Value Selling
What Anthropic's IPO Price Asks of Its Customers

Contents
- 01A take-or-pay cost base and a pay-as-you-go revenue line
- 02What the price implies
- 03From Anthropic's revenue to the customer's benefit
- 04What the filing shows about customer return
- 05Where $850 billion a year can come from
- 06The case that the bar gets cleared
- 07What I would watch
- 08Method, assumptions and disclosure
- 09Sources
Anthropic's compute commitment stands at about $900.9 billion. The revenue to pay for it is almost entirely uncommitted. Customers pay by the token, month to month, and can stop the moment the math stops working. What keeps them paying is the real story of this IPO.
That is the part of the S-1 I keep coming back to. At $70 to $78 a share, the IPO values Anthropic at roughly $916 billion to $1.02 trillion. Most of the commentary asks whether that price is too high. I think the more useful question is what the price asks of the people paying the bills.
By my arithmetic, the midpoint asks Anthropic's customers to capture about $850 billion a year of economic benefit within five years. This piece shows where that number comes from, how much of it today's evidence supports, and where the rest has to be found.
$853B a year
Low to high case, $ a year
Low to high case, $ a year: $369B to $1.97T, base $853B base.
Serre projection from Anthropic S-1/A Amendment No. 2 (Nov 10, 2026). Range $369B to $1.97T.
A take-or-pay cost base and a pay-as-you-go revenue line
Anthropic's costs are contracted for years. Its revenue is not. Amendment No. 2, filed November 10, puts aggregate committed spend at about $900.9 billion, nearly all of it compute and infrastructure. About $451.6 billion lands before the end of 2030, and $70.7 billion is due within twelve months of September 30.
The risk factors describe much of this as minimum-purchase and take-or-pay: the payment is owed whether or not the capacity gets used. The same section says revenue is mostly consumption-based and largely not under long-term contract. Remaining performance obligations, the revenue customers are contractually committed to, were $13.8 billion at September 30.
What Anthropic has committed to spend, against what customers have committed to pay
$ billions. Every bar is drawn to the same scale.
Committed spend, total
Purchase commitments and leases, running to 2045
Of which before 2031
About $106B a year (my arithmetic)
Of which due within 12 months
87% of run-rate revenue (my arithmetic)
Run-Rate Revenue, Oct 31, 2026
Non-GAAP, not independently verified
Q3 2026 revenue, annualized
$16.96B x 4 (my arithmetic)
Remaining performance obligations
Revenue under non-cancelable customer commitments
Cash and short-term investments
At September 30, before IPO proceeds
Think of a power plant that has signed twenty years of fuel contracts and sells electricity by the kilowatt-hour, with no customer under contract. The plant is fine for as long as customers keep the lights on. Customers keep the lights on for as long as the electricity earns them more than it costs.
How unusual is that? Large software companies typically have well over a year of revenue under contract. Anthropic has about ten weeks.
Revenue under contract, as a share of a year's revenue
Contracted future revenue (RPO) against the latest quarter's revenue, annualized
Snowflake is the fair comparison, because it also bills on consumption. Its customers still pre-commit roughly a year and a half of revenue. Anthropic's customers have not yet been asked to, or have not yet agreed to.
Some fairness on the big number. The $900.9 billion is a ceiling across two decades, not a bill due this year, and parts of it carry termination rights or depend on capacity being delivered. Near-term coverage looks comfortable: cash on hand is almost twice the next twelve months of commitments. The issue is the shape, not the solvency. Fixed obligations on one side, month-to-month customers on the other, and customer return on investment is the only thing holding the two together.
What the price implies
The midpoint price implies Anthropic needs about $284 billion of annual revenue in five years, 3.5 times today's run-rate. That is a projection, and here is how it is built.
The filing shows about 13.09 billion shares outstanding after the offering and the $26.0 billion concurrent private placement. Multiply by the range and basic equity value is $916 billion to $1.02 trillion, or 11 to 13 times run-rate revenue. Press reports of $1.3 to $1.4 trillion use a broader, fully diluted share count.
To back out the revenue that price needs, I asked a simple question. If investors require a normal equity return for five years, and the business then trades at a mature revenue multiple, how much revenue must it have?
The revenue the IPO price needs in year five, and the benefit behind it
Start from one of the three cases, then move any input.
Basic equity value
$968B
Times run-rate revenue
12.0x
Year-5 revenue the price needs
$284B29% a year from $81B
Benefit customers must capture by year five, at 3x
$853B a year
Three checks on whether that base case is reasonable.
- The multiple. Six times revenue at a 25% net margin is 24 times earnings, close to where large technology companies trade. A richer multiple lowers the bar; a thinner margin raises it.
- Other people's math. Economists Bernstein and Cummings ran an earnings-based version on the earlier $1.5 to $2 trillion talk and got $519 billion of revenue by 2031. Scaled to today's lower price, their method and mine land in the same neighborhood.
- The compute bill. Pre-2031 commitments average about $106 billion a year. Compute ran at 57% of revenue through nine months of 2026, so that bill pairs with roughly $186 billion of revenue. The base case sits above it.
Notice what the base case does not require: anything like the current growth rate. Third-quarter revenue grew 1,127% year over year. The price needs 29% a year. On revenue alone, this does not look like a heroic ask.
From Anthropic's revenue to the customer's benefit
Revenue is the wrong place to stop, because nobody pays a dollar to get a dollar back. Every dollar of Anthropic revenue is a dollar of customer cost, and a customer keeps paying only while the benefit clearly exceeds it.
How much more is a judgment call. I use three dollars of gross benefit per dollar of fees, with a range of two to four. That is my assumption, and it is a low bar on purpose. Claude fees are only part of what a deployment costs once you count integration, review, training and the other vendors in the stack.
What customers spend with Anthropic, and the benefit that spend has to buy
$ a year, both views on one scale. Serre projection.
Today, at run-rate
Range at 2x to 4x: $162B to $324B
Year 5, base case
Range, low to high case: $369B to $1.97T
This is what I mean by what Anthropic is on the hook for. Not legally. The legal obligations are the compute contracts. Economically, the company owes its customers enough value that they keep the meter running. At the IPO midpoint, that is about $850 billion a year by 2031.
For scale, $853 billion is about 4% of the combined operating and capital spending of the Fortune 500, and about 2% of the $40 trillion the filing gives as the value of global knowledge work. Anthropic's own penetration metric points the same way from the other side: its revenue from Fortune 500 customers averages under 0.1% of their total expenses, and under 1% for its ten largest.
One wrinkle makes the bar higher than it looks. The filing says cloud marketplaces carry 44% of revenue and the top five customers about a quarter. Some of that is likely other companies reselling Claude inside their own products, though the filing does not size it. Where that is so, the end customer has to clear a return on Anthropic's fee plus the reseller's markup.
What the filing shows about customer return
The filing asserts that customers earn a return and then measures something else. Across more than 400 pages, the phrase appears in one sentence, repeated: the company believes enterprise customers achieve a "compelling return on investment" from Claude. The evidence offered is net revenue retention of about 900%.
Retention is the right metric for an investor. It is the wrong evidence for a customer's return. A 900% figure proves customers spent nine times more. It does not show what they got back. The risk factors add that these operating metrics come from internal systems and have not been independently verified.
The twenty customer case-study panels are where the value evidence lives, and they are useful. They are also mostly hours, speed and adoption. Two of the twenty put a dollar figure on customer value: an unnamed broadband provider's reported sevenfold return, and a legal-tech study's potential billing per hundred lawyers.
So we priced them. We took the thirteen cases with a recurring metric and built each one from the filing's number, the customer's own public financials, and labeled assumptions. The full walk-through is in What Anthropic's S-1 Case Studies Are Actually Worth.
Four of the priced cases, valued two ways
$ a year. Switch the lens to see what an hours-only reading misses.
Salesforce
8.1M hours of estimated annual productivity
$605M$91M cash
SAP
25 hours saved per consultant per month
$306M$123M cash
DoorDash
Major launch timelines cut by more than half
$4M$132M cash
AIG
40% more submissions bound in one business line
$0.7M$13M cash
All 13 cases
Time freed, at cost: $1.25B
Projected cash value, base: $622M (range $132M to $2.4B)
Two patterns matter for the valuation question.
Capacity is not cash. Where the only route to value is time saved, under a quarter of the freed capacity turns into money in our base case. An hour saved becomes a dollar only when someone names what it replaces: a hire not made, a contractor not renewed, an hour billed.
The big money is on the revenue line. DoorDash and AIG barely register as hours saved. Their value comes from products reaching customers sooner and more business written. An hours-only lens misses almost all of it.
None of this says customers are not earning a return. It says the filing does not show one in a form a finance team could check. Anthropic says as much from the other direction, writing that today's revenue does not reflect the full economic value Claude creates.
Where $850 billion a year can come from
Efficiency alone covers about half of the bar in my base case. The rest has to be growth. That is the finding I would put in front of any board approving an AI budget.
Customer benefit comes from two places. The first is capacity: the same work done in fewer hours. The second is expansion: revenue that would not otherwise exist.
How far capacity goes
Start with what has been measured. The St. Louis Fed's worker survey put generative AI time savings at 1.6% of all US work hours in 2025, rising to 2.2% by the second quarter of 2026. Those are self-reports. Forward estimates of the annual productivity lift run from about 1.0 to 1.8 points, with Goldman Sachs near 1.5. Daron Acemoglu's skeptical bound is far lower.
Then comes the step most business cases skip. A study of Danish workers found chatbot users reporting time savings of 2.8% of hours, with no measurable change in earnings or recorded hours. Workers estimated that only 3% to 7% of the time they saved translated into higher earnings. Hours saved are not cash saved until someone acts on them.
Who captures a saved hour?
A saved hour that never reaches a paycheck has not necessarily vanished. It can land in four places, and only some of them are a return the buyer can count.
The owners
Keeps a budget fundedThe firm does the same work with fewer paid hours, or more work with the same payroll. The gain shows up as margin, and it accrues to management and investors.
The customer
Competition turns the saving into lower prices or better service. Real value, but the buyer of the AI does not keep it.
The worker
Higher pay, or the same pay for an easier day. The Danish data says the pay route is rare so far.
Nowhere
Gets it cutThe hour dissolves into meetings, email and slack time. Nobody decided what to do with it.
For Anthropic's bar, the first route is the one that keeps a budget funded, and the fourth is the one that gets it cut. A benefit does not have to reach the whole workforce to count. It does have to reach someone who can find it in the numbers.
One caution on reading the study. It shows the gains did not reach wages. It does not show they reached profits instead. That part is still unmeasured.
My base case is generous on both steps. It frees 7.2% of global knowledge-work hours by year five, turns half of them into cash, and credits Anthropic with 46% of the result.
How much of the year-5 benefit bar capacity covers
$ a year, four scenarios on one scale. Serre projection.
Easy: low bar, high capacity
Bar $369BCapacity $859BGap None
Base, gains hold steady
Bar $853BCapacity $660BGap $193B (23%)
Base, gains taper each year
The case I lead with
Bar $853BCapacity $403BGap $451B (53%)
Hard: high bar, low capacity
Bar $1.97TCapacity $104BGap $1.86T (95%)
I lead with the taper row, because efficiency gains usually fade. The first hours automated are the easy ones. Put differently: for capacity alone to meet the base bar, about 9.3% of the world's knowledge-work hours would have to be freed and half of them converted to cash. The most optimistic path I can cite frees 8.5%.
From the value of freed hours to the benefit bar
Year 5, base case, $ a year. Serre projection, not a forecast.
Freed hours, at wage value
7.2% of $40T
Never turned into cash
Half is realized
Credited to other AI vendors
Anthropic gets 46%
Gains fade each year
25% smaller yearly
Covered by efficiency
What survives
Benefit bar at the $74 midpoint
3x on $284B revenue
Gap growth has to fill: $451B, 53% of the bar
Read it left to right. The hours AI frees are worth a lot at wage value, but each step takes a cut: half never becomes cash, more than half of what does belongs to other AI vendors, and gains fade as the easy work runs out. What survives covers less than half of the bar. The purple block is what growth has to supply.
What the gap asks for
The takeaway
Cost savings get customers about halfway. The other half has to be new revenue. Across the whole economy, that means new products and bigger markets, not companies taking business from each other.
The logic runs in three steps.
- Efficiency leaves a gap. About $451 billion a year of the bar is not covered by freed hours.
- Revenue only counts at its margin. A dollar of new sales is not a dollar of benefit. The customer keeps only the profit on it. At a 25% margin, closing the gap takes about $1.8 trillion of new sales a year.
- That is a very large number. It equals 8.6% of Fortune 500 revenue, or 23 times the $77 billion Menlo Ventures counts as today's enterprise AI application market.
So where can $1.8 trillion of new sales come from? There are four places, and only two of them add up.
Where the growth comes from
Does it add up across all customers?
Does not add up
Swapping old software for AI tools
A buyer replaces a $10M vendor with a $7M AI-native one
No. Only the $3M saved is benefit. The rest moved between vendors
Winning share from competitors
A fast adopter takes business from a slower rival
No. One customer's gain is another customer's loss
Adds up
New products and markets
Something that was not viable to build or sell before
Yes
Growing the existing business
Better pricing, conversion and retention bring in more demand
Yes, if total demand grows
So what. Any one company can earn its return by beating a rival. All of Anthropic's customers together cannot, because the rival is usually a customer too. For the IPO price to hold, AI has to make the pie bigger. That gives a simple test for any AI business case: does it create something new, or only move it from someone else?
The case that the bar gets cleared
There is a strong case that customers clear this bar, and it deserves a straight hearing.
- The revenue is arriving faster than the price needs. Run-rate revenue rose about $10 billion in October alone. At that pace, the base case's year-5 revenue shows up in under two years.
- Customers are widening, not just deepening. Customers paying over $100,000 a year went from about 950 to more than 11,900 in twelve months. About 250 pay over $10 million.
- The money is mostly new. Menlo Ventures' 2026 survey found 55% of incremental enterprise AI spend came from net-new budgets and only 22% from software budgets. Menlo is an Anthropic investor, so weigh that accordingly. Still, it is the pattern expansion would produce.
- The unit economics improved. Revenue per dollar of compute rose about fivefold from early 2025 to mid 2026. The second quarter produced the first operating profit, and nine-month operating cash flow was positive.
- Wallet share is tiny. Under 0.1% of a Fortune 500 customer's expenses is a small number to grow from.
- Some value never shows up as hours. Mozilla shipped fixes for hundreds of vulnerabilities Claude found. Bristol Myers Squibb cut a two-month data question to minutes. Neither fits a labor-savings formula, and both are real.
How much room there is to grow
Usage is wide but shallow, and that is the best argument for revenue expansion.
How widely and how deeply US workers use AI, mid-2026
Share of US workers, and of work hours
How often
Daily use, by industry
Hours
Half of US employees have tried AI at work. One in seven uses it daily. It touches about 6% of the hours worked. The filing shows the same pattern in dollars: enterprise users without Claude Code spend under $75 a month on average, while the top 5% of Claude Code users spend about twenty times that.
So expansion does not depend on signing new logos. It depends on turning occasional users into daily ones. That cuts both ways for the benefit bar. More usage means more fees, and every added dollar of fees needs its three dollars of benefit.
I find most of that persuasive on revenue. My caution is about what revenue proves. Enterprises blew through their AI budgets this year, by Menlo's account, and overshoot spending gets reviewed. Spend can run ahead of measured value for a year or two. It cannot for five.
The filing shows how quickly the meter can move. Customer spend fell about 20% in the two weeks around New Year, because usage follows working days. Third-quarter compute went back up to 58% of revenue, and the quarter swung to an operating loss. Revenue that can be turned down in a week is only as durable as the value behind it.
What I would watch
The stock will trade on revenue for a while. The value underneath it will be decided in customers' finance reviews. Each group has a different thing to track.
If you own or are weighing the stock
- Compute as a share of revenue. It was 57% through nine months. That ratio is the bridge between the commitments and the revenue they need.
- Contracted revenue against contracted cost. $13.8 billion against $900.9 billion today. Watch whether customers start signing for longer.
- Retention as cohorts age. A 900% figure in a year of first deployments says little about year three.
- Price. The filing attributes part of a 40% drop in cost per task to a lower price. Customer savings that come from price cuts are revenue Anthropic gave up.
If you are buying AI
- Name the capture action before you approve the spend. A hire avoided, a contract cut, hours billed, a product shipped sooner.
- Keep capacity and cash on separate lines, and count each benefit once.
- Measure cost per completed task. One customer in the filing, Rakuten, describes doing exactly this. It is the cleanest practice in the document.
If you are selling AI
- Lead with the growth case. Half of the bar cannot be met with efficiency, so half of your business cases should not be about efficiency.
- Bring a baseline finance can check and a number they can open: the formula, which inputs are confirmed, which are assumed.
Anthropic's founders write in their letter that the advice which stuck with them was never to let the company be valued above what it is really worth. I would hold AI value claims to the same rule. At this price, the market is saying Anthropic's customers will find about $850 billion a year of real benefit. Some of them already have. The rest of the proof is not in the S-1, and it will not come from Anthropic. It will come from the customers, one business case at a time.
Method, assumptions and disclosure
Every projection above comes from the inputs below. Change one and the answer moves, in some cases by a lot.
Every input behind the projections
Low, base and high cases
| Input | Low | Base | High | Basis |
|---|---|---|---|---|
| IPO price | $70 | $74 | $78 | Filing |
| Shares after offering and placement | 13.09B | 13.09B | 13.09B | Filing |
| Run-Rate Revenue, Oct 31, 2026 | $81B | $81B | $81B | Filing; non-GAAP, unverified |
| Required equity return | 10% | 12% | 14% | My assumption |
| Year-5 revenue multiple | 8x | 6x | 4x | My assumption |
| Customer benefit per $1 of fees | 2x | 3x | 4x | My assumption |
| Global knowledge-work wage base | $40T | $40T | $40T | Filing's estimate |
| Productivity lift per year | 1.0 pts | 1.5 pts | 1.8 pts | Anthropic research; Goldman Sachs |
| Share of freed hours turned into cash | 25% | 50% | 75% | My assumption |
| Anthropic's share of AI-driven benefit | 35% | 46% | 55% | Run-rate over Menlo's market estimate |
| Taper: each year's gain vs. the last | 25% smaller | 25% smaller | 25% smaller | My assumption |
| Customer margin on new revenue | 15% | 25% | 40% | My assumption |
The weakest inputs, in order: Anthropic's share of the benefit, the hours-to-cash rate, and the pairing of required return with exit multiple. The bar moves more than fivefold between the low and high hurdle. I also apply US productivity estimates to a global wage base that is Anthropic's own figure.
Case-study values are Serre projections, not customer-reported results. Only the companies know their own denominators.
Key takeaways
- 1Anthropic has about $900.9B of committed spend against $13.8B of contracted customer revenue. Customer ROI is what holds the two together.
- 2At the $74 midpoint, the price implies about $284B of revenue in five years. That is 29% growth a year, far below today's pace.
- 3At three dollars of benefit per dollar of fees, customers must capture about $853B a year by year five.
- 4Efficiency covers about $403B of that in the base case. The $451B gap needs roughly $1.8T of new customer sales a year.
- 5Share shift between companies does not add up across the customer base. The bar is met by new products and larger markets.
- 6For buyers and sellers: name the capture action, keep capacity apart from cash, and lead with the growth case.
Sources
Filings
- Anthropic S-1/A, Amendment No. 2, November 10, 2026: price range, share counts, commitments, run-rate revenue, customer counts, remaining performance obligations
- Anthropic S-1/A, Amendment No. 1, November 2, 2026: third-quarter and nine-month results
- Anthropic S-1, October 26, 2026: risk factors, case-study panels (pp. 208 to 227), founders' letter
- ServiceNow Q3 2026 results, Salesforce Q2 FY27 results, Snowflake Q2 FY27 results: remaining performance obligations and revenue
Valuation and commitments analysis
- Bernstein and Cummings, The Fantastical Promises of a $2 Trillion IPO, October 30, 2026
- 9fin, Anthropic S-1 analysis: commitment terms by counterparty
- NBC News, Anthropic sets $1 trillion IPO valuation target: diluted valuation and offering size
AI productivity and spending
- St. Louis Fed, The State of Generative AI Adoption in 2025 and FRED Blog, August 2026 update
- Menlo Ventures, 2026: The State of Enterprise AI, November 3, 2026
- Gallup workplace AI survey, May 2026, as reported by U.S. News
- Humlum and Vestergaard, Large Language Models, Small Labor Market Effects
- Acemoglu, The Simple Macroeconomics of AI
- Goldman Sachs, Generative AI could raise global GDP by 7%
- Anthropic, Estimating AI productivity gains and Economic Index
- Fortune 500, 2026 list: revenue and profit totals
Turn every deal into a defensible case for change.
Value engineering for every seller. A person approves what enters the library.

