Anthropic's revenue jumped. Can it carry the buildout?
The AI-debt series, Part 4
The AI-debt series, Part 4
American railroads doubled their freight in the 1880s and grew revenue by half. How much of today’s AI growth turns into money depends on the same two numbers.
PIE
What’s going on. Anthropic, the company behind the Claude AI models, filed on Oct 26 to sell shares to the public. Figures in that S-1 filing show $11.555bn of recognized revenue in the second quarter of 2026 and $16.286bn for the first half. The unaudited Quarterly Consolidated Statements of Operations on printed page 158 show $10.581bn of total operating expenses and $974m of operating income for that quarter. For the first half as a whole, the company recorded a $693m operating loss, took in $3.929bn of cash from operations and held $131.651bn in cash, cash equivalents and short-term investments at June 30.
Why it matters. Anthropic has signed compute deals worth tens of billions of dollars over many years, and Part 2 described lenders financing chips for its use. One quarter with an operating cushion can’t show whether revenue will carry that. Our three June-baseline sensitivities, which begin before an already elapsed quarter, put revenue for the four quarters ending June 2029 between about $25bn and $194bn. We chose every assumption behind that spread. These are not October-calibrated forecasts, and even the upper path cannot establish repayment capacity without future costs and payment dates.
The history. From 1882 to 1890, American railroads more than doubled the freight they hauled. Freight revenue rose 51% because the average price per ton-mile fell 25%. More use brought more money, though not one for one.
Three questions that get blurred
“Is the revenue enough?” bundles three questions. Does the service business earn more than it costs to run? Does a batch of chips earn back its cost over its working life? Can lenders be repaid on schedule from the cash that actually arrives? A yes to the first says little about the third, and our model touches only revenue.
The $974m speaks to the first question, for the second quarter of 2026 only. Operating profit is an accounting result. It is neither net profit nor the cash left for lenders.
The first half shows why the categories matter. Anthropic posted an operating loss and still brought in $3.929bn of operating cash. The filing’s cash-flow statement shows how collection and payment timing, together with noncash adjustments, helped explain the difference. Accounts receivable, money customers owed, reduced operating cash flow by $4.982bn. Changes in prepaid expenses and other assets added $2.094bn; payables, accruals and other liabilities added $2.679bn. Deferred revenue, payments received before the related revenue is recognized, added $1.380bn. These selected items are part of the bridge, not a complete reconciliation, and advance payments alone do not explain the gap.
How thin the cushion was
Hold the quarter’s $10.581bn of total operating expenses fixed. Revenue could then have come in 8.43% lower before the quarter’s operating profit hit zero: $11.555bn times 0.9157 is about $10.581bn.
That is a conditional stress test on one quarter. Some costs, like the computing that answers each customer request, would likely shrink with demand. Others can arrive early. Anthropic expects new Google and Broadcom chip capacity to start coming online in 2027, and whenever capacity costs begin before the corresponding customer revenue is recognized, the cushion narrows.
The commitments and their limits
Anthropic has announced its largest obligations in round numbers. In April it said it was committing more than $100bn over ten years to AWS for up to 5 gigawatts of capacity. In November 2025 it committed to buy $30bn of Azure compute. Dividing $100bn by ten gives an average above $10bn a year, but the announcement sets out no payment schedule, and nothing we’ve seen shows how far these deals overlap with each other or with financing described elsewhere.
The evidence used here does not establish a matched schedule of rent, notes and purchase commitments. A financing commitment is not itself a full calendar of cash payments. Without that schedule we don’t estimate a lender coverage ratio or a default probability.
The same gap rules out a tempting shortcut. Subtracting Part 2’s $63bn financing from a quarter’s revenue mixes two categories and two clocks: capital arranged for equipment over years, and sales booked in three months. Lease costs already inside Anthropic’s expenses shouldn’t be added on top. And the lease expense in the accounts can differ from the cash paid to a lessor in the same quarter, so one can’t stand in for the other.
Revenue is volume times price
The Census Bureau’s historical statistics, series Q274 to Q283, show railroads hauling 39,302 million ton-miles of freight in 1882 and 79,193 million in 1890, up 101.5%. A ton-mile is one ton carried one mile. Freight revenue rose from $485m to $734m, up 51.3%. The average price fell from 1.236 cents per ton-mile to 0.927 cents, down 25%.
Multiply the changes: about 2.01 times the freight at 0.75 times the price gives roughly 1.51 times the revenue. The published figures are rounded, so the products miss by a hair. The rule holds: revenue equals paid volume times realized price.
Railroad rates say nothing direct about AI prices. The comparison teaches the multiplication. The Anthropic revenue figures we have don’t split into volume and price, so from them alone we can’t tell which one did the work.
How fast AI is spreading
Economists Alexander Bick, Adam Blandin and David Deming, writing for the St. Louis Fed in November 2025, found 54.6% of US adults aged 18 to 64 had used generative AI in August 2025, up from a revised 44.6% a year earlier. On a matched measure, the internet had reached 30.1% about three years after it opened commercially. Work use stood at 37.4%, covering 5.7% of work hours.
The authors’ tracker now shows August 2026 figures: 60.8% of adults 18 to 64, 40.5% of employed people using it for work, 5.3% of work hours. We haven’t audited the tracker’s definitions or revisions, so we keep the two vintages separate. That includes the lower hours figure, which could reflect changing use or changing method.
None of these counts paying customers.
Fast early adoption tends to slow. Census data show 18.0% of households online in 1997, 54.7% in 2003 and 71.7% in 2011: about 20.35% a year of share growth, then 3.44%. Households are a different denominator from individuals, and share growth isn’t revenue growth.
Changing how factories run took longer. Electric motors made up 5% of the rated horsepower installed in US manufacturing in 1899 and 82% in 1929, according to the 1930 Census. That measures installed capacity, not output or the share of plants using electricity. The series has a seam: the minimum product value for a factory to be counted rose from $500 in the 1919 and earlier surveys to $5,000 from 1923, so the points can’t be drawn as one smooth curve.
Business surveys have a seam too. The Census Bureau revised its core AI questions in November 2025, a break Federal Reserve researchers also flag. Our source note records the October release separately, with its collection dates and a further wording caveat. These surveys cannot be joined without checking what each question counts.
A model you can check with a calculator
Scenario planning makes the assumptions checkable. You build several conditional paths, run the numbers for each, and see which conclusions survive. These are conditional projections, not a single best estimate. We assign no probabilities.
Timing matters. This is a June-baseline sensitivity, not an October-calibrated forecast. The model starts with reported Q2 2026 revenue and projects Q3 2026, a quarter already over when this article was prepared. Its first modeled Q3 ranges from about $10.93bn to $13.48bn. The filing’s rounded 86-times narrative implies approximately $16.94bn for Q3, not an exact disclosed quarterly result. That approximation is above every original path. We retain the June-baseline table to show how the assumptions work, not as today’s revenue forecast. The model note separately shows how mechanically anchoring each path to approximately $16.94bn changes the forward amounts.
Our model starts from the $11.555bn quarter and moves it with five multipliers, each defined so nothing is counted twice:
- Pool: how much of the potential market is adopting AI at all.
- Conversion: what share of that pool pays for this kind of AI.
- Usage: how much work each paying user runs, in a fixed unit we call normalized workload.
- Price: what Anthropic collects per unit of that work after discounts and shifts in model mix. This is no list price per token.
- Share: Anthropic’s slice of the paid work.
Every driver is an assumption, not a measured Anthropic growth rate. We allow adoption to slow as the pool fills. The technical settings and limits are in the model note.
| Path | Adopting pool, first modeled year | Share who pay | Work per paying user | Price per unit of work | Anthropic share |
|---|---|---|---|---|---|
| Internet-like | +9.5%, then slowing | +2% a year | +25% a year | -15% a year | flat |
| Faster AI integration | +17.7%, then slowing | +8% a year | +50% a year | -10% a year | +5% a year |
| Price-compressed | +5.8%, then slowing | flat | +20% a year | -30% a year | -10% a year |
Here is one step worked out. Take the internet-like path to the second quarter of 2027. The adopting pool has grown about 9.5%, a multiplier of 1.095. Multiply the five dials: 1.095 × 1.02 × 1.25 × 0.85 × 1.00 is about 1.186, and $11.555bn times 1.186 is about $13.71bn for that quarter. Every figure below repeats that step quarter by quarter through mid-2029.
| US$bn, conditional on chosen dials | Internet-like | Faster AI integration | Price-compressed |
|---|---|---|---|
| Calendar 2027, sum of four quarters | 55.96 | 90.07 | 36.00 |
| Calendar 2028, sum of four quarters | 65.07 | 151.60 | 28.59 |
| Four quarters ending Q2 2029 | 69.77 | 193.66 | 25.41 |
| Q2 2029 quarter times four (reference only) | 73.34 | 229.76 | 23.20 |
These are amounts Anthropic could recognize as revenue if the paid work were served and booked in the quarter. The calculations and assumptions are in the source note.
What the paths teach
Adding up a year. The honest annual figure sums four quarters. Multiplying the latest quarter by four gives a run-rate, which overstates the past year when revenue is rising. In the faster path, the last quarter times four is $229.8bn, while the four quarters sum to $193.7bn. Using the baseline second quarter of 2026, multiplying by four gives $46.22bn, a reference point for scale. In May the company said its run-rate revenue crossed $47bn, an annualized pace for that month. Its recognized first-half revenue was $16.286bn.
Where the spread comes from. Falling prices can cancel rising use. The size of that effect depends on the assumptions we chose, not on a measured law of AI markets. Price and usage can also change together; our simple model treats them separately.
The jump we didn’t extrapolate. Revenue went from a directly disclosed $4.731bn in the first quarter to $11.555bn in the second, about 2.4 times. None of our paths grows that fast; the fastest adds about 17% in its first quarter. One quarter can’t show whether a jump was a step change, a trend or partly one-time.
The limits. These paths stop after three years. They require enough customers willing to pay, a feasible market share and enough working capacity to serve them. The model note spells out those conditions. None establishes profit, cash available for lenders or repayment capacity.
The case that it’s plenty
The strongest argument for yes is the record. On the figures we have, revenue more than doubled in a quarter. The first half brought in $3.929bn of operating cash despite the accounting loss. The company held $131.651bn at June 30, and personal adoption has outrun the internet’s early pace. In the faster path, revenue for the four quarters to June 2029 reaches $193.7bn, more than four times the baseline quarter-times-four reference.
That path needs a lot to go right. By mid-2029 Anthropic would have to serve about 6.8 times the baseline quarter’s paid workload, in normalized units. That describes the work, and it isn’t a chip count: efficiency gains and product mix change how much hardware the work takes. Chips still have to arrive, power has to connect, customers have to be able and willing to pay, and the revenue has to be recognized on time. As Part 3 laid out, delivered capacity is its own constraint, and our model doesn’t check it.
The case that price eats the growth
The strongest argument for no is the railroad run in reverse. In the price-compressed path, usage per payer still grows 20% a year. Price falls 30% a year and Anthropic’s share slips 10%, so those dials multiply to about 0.76 a year before the pool grows. Quarterly revenue drifts to $5.80bn by mid-2029, half of the baseline quarter’s revenue.
Usage rises in every path because we assumed it would. Our working hypothesis is that cheaper AI work draws more use, and we have no estimate of how strongly. The railroads’ volume outran their price cuts by enough to grow revenue by half. Our harshest path assumes AI volume doesn’t.
What would change this read
Several disclosures would sharpen it: revenue per unit of work alongside paid volume, so price and usage can be tracked separately; a schedule of rent, notes and purchase commitments with overlaps netted out; and a split between paid and free use. A run of quarters at or above the faster path, with capacity arriving on time, would push toward yes. Realized prices falling faster than paid usage grows, business adoption stalling once survey definitions settle, or new capacity costs landing ahead of the customers would push toward no.
The railroads hauled twice the freight and took in half again the money. For Anthropic, we can see the money. We can’t yet see the freight or the fare.