Technology Parks and Machine Investment Group, tied to an agreement to codevelop a data center for CoreWeave.

The terms cap water use. They also set requirements for clean energy, noise, emissions, and local hiring.

Bain cites Epoch AI estimates showing the largest AI data center at any given time has roughly doubled in size and cost every 12 to 16 months. The largest facilities operating today are approaching 1 GW.

If that doubling trend holds, many facilities will approach 2 GW by 2027. The same trend points to 9 GW campuses by the end of the decade.

Hyperscalers’ data center capex alone could top $1 trillion by 2028. Total annual AI infrastructure spending, including upgrades to installed GPUs, memory, and networking, could reach $1.5 trillion by 2031.

Bain’s second report asks whether the revenue will show up to cover all of this. Assuming capex runs about 25% of industry revenue, sustaining that spend would require an AI market of roughly $6 trillion a year by 2031.

Consumer AI could bring in $200 billion to $400 billion through subscriptions and advertising. Enterprise adoption could add $1 trillion to $1.4 trillion for AI providers.

Together that’s $1.2 trillion to $1.8 trillion. Bain puts the remaining gap at about $4.2 trillion in annual revenue by 2031, which has to come from new sources of economic value.

Bain names four places that money could come from. Search and advertising could add $100 billion to $200 billion or more as AI replaces a large share of traditional search.

Autonomous cars, trucks, drones, and industrial automation make up a $400 billion opportunity. Physical AI, covering simulations, digital twins, and robotics, could be worth about $900 billion, assuming a 10% cut in R&D and manufacturing costs.

The fourth category is new products, such as drug discovery powered by AI and breakthroughs in materials science. Bain doesn’t put a dollar figure on that one.

Bain’s verdict is that enterprise productivity gains won’t be nearly enough on their own. Funding the buildout sustainably would require adding about 1% to the annual global GDP growth rate.

That’s a tall order, because Bain says the infrastructure is being built ahead of the demand curve. Their open question is whether the applications arrive in time to pay for it.

If you pull wire, set gear, or run pipe, Bain’s numbers point straight at you. They name electrical workers, mechanical tradespeople, and cooling specialists as the trades running short. Expect power to drive sequencing, since grid capacity takes four years or more and transformers are scarce. Onsite generation adds its own equipment lead times and complexity. Community agreements like Lancaster’s add water, noise, and local hiring requirements that shape how the site runs.

Bain’s base case calls for nearly 150 GW of new capacity by 2030. Hyperscalers’ data center capex could top $1 trillion by 2028. If the doubling trend Bain cites from Epoch AI holds, many facilities will approach 2 GW by 2027, with 9 GW campuses by the end of the decade. Bain’s open question is whether new AI applications arrive in time to pay for it all.

On these jobs, power sets the schedule. When grid capacity takes four years or more and transformers are scarce, gear with long lead times often gets ordered before design is finished. Subs and suppliers who can commit equipment and crews early often have the edge.

Both reports point to risk beyond the jobsite. Bain says the infrastructure is being built ahead of demand. Local opposition also blocked or delayed $130 billion in a single quarter. Weigh announced projects by where they stand on permits, and check who’s actually funding the job before you commit crews.

🥾Want to advertise in TheJobWalk? Learn More🥾