What AI is actually worth
Last summer, researchers from Stanford made a striking claim in a Wall Street Journal op-ed: Americans captured roughly $97 billion of value from generative AI tools like ChatGPT in 2024, while the companies building those tools captured only about $7 billion in US revenue. A 14-to-1 gift from producers to consumers. By April 2026, the Stanford team had updated the number to $172 billion, and the argument that AI is a historic gift to consumers had become a staple of the AI investment debate.
The number is probably overstated by five to ten times. Once you understand how it was measured, and compare it to what people actually do when offered AI for money, the picture changes. The real value is much smaller than the headlines suggest, and AI companies have probably already captured more revenue than consumers are getting in free value.
This post walks through how the $172 billion was built, the four reasons to discount it, where revenue actually stands, and what that changes about the AI investment story.
How the $172 billion was measured
The research, led by Erik Brynjolfsson at Stanford’s Digital Economy Lab with Avinash Collis at Carnegie Mellon and Felix Eggers at Copenhagen Business School, uses a survey method called “willingness to accept.” The plain-English version: they ask a representative sample of US adults how much money they would need to be paid to give up AI chatbots for one month. They average the answers, multiply by 115 million users and 12 months, and arrive at $172 billion.
The average answer in their March 2026 survey was $124.50 per month. That implies the typical American values generative AI at roughly $1,500 per year. The middle person in the sample (the median) answered $11.40 per month, or $137 per year.
The gap between those two numbers is the first signal that something is off. Use the median instead of the average and the aggregate becomes $16 billion, not $172 billion. A small number of very enthusiastic users is pulling the average up by a factor of ten.
The research is not fringe. The same methodology has been used on Facebook, Google Search, and email, and the underlying 2019 PNAS paper was peer-reviewed. The AI-specific paper is a working paper. The Stanford team is serious. This is the most-cited estimate of what AI is worth to American consumers.
The problem a careful reader spots in 30 seconds
ChatGPT has about 900 million weekly users as of February 2026. Roughly 50 million of them pay for a subscription. That is a conversion rate of 5.6 percent. The cheapest paid tier costs $20 per month.
If the typical user really valued ChatGPT at $124.50 per month, why are 94 percent of weekly users not paying $20 a month for it? Either the survey number is wrong, or OpenAI is leaving enormous money on the table with its pricing. Both cannot be true.
This is the revealed preference problem. What people actually do usually tells you more than what they say in a survey. When 94 percent of regular users pass on a $20 offer for something they say they would need $124 to give up, the behavior and the survey are incompatible. One of them is mismeasured.
Four reasons the survey number is too high
Economists have understood for decades that asking people “how much would you need to be paid to give up X” produces bigger answers than asking “how much would you pay for X.” The gap can be two to ten times. Daniel Kahneman, the Nobel economist, called this loss aversion: we hate losing things more than we like getting the equivalent gain.
When researchers average across many goods, the ratio between these two questions runs about 7 to 1, based on a 2002 meta-analysis by Horowitz and McConnell. Apply that correction to Brynjolfsson’s $124.50 and the more honest valuation drops to roughly $17 per month, and the aggregate drops to about $24 billion. That single correction takes the story from $172 billion down to something ten times smaller.
Three other problems stack on top.
Substitutes are free. The survey asks what it would cost to give up AI chatbots entirely. But if ChatGPT disappeared tomorrow, people would use Google Search, Wikipedia, Stack Overflow, colleagues, and rival AI models. The relevant question is not “what would it cost to live without AI,” it is “what would it cost to live without this particular AI when the rest of the internet still exists.” The paper’s own data show users keep an average of 2.6 AI products at once, which is direct evidence that AI is substitutable. Correcting for substitutes probably cuts the estimate by another half.
Work use inflates the number. About half of respondents say they use AI for work. Their employers sometimes pay for it. That value belongs in business accounting, not in consumer welfare. The paper itself says work use adds about $9 per month to the median user’s valuation. Stripping that out further reduces the estimate.
The average is dominated by outliers. The median respondent values AI at $11.40 per month. The average is $124.50. The 11-fold gap means a small number of enthusiasts are pulling the average up dramatically. If you trust the median, the aggregate is $16 billion. If you trust the mean, it is $172 billion. The methodology is only as trustworthy as the tail.
Put the four corrections together and a defensible estimate of US consumer surplus from generative AI is about $25 billion per year, with a plausible range of $5 billion to $60 billion. Still real. Still growing. But not $172 billion.
Meanwhile, revenue caught up fast
The original 2024 headline compared $97 billion of consumer value against $7 billion of US revenue captured by four companies (OpenAI, Microsoft, Anthropic, Google). That $7 billion number was narrow by construction. A broader bottom-up count was already larger in 2024, and the gap exploded from there.
The best bottom-up count comes from Menlo Ventures, which surveys about 500 US enterprise buyers and builds a detailed model of actual spending. Their December 2025 report:
- 2023 US enterprise AI spending: $1.7 billion
- 2024: $11.5 billion
- 2025: $37 billion
On top of that, consumers spent roughly $12 billion on AI subscriptions in 2025 (ChatGPT Plus, Claude Pro, Gemini Advanced). Total US AI revenue for 2025: about $49 billion.
For early 2026, the individual companies are moving fast. Anthropic announced on April 6 that its annualized revenue had crossed $30 billion, up from $9 billion at the end of 2025. They more than tripled in four months. OpenAI disclosed $25 billion in early 2026. Microsoft’s AI business is targeting $25 billion in fiscal year 2026. Add Google AI products, Meta AI, xAI, and smaller players, and the US AI revenue run rate is probably above $80 billion.
A separate signal that revenue is real: according to a December 2025 Tropic procurement report, enterprise software vendors are asking for 20 to 37 percent price increases on AI-enabled renewals, with negotiated outcomes landing about 12 percent above pre-AI baselines. That is well above the single-digit annual increases that had been the norm for enterprise software. That is the opposite of what you see in commoditizing markets.
Two ways to read the ratio
Put the numerator and the denominator together and two very different stories emerge depending on which welfare estimate you believe.
If you take the famous $172 billion figure at face value:
| Year | Consumer value (survey average) | US AI revenue | Ratio |
|---|---|---|---|
| 2024 | $97B | $11.5B | 8 to 1 |
| 2025 | $116B | $49B | 2.4 to 1 |
| 2026 run rate | $172B | $80 to $100B | 2 to 1 |
Even in this generous reading, the welfare gap has closed fast. Eight to one down to two to one in two years is an extraordinary rate of producer capture.
If you apply the standard corrections (the defensible estimate):
| Year | Corrected value | US AI revenue | Ratio |
|---|---|---|---|
| 2024 | ~$15B | $11.5B | About even |
| 2025 | ~$20B | $49B | Revenue 2.5x larger |
| 2026 run rate | ~$25B | $80 to $100B | Revenue 3 to 4x larger |
Under this reading, producer capture has already exceeded consumer welfare, probably since 2024. AI was never a huge gift to consumers in the sense the headlines suggested.
Both tables are defensible ways to read the data. The first uses the standard welfare-economics recipe. The second uses the same recipe with the corrections economists apply when they are being careful.
Why the bearish AI take is weakening
For the last year, a popular bearish view on AI investing has gone like this: AI creates value, but the value flows to consumers for free. The companies building AI spend hundreds of billions of dollars and can never earn it back. This is the fiber-glut analogy. The technology is real, but the producers who funded it go bankrupt.
The data above is hard on that view. Under any honest reading:
- AI company revenue is growing rapidly (Anthropic from $9 billion to $30 billion in four months, OpenAI from $6 billion in 2024 to $25 billion in early 2026, Microsoft AI from a $13 billion run rate first disclosed in January 2025 to a $25 billion target by the end of fiscal year 2026)
- Enterprise software vendors are pushing 20 to 37 percent renewal uplifts for AI features, with negotiated outcomes around 12 percent, well above the single-digit historical norm
- Paid consumer subscribers sit at 5.6 percent conversion and climbing
- The welfare gap, whether you measure it generously or conservatively, has compressed fast
The open question is whether the compression continues. If Anthropic’s growth rate slows, enterprise pricing plateaus, and consumer conversion stalls, producer capture stabilizes somewhere below what hyperscalers need to earn a return on $660 billion of 2026 infrastructure spending. That is still possible. But the trajectory so far does not support the framing that dominated 2024 discourse.
Three things to watch over the next year
Rather than predictions, here are three signals that would tell a careful reader which way this is going.
Paid conversion on ChatGPT. Currently 5.6 percent. If it drifts toward 10 or 15 percent as OpenAI rolls out tiered pricing, consumer willingness to pay is higher than revealed preference suggested and producer pricing has real room to grow. If it stays flat or falls, consumer pricing has hit a ceiling.
Enterprise renewal pricing. Late-2025 data showed negotiated AI-related renewal uplifts around 12 percent against a single-digit historical norm, and vendor initial asks in the 20 to 37 percent range. If negotiated uplifts hold or rise through the next renewal cycle in late 2026 and early 2027, enterprise buyers have accepted AI as a durable cost rather than an experiment. If prices roll back, it was a one-time upcharge on novelty.
The next Stanford welfare survey. If the mean willingness-to-accept number keeps climbing, the welfare story is durable. If it flattens or falls, what looked like welfare growth was partly salience and novelty.
None of these is a crystal ball. All three will move meaningfully in the next 12 months. Watching them is cheaper than having an opinion.
The story changed and not many people noticed
The $172 billion number is not fraud. It is a serious estimate built with a serious method. But like all welfare estimates from survey data, it is fragile at the edges, and the edges are where aggregates live. When you apply the standard corrections economists use when they are being careful, the true number is probably around $25 billion. Much smaller. Still real. Still growing.
What has changed since the 2024 headlines is not whether AI creates value. That is clear. It is how much of that value is still flowing to consumers and how much is being captured by the companies building AI. The answer is that producer capture has been catching up fast, probably faster than either side of the AI debate has noticed. The “AI is a historic gift to consumers” framing was true for a brief moment. It is not true now.