Bain: AI needs $6T a year by 2031. Guess who pays

Bain put a number on the bill: $6 trillion in annual AI revenue by 2031, or the data centers do not pay for themselves. Existing AI products — consumer subscriptions, ads, enterprise software — get the industry to somewhere between $1.2 and $1.8 trillion of that. The remaining $4.2 trillion does not exist yet. That gap is the single most useful fact for anyone signing an AI vendor contract this quarter.

What actually happened

Bain & Company published its 7th Global Technology Report on September 29. Annual AI infrastructure spending could reach $1.5 trillion by 2031 — new data centers and compute, plus the running cost of refreshing installed GPUs, memory and networking. To justify that, the industry needs $6 trillion a year coming back.

Bain's read on where the missing $4.2 trillion comes from: nascent categories — autonomous machines, robotics, drug discovery, mental health, energy generation. Bloomberg and the Japan Times both note the report's conclusion that infrastructure is being built well ahead of the demand curve.

The report also contains the number that explains every AI-branded product launch you have seen this year: hardware and semiconductor stocks compounded at 24% a year from 2020 to 2026. Software stocks managed 6%. And Bain's measured productivity gains from enterprise AI today land at 20–27% — real, but not $4.2 trillion real. David Crawford, chairman of Bain's technology practice, put it plainly: the economics demand trillions in new revenue beyond productivity gains.

Why the AI funding gap matters for your business

You are not going to close a $4.2 trillion gap. You are going to be billed toward it. That is what this report is for, if you run a small business: it is a forecast of your vendor invoices.

Three moves that pay off regardless of whether Bain is right:

Assume the per-seat AI line item goes up. Every SaaS tool you use added an AI tier in the last 24 months, most of them priced below cost to drive adoption. Infrastructure amortization does not care about your renewal date. Before you sign anything multi-year, get the price-increase cap in writing, and check whether the AI features are in the base tier or a bolt-on that can be repriced separately.

Know your token cost per transaction, not per month. A $200 monthly AI bill is invisible. A $0.04 cost per support ticket resolved is a number you can defend, renegotiate, or engineer down by switching models. Instrument it now, while switching is cheap. Our ROI calculator exists for exactly this arithmetic.

Build so the model is replaceable. Frontier model pricing has spread across more than two orders of magnitude, which means the same workload can get dramatically cheaper or dramatically more expensive on one config value. If your prompts, retrieval and test set live in your own repo, a repricing is an afternoon. If they live inside a vendor's workflow builder, it is a migration.

The honest version: Bain is not predicting a crash, and neither are we. Capacity built ahead of demand is how railroads, fiber and cloud all started. But every one of those buildouts was paid for in the end by the people using the service. Price your contracts like you are one of them, because you are.

Sources: Bain & Company via PR Newswire, Bloomberg, The Japan Times.