The $172B Question: What Generative AI Is Actually Worth

Most AI coverage treats the technology as a cost problem: training compute, inference pricing, API margins. That frame misses the real story. In April 2026, Stanford's Digital Economy Lab published a working paper that tried to measure generative AI the way economists measure welfare — not by what firms charge, but by what users would refuse to give up.

The answer changed the baseline.

The study

*What is Generative AI Worth?* (Brynjolfsson, Collis, Eggers, Kazinnik, Nguyen, April 2026) ran two waves of willingness-to-accept surveys on Prolific, sampling 1,491 U.S. adults in July 2025 and 1,908 in March 2026. Respondents were asked how much compensation they would need to forfeit one month of access to any generative AI tool for one month starting tomorrow.

The question is deceptively simple. It is a binary choice experiment with randomized price points — $1, $10, $20, $50, $100, $200, $500 — fit to a logit model. From the fitted demand curve the authors recover a median willingness-to-accept (WTA) and a mean, then multiply by the adult user base to get aggregate annual consumer surplus.

What makes the paper interesting is not the method. It is the gap between the result and the narrative the industry tells itself.

The numbers

| Metric | July 2025 | March 2026 | Change | |---|---|---|---| | Mean WTA per user | $98.00 | $124.50 | +27% | | Median WTA per user | $3.39 | $11.40–$11.48 | +235–238% | | U.S. adult users | 98.78M | 115.33M | +21% | | Aggregate annual consumer surplus | $116.2B | $172.3B | +48–50% |

The aggregate figure is the one that should reframe boardroom conversations. U.S. consumer surplus from generative AI reached $172 billion annually by early 2026. Leading generative AI firms — OpenAI, Anthropic, Google, Microsoft — captured roughly $14.2 billion in consumer revenue in 2025. That puts consumer surplus at roughly twelve times producer revenue.

Economists will recognize the ratio. It aligns with Nordhaus (2004), who found that innovators capture only about three percent of the total social returns from major twentieth-century technologies. The rest flows to users as welfare gain.

Mean and median tell different stories

The mean WTA is $124.50. The median is $11.40. That fifty-to-one spread is the paper's most important structural fact.

A small tail of heavy users — workplace users, paid subscribers, daily adopters — value generative AI in the high hundreds. They pull the mean upward. The median user is more representative: someone who uses the tools occasionally, often for free, who would miss them but whose compensation demand is modest.

For product teams and policy makers, the median is the more honest central tendency. It says the average American values gen AI at the cost of a decent lunch, while the power user values it like a productivity infrastructure. Both are true at the same time.

Adoption outran investment

Generative AI adoption in the U.S. hit 28.3% by 2025 — 24th in the world. The country leads in model development and private investment (23× China's private AI spending in the same period) but lags in actual use. China and Europe posted the highest year-over-year increases in organizational adoption.

That gap is the commercial condition. The technology is already generating substantial welfare. Its producers are not capturing most of it. That means the value is still escaping into user surplus rather than into revenue lines. The companies that figure out how to hold more of it without breaking the free-or-near-free consumer contract will own the next decade.

The labor signal

The paper documents a counterintuitive labor pattern. Employment for software developers aged 22 to 25 fell nearly twenty percent from 2024 to 2025. One third of employers expect further workforce reductions in the coming year. Yet almost half expect little to no change.

Anticipated cuts concentrate in service operations, supply chain, and software engineering — the structured, measurable tasks where AI shows the largest productivity gains. Customer support climbs 14–15%, software development 26%, marketing output 50%. Gains are smaller for tasks requiring deeper reasoning.

Productivity is up. Entry-level employment is down. That is the regime the economy is navigating.

Why this matters for Edgeless

The $172B figure is a floor, not a ceiling. The surveys were fielded when gen AI tools were still novel and often free. Valuation should rise as usage deepens, tools become more capable, and workflows shift from experimentation to infrastructure.

The 12× surplus-to-revenue ratio is the argument for building on top of existing models rather than training new ones. The welfare is already being created. The question is who captures it.

That is the operating assumption for every agent product we ship.

What the reel got wrong

A popular Instagram reel citing this paper claimed a "$150 median consumer surplus per user" and a "97% get value" stat. Neither appears in the paper. Median WTA is $11.40, not $150. The $150B figure belongs to Google's annual capex in the broader AI Index report. The aggregate consumer surplus is $172B, but that is total, not per-user.

The core insight survived the misquote. Generative AI is already worth a lot to a lot of people, and the market has not come close to capturing it.