AI Demand and Potential
How fast people and companies are taking up AI, where use is concentrated, what it is doing to the economy so far, and what that implies for a model maker's revenue.
Summary
Demand for AI is rising on every measure in the public record. The share of adults who use it is up, and so is the share of firms. Job postings ask for it more often, and annual reports discuss it more. Use is heaviest in high-income countries and in desk work.
The economic payoff is harder to see. Investment in computers and software supplied about an eighth of U.S. GDP growth over the past year. Productivity growth has picked up since 2023. The data cannot yet credit AI for either.
For Anthropic, six quarters of revenue fit a market ceiling of $150 billion a year about as well as a ceiling of $1 trillion. The size of the market decides the forecast. The sections below estimate that size from the wage bill of the work people use AI for.
Model capability
Demand follows what the models can do. Epoch AI's Capabilities Index (ECI) combines many benchmarks into one score. The best available model has gained about 15 points a year since early 2023.
The lead changes hands often. In September 2026, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 tied at the top with an ECI of 168.8.
The inputs grow faster than the scores. Training compute for frontier models has grown about 4.9 times a year since 2018, across the 54 frontier models in Epoch's dataset. Each new generation costs more to train. Model makers sign multi-year compute contracts years ahead for that reason, as Anthropic's S-1 shows.
Adoption by people and firms
People
The Real-Time Population Survey asks a national sample whether they used generative AI. Economists Alexander Bick, Adam Blandin, and David Deming run it. Use among adults 18 to 64 rose from 44.6% in Q3 2024 to 60.8% in Q3 2026.
Use for work rose from 33.3% to 45.2% by Q2 2026, then fell to 40.5% in the latest quarter. It is the first decline in the series. One quarter does not make a trend.
Firms
The Census Bureau's Business Trends and Outlook Survey asks firms every two weeks whether they used AI in any of their business functions, a question Census broadened in November 2025. The current series starts in November 2025. Since then, the share rose from 17.3% to 24.6% by early October 2026.
Plans run ahead of use: 28.6% of firms expect to use AI within six months. Large firms lead. Among firms with 250 or more employees, 43.6% use AI. Among firms with fewer than 5 employees, 24.3% do.
Hiring and disclosure
Employers ask for AI skills more often. In Indeed's postings, the share of U.S. job ads that mention AI was 1.8% in September 2023, 3.6% in September 2025, and 7.2% in September 2026.
Annual reports show the same turn. The number of 10-K filings that use the phrase "generative AI" went from 69 in 2023 to 1,201 so far in 2026. Filings that mention "AI agent" went from 3 to 59.
Claude usage by country and occupation
The Anthropic Economic Index measures Claude usage relative to each country's working-age population. A score of 1.0 means usage in proportion to population. Australia leads at 6.4, followed by Singapore at 5.8 and Switzerland at 5.0. The U.S. ranks 12th at 3.9.
Income explains most of the pattern. Across 120 countries, the log of the usage index rises almost one for one with the log of GDP per person (elasticity 0.95, R² 0.81). A country with 10% higher income per person tends to show about 10% higher usage per working-age adult.
The relationship compares countries at one point in time. It does not prove that rising income causes more use. It does give a baseline for spotting outliers. Australia, New Zealand, and Tunisia sit well above the line. Saudi Arabia and Qatar sit well below it.
Occupations
The index also matches conversations to job tasks from O*NET, the U.S. catalog of what each occupation does. The match classifies the work in each conversation. It does not identify the user's job.
Tasks common in computer and math occupations make up 23.8% of Claude usage. Those occupations earn 5.8% of U.S. wages. Management runs the other way: 14.9% of wages and 5.9% of usage. Healthcare practitioners earn 9.8% of wages and account for 3.3% of usage.
The gap marks where usage has room to grow. Management, healthcare, and business operations hold the largest wage pools and the least usage relative to their size. The index's classifier estimates that a task that takes a person about 5 hours alone takes about 40 minutes of conversation with Claude. Usage splits close to evenly between augmentation, where the person stays involved (51%), and automation, where Claude completes the task (49%).
Effects on the U.S. economy
Spending is the clearest macroeconomic effect so far. Over the last four quarters, U.S. real GDP grew 2.2%. Business investment in information processing equipment and software added an average of 0.26 percentage points a year to that growth.
That is about an eighth of all growth from one category of investment: computers, servers, communication equipment, and software. BEA counts data center buildings separately, as structures.
Productivity matters more and is harder to read. Output per hour in the nonfarm business sector grew 2.45% a year on average from 2023 to mid-2026. From 2015 to 2019 it grew 1.41% a year.
Economists have seen this kind of gap before. In 1987, Robert Solow noted that computers showed up everywhere but in the productivity statistics. The gains came in the late 1990s, a decade or more after the investment.
Erik Brynjolfsson and colleagues describe a "productivity J-curve." Firms first spend on training, process changes, and software that the statistics treat as costs, so measured productivity can lag before it rises. The recent pickup may have many causes. These figures alone can't credit AI.
Market size
You can size the market from the work AI does. U.S. employers paid about $10.8 trillion in wages in 2025 across 155 million jobs. The 8 occupation groups that account for 83% of Claude usage earn $5.8 trillion of that.
AI revenue is some fraction of that labor cost. Multiply the share of tasks AI handles by what buyers pay for AI relative to the labor cost of those tasks. Each slider below holds an assumption. Move it to see what it does to the answer.
For scale, Anthropic's revenue in Q2 2026 ran at $46 billion a year. The scenario chart in the next section can use this result as its ceiling.
Anthropic revenue scenarios
Anthropic's quarterly revenue went from $376 million in Q1 2025 to $11.6 billion in Q2 2026. Analysts usually model revenue that grows fast and then levels off as a market fills with a logistic curve. The curve has a ceiling, the most annual revenue the market will support, and a speed. Given a ceiling, the six quarters pin down the speed.
The six quarters cannot pin down the ceiling. Fits with ceilings from $150 billion to $1 trillion a year all match the history about equally well, with a typical error of 14% to 17%. Every one of them puts Q3 2026 revenue near $17 billion. The amended S-1 tests the method. It cannot choose among ceilings. After 2026 the paths split: 2027 revenue ranges from $129 billion to $442 billion depending on the ceiling.
A second check drops the ceiling and asks how fast growth slows. From Q4 2025 to Q2 2026, revenue grew an average of 138% a quarter. If that rate halves every two quarters, 2027 revenue reaches $343 billion and levels off near $670 billion a year by 2030.
If the rate halves only every four quarters, revenue passes $3 trillion in 2028. That is more than a quarter of all U.S. wages. Arithmetic like that is why the ceiling matters more than the recent pace.
These would move the estimate: reported Q3 2026 revenue in the amended S-1, Run-Rate Revenue, net revenue retention as the customer base matures, the price per token as models get cheaper to serve, and whether the compute Anthropic contracted for 2027 and 2028 arrives on time.
Economics of AI demand
Diffusion
New technologies spread along S-curves. Adoption is slow while early users experiment, fast once the gains are clear, and slow again as the last holdouts adopt. Use by people is in the fast middle. Use by firms is earlier, near a quarter of firms.
An S-curve fitted to the Census series reaches half of firms around 2029. That fit rests on 11 months of data, so read it as a sketch. Large firms adopt first because they can spread the fixed costs of integration over more work. Those costs include data plumbing, security review, and retraining.
Price and quantity
When a resource gets cheaper, total spending on it can rise, because new uses become worth paying for. Economists call this the Jevons effect. The S-1 shows the mechanism inside one company. Q2 2026 revenue grew 1,368% from a year earlier while compute expense grew 300%. Each dollar of compute produced far more paid work.
If buyers' demand is elastic, falling prices per task expand the market faster than they cut revenue per task. The market size model captures this through the share of work AI handles, which rises as the cost per task falls.
Value capture
The value AI creates splits three ways: what buyers pay the model makers, what buyers save, and what workers gain in output and pay. The price slider in the market size model sets that split.
William Nordhaus estimated that innovators captured about 2% of the total value of U.S. innovations from 1948 to 2001. Users got the rest. Model makers' share depends on how far apart the leading models stay. When two labs tie on capability, as they did in September 2026, price competition favors buyers.
Constraints
Supply sets the limit today. Anthropic says it has run short of compute, buys capacity 12 to 18 months ahead, and pays more for short-term capacity when demand outruns its plan. Power, chips, and data center construction set the pace. Training compute can't keep growing near 5 times a year on any fixed budget, so the cost of the next model generation constrains every lab's growth.
Indicators to watch
Watch these to see whether demand keeps broadening: firm adoption past 30%, the direction of work use after its first quarterly decline, the share of job postings that mention AI, and productivity growth in 2027. For Anthropic, the next hard data is the amended S-1's Q3 revenue and Run-Rate Revenue.