AI needs to find $4.2 trillion in new markets by 2031. Bain says it could come from four places

By 2031, the cloud companies underwriting the AI data-center build will spend $1.5 trillion a year on chips, servers, and power. A typical 25% capital-expenditure-to-revenue ratio, where capital expenditure is the share of revenue a business plows back into physical infrastructure, means that level of spending requires $6 trillion in annual AI revenue. The markets AI can plausibly serve today, consumer and enterprise software combined, top out at about $1.8 trillion. The arithmetic leaves a $4.2 trillion gap.

That is the number at the center of Bain & Company's 2026 Technology Report. The consultancy is not arguing the buildout is irrational. It is arguing that the buildout has to be paid for, and the bill is too large for the existing customer base. Bain's own finding is that current productivity gains from AI services do not justify the capital flowing into the sector. The math has to close somewhere, and Bain's conclusion is that somewhere is brand-new product categories.

Those categories, by Bain's count, are four: AI-native search; autonomous vehicles and drones; "physical AI," which includes robotics and digital twins of factories, warehouses, and supply chains; and AI-driven product development, with pharmaceutical R&D as the leading example. Each is a construction site. None of them is a market today in the sense that consumer software or cloud services are. Each has to be invented at a scale the technology industry has rarely attempted.

The math

Bain estimates hyperscaler capital expenditures, the cash the largest cloud companies spend on data centers, GPUs, and the power to run them, could reach $780 billion in 2026, a fivefold increase in three years. Annual spending on AI infrastructure could climb to $1.5 trillion by 2031, according to a NetworkWorld summary of the report. The 25% capex-to-revenue benchmark is a stated assumption, not an externally audited figure, and it reflects how capital-intensive network operators, telecoms, and large industrial firms have historically run their balance sheets. AI hyperscalers have not yet operated at this scale long enough for the ratio to be tested.

The comparison point is the cloud business itself. Hyperscaler capex in 2023 was roughly $150 billion, against combined cloud-segment revenue of more than $200 billion, a capex ratio closer to 60% during a heavy build phase, which has since fallen as utilization improved. The 25% figure assumes a mature steady state, not a build year.

The four sites

AI-native search. Today's search market is roughly $200 billion in advertising revenue, dominated by Google. An AI-native replacement would not just slot into that market. It would have to displace it, or open a parallel market for agents that browse, transact, and execute tasks on a user's behalf. The construction question is whether agents become a new commerce surface that advertisers pay to reach, or whether they compress the existing market by skipping the click.

Autonomous vehicles and drones. The global automotive industry sells roughly $3 trillion in vehicles and parts a year, and the drone and robotics industries are growing fast off a small base. Bain's category is broad enough to include robotaxi fleets, autonomous trucking, last-mile delivery drones, and industrial drones for inspection and agriculture. The construction question is regulatory clearance, public acceptance, and unit economics, none of which are settled.

Physical AI. This is the catch-all for AI that operates in the physical world: factory robots that learn from simulation, digital twins of buildings and supply chains that let companies model changes before committing capital, warehouse automation that adapts to new SKUs without reprogramming. Bain's analysts treat this as the most fragmented of the four categories, with the longest path from demonstration to recurring revenue.

AI-driven product development. The clearest case is pharmaceuticals, where foundation models are being used to design molecules, predict protein structures, and run virtual screens before wet-lab work. The construction question is whether the time and cost savings translate into drugs that reach clinical trials, get approved, and generate revenue at a scale that the AI infrastructure can be billed against.

What has to be true

For the math to close, one of two things has to happen. Either the four candidate markets scale fast enough that, by 2031, they collectively generate $4.2 trillion in annual revenue, an average of more than $1 trillion per category, or hyperscaler capex slows to match the markets that actually exist. Bain's framing is closer to the first scenario: the report's title argues that "new innovation is required," not that the spend should be cut.

Neither outcome is foreordained. The category-by-category analysis is a future-modeling exercise, and Bain is explicit that the four markets are candidates, not forecasts. The consultancy has not published a probability for any one of them reaching trillion-dollar scale. What it has published is a constraint: at the spending levels the cloud giants are already committing to, the existing AI market is not large enough. The next five years of AI will be defined less by which model wins the benchmark and more by whether the industries being built around it can grow into the bill.