Even Bain’s optimistic estimate for existing AI applications leaves $4.2 trillion in annual revenue missing from the infrastructure boom’s economic story.
Not missing GPUs. Missing customers willing to pay.
That’s the awkward bit behind [The National’s $6 trillion headline](https://www.thenationalnews.com/future/technology/2026/09/29/ai-industry-needs-to-earn-6-trillion-by-2031-to-justify-data-centres/). We know how to order chips, arrange financing, and announce enormous campuses. Finding several trillion dollars of additional annual demand is a less cooperative engineering problem.
The spreadsheet behind the scary number
Bain’s September 2026 framework estimates annual AI-infrastructure investment could reach $1.5 trillion by 2031. That includes facilities and upgrades to GPUs, memory, and networking—not merely buildings with excellent air conditioning.
Assume capital expenditure equals 25% of revenue, and the arithmetic is:
<> $1.5 trillion ÷ 0.25 = $6 trillion in annual revenue./>
This is a conditional requirement, not a revenue forecast. Nor is it a complete investment-return model: revenue alone tells you little about electricity bills, margins, utilization, financing costs, or how quickly expensive hardware becomes yesterday’s hardware.
Change the assumed capex share to 30%, and the implied requirement falls to $5 trillion. Set it at 20%, and it rises to $7.5 trillion.
The headline has a denominator. Read it.
Bain estimates enterprise productivity applications could generate $1 trillion–$1.4 trillion, with consumer subscriptions and advertising adding $200 billion–$400 billion. Together, that leaves a $4.2 trillion–$4.8 trillion gap.
Robotics, autonomous vehicles, industrial automation, and AI-mediated advertising are proposed ways to fill it. Plausible categories? Sure. Contracted demand? No.
What Nobody Is Talking About
Useful AI and profitable AI infrastructure are different bets.
A company might save $100,000 using an AI workflow. That does not mean its model provider can charge $100,000. Competition, procurement, integration costs, and the customer’s perfectly reasonable desire to keep the savings get in the way.
Sequoia’s David Cahn made this distinction in [“AI’s $600B Question”](https://sequoiacap.com/article/ais-600b-question): cheaper compute can benefit founders even when infrastructure investors face disappointing returns.
That’s my central objection to the boom’s sales pitch. Demonstrating that AI creates value does not demonstrate that enough of that value flows back to whoever financed the GPUs.
There’s also an accounting trap worth watching: infrastructure, model services, and applications can all report revenue from the same underlying customer spending. An ecosystem-wide target needs consistent boundaries. That’s a caution—not evidence that Bain double-counted.
The financing gets creative before the products do
Meta’s Hyperion arrangement shows how widely the exposure can spread. Its October 2025 deal with Blue Owl involved approximately $27 billion in development costs, with Blue Owl-managed funds owning 80% and Meta 20%.
Debt investors, including PIMCO, were involved. Meta also committed to operating leases and a conditional 16-year residual-value guarantee.
Risk hasn’t evaporated. It has acquired paperwork.
One correction also matters: Epoch AI’s $200 billion in hardware and 9 GW extrapolation concerns a possible leading AI supercomputer in 2030. It is not a verified construction plan for Meta’s Prometheus facility, despite the news article’s attribution.
Ship outcomes, not token bonfires
For developers, the sensible response isn’t refusing AI. It’s refusing sloppy economics:
1. Measure cost per successful outcome, including retries and human review.
2. Route simpler work to smaller models; test caching and bounded agent workflows.
3. Keep providers replaceable. Today’s pricing is not a lifetime guarantee.
A support workflow resolving issues reliably is more convincing than an agent producing magnificent logs while accomplishing nothing.
My bet: AI will deliver substantial value, but some infrastructure investments will still disappoint. Those positions aren’t contradictory.
Build something customers pay for repeatedly. A GPU purchase order is not product-market fit.
