Last week we covered Bain's number: AI needs $6 trillion a year in revenue by 2031 to pay for its own buildout, with at least $4.2 trillion of that coming from product categories that don't exist yet. a16z just published the consumer-side evidence for why that gap is real — and the number it found cuts two ways.
As of April 2026, only 2.2% of US households pay for any AI service, according to a16z's analysis of PNC transaction data. That's up from near zero in January 2023, but it still means 98% of American households pay nothing. Among the minority who do pay, average monthly spend rose from $22 to $31 over two years — so the paying base is spending more, but the base itself is barely growing. On the enterprise side, the same report found 69% of S&P 500 companies have a live AI deployment, but only 2% disclose a metric they actually track over time. Adoption and measurement are different things, and almost nobody's doing the second one.
Read one: the spend is futuristic, not current
The uncomfortable version of this story is the one Bain's $4.2 trillion gap already told: hundreds of billions in infrastructure capex is being justified by product categories — autonomous systems, physical AI, agentic software, ad-supported chat — that mostly don't generate revenue yet. The 2.2% household number is the receipt for that. Hyperscalers aren't scaling GPU fleets to meet today's $31-a-month subscriber base. They're building for a demand curve that assumes today's usage is a rounding error against where this goes. That's a bet on the future, priced and spent today, against a present that doesn't yet support it.
Read two: it's extremely early
The other way to read the same number is less alarming. Paid penetration for most consumer technology categories looked like this once. Early broadband, early streaming, early smartphone data plans all had years where usage was real but paid conversion was a sliver of the population, before the curve bent. a16z's own chart shows exactly that shape: near zero to 2.2% in a little over three years, a compounding curve, not a flat one. Enterprise is earlier still — 69% deployment with only 2% measurement isn't evidence AI doesn't work in the enterprise, it's evidence most companies haven't finished the unglamorous work of instrumenting what they've already rolled out. That's a maturity gap, not a demand failure.
Both readings are true at once
That's what makes this number worth sitting with instead of resolving into a single verdict. The infrastructure spend genuinely is priced for a future that hasn't arrived — that part of Bain's gap is real, and 98% of households not paying is the concrete evidence of it. At the same time, three years in is not when you'd expect a genuinely transformative technology to already have mature monetization, and the growth curve underneath the 2.2% is steep, not stalled. The honest conclusion isn't bubble or not-bubble. It's that the infrastructure is being built on a timeline that's running well ahead of the revenue curve underneath it — and whether that gap closes in time is still an open question, not a settled one in either direction.