Anthropic has committed to spend ~$900B on compute. Its current annual run-rate is about $81B.
This post digs into how much revenue it takes to honor those commitments, where the revenue comes from now, and where the rest could come from.
The sources are Amendment No. 3 to Anthropic's S-1, Anthropic's Economic Index, US wage data, and some rough models of my own.
Treat every figure that isn't from the filing as an illustration.
Summary
Anthropic needs roughly three times today's revenue by 2030. If we naively model diverting wages to AI, there is enough US wage spend to cover Anthropic's needed growth. Most of the adoption and revenue growth would need to be in non-coding occupations.
- Anthropic has $900.9B in commitments, mostly for compute. $70.7B is due in the 12 months after September 2026 and $451.6B before the end of 2030.
- Revenue needed. On the payment schedule I estimate, Anthropic needs $140B–$160B in 2027, $219B–$250B at the peak in 2029, and $211B–$240B in 2030. That's 1.7-2.0x today's run-rate in 2027 and roughly 2.6-3.1x in 2029-2030.
- Growth pace. It won't take exponential growth in adoption. Adding about $7B to $9B of run-rate every month through 2027, and less after that, would be sufficient. Anthropic added about $3B a month in August and September, and $10B in October.
- Revenue today. About 71% of revenue is billed in the US, and roughly 40–50% comes from coding. About 74% of revenue comes from 11,900 customers. There's still a lot of room for growth in use cases outside coding and outside the US.
- Usage against wages. Relative to each occupation group's share of US wages, Claude is used heavily in software, writing and science, and lightly in management, healthcare and business operations. Together, those last three groups account for about 34% of the estimated US wage bill.
- Wage ceiling. Based on O*NET's categorization of tasks and occupations, I model $6.2T of US wages being exposed to capture by Anthropic. With my assumptions, Anthropic's revenue from software is about 4.3% of US software wages. If every exposed occupation reached that level of adoption, Anthropic would earn about $378B a year. This is more than it needs to meet its commitments by 2030.
- Non-coding growth. Non-coding revenue is about $43B today. If coding revenue grows 15% a year, covering the 2030 commitments takes 3.4x to 4.1x today's non-coding revenue, or 39% to 47% growth a year. With coding flat it would be 46% to 52%.
- New capabilities and demand. All of this assumes no new capabilities and no new demand. In a toy model where half of revenue growth comes from new uses, non-coding revenue from existing work needs 29% growth a year, and the new uses need to approximately double Anthropic's number of large customers and subscribers.
What Anthropic has committed to spend
The filing, dated 17 November 2026, puts Anthropic's committed spend at $900.9B, substantially all of it for compute and infrastructure.
The filing gives two totals for when the money is due: $70.7B in the 12 months after 30 September 2026, and $451.6B before the end of 2030. The rest (about $449B) is due between 2031 and 2045.
See the appendix for the full breakdown by component, counterparty, and terms.
Anthropic says it intends to fund the commitments through the revenue it generates and the capital it raises. On 30 September 2026 it had $146.6B of liquidity: $95.6B of cash, $36.3B of short-term investments and a $14.8B undrawn credit line. Three suppliers had also agreed to lend it up to $77.0B. The IPO and a concurrent $26.0B private placement are set to raise about $105B before fees (and up to $117B if underwriters exercise their full over-allotment option).1
Together that is about $319B against the $451.6B due before the end of 2030. The rest has to come from revenue or from capital it hasn't raised yet.
How much revenue that needs
At most 26% of compute and infrastructure cost in the first nine months of 2026 went to serving customers. The rest went to training, research, and internal use.2 This bill is fixed in advance and, in the filing's words, "generally must be satisfied regardless of our revenue or utilization".
Compute and infrastructure cost 160% of revenue in 2025, 57% in the first nine months of 2026, and 58% in the third quarter. The best quarter so far is 48%, the second quarter of 2026.
I use two shares: 57% (the nine-month figure) and 50% (close to the best quarter).
The filing gives an annual schedule for $171.6B of the $451.6B due before the end of 2030, and only two aggregate totals for the rest ($70.7B in the 12 months after September 2026 and $451.6B through 2030). Using the disclosed schedule as a base and fitting the remaining $280.0B to those two totals, I estimate $14.6B is due in Q4 2026, $80B in 2027, $112B in 2028, $125B in 2029 and $120B in 2030.3
| Year | Compute due (est.) | 57% compute | 50% compute | Multiple of today's ARR |
|---|---|---|---|---|
| 2027 | $80B | $140B | $160B | 1.7x to 2.0x |
| 2028 | $112B | $196B | $224B | 2.4x to 2.8x |
| 2029 | $125B | $219B | $250B | 2.7x to 3.1x |
| 2030 | $120B | $211B | $240B | 2.6x to 3.0x |
The last column sets the 57% and 50% figures against $81B, the run-rate at the end of October 2026. At 57%, revenue has to reach about 1.7 times that in 2027, peak at 2.7 times in 2029, and fall to 2.6 times in 2030.4
Caveats:
- Cash and expense are different. The amounts due are cash payments. I treat a dollar due as a dollar of expense in the same year, which isn't how leases are booked; the cost of hardware leases (such as Broadcom's, detailed in the appendix) shows up as depreciation and interest over the life of the hardware.
- The amounts due are minimums. They leave out $150B to $175B of uncounted arrangements (also in the appendix) and any new deals.
- Some of it can be canceled. I don't know how much of the $186.2B with termination rights or conditions falls before 2030.
What revenue is today
Quarterly revenue went from $376M to $16,956M in six quarters.
| Quarter | Revenue ($M) | Quarterly growth | Compute and infrastructure ($M) | As % of revenue |
|---|---|---|---|---|
| Q1 2025 | 376 | 944 | 251% | |
| Q2 2025 | 787 | +109% | 1,390 | 177% |
| Q3 2025 | 1,382 | +76% | 2,047 | 148% |
| Q4 2025 | 2,040 | +48% | 2,945 | 144% |
| Q1 2026 | 4,731 | +132% | 3,387 | 72% |
| Q2 2026 | 11,555 | +144% | 5,558 | 48% |
| Q3 2026 | 16,956 | +47% | 9,863 | 58% |
Revenue for the first nine months of 2026 was $33,242M, against $4,585M for all of 2025. The run-rate was $81B at 31 October, up $10B from 30 September (after adding about $3B a month in August and September).5 Reuters reported it at about $9B at the end of 2025. The filing gives no revenue guidance, and says more than once that it expects growth to slow.
Very little of this revenue is contracted in advance. The filing says many of its top customers have no long-term contract or have one that expires soon.
On profit, Anthropic is close to breaking even before non-cash items. The operating loss for the nine months was $1,260M, and the second quarter was profitable. Cash from operations was $6,569M. The net loss was $104.0B, of which $94.3B was a non-cash change in the fair value of convertible notes and similar instruments.
Now set that against the revenue needed. Reaching $140B to $160B of revenue in 2027 takes growth of about 6.4% to 7.9% a month from October 2026 to December 2027—or roughly $7B to $9B of run-rate added each month through 2027, and $2.9B to $3.6B a month on a straight-line path to 2030. Recent growth is about 14% a month: 13.6% a month across the third quarter, and 14% in October alone.6
So 2027 needs about half the current pace. After 2027 the required growth slows sharply on my schedule: about 40% in 2028, 12% in 2029 and roughly flat (-4%) in 2030 as existing commitments peak.
Where the revenue comes from
What Anthropic discloses
The filing cuts revenue five ways.
| Cut | What the filing gives | Period |
|---|---|---|
| Kind | API $26.3B (79%), Subscription $6.9B (21%) | Jan-Sep 2026 |
| Channel | 44% sold through cloud marketplaces | Jan-Sep 2026 |
| Customer size | More than 11,900 customers spend over $100,000 a year and account for 74% of revenue. About 250 are above $10M | 12 months to Sep 2026 |
| Concentration | Top five customers about 25% of revenue. None above 10% | Jan-Sep 2026 |
| Geography | United States $23,470M (71%), International $9,772M (29%) | Jan-Sep 2026 |
API (consumption) revenue is billed by usage, and subscription covers both business seats and personal plans. The filing doesn't cross its tables, so I can't say how much subscription revenue is international or how much API revenue comes through marketplaces. Marketplace revenue is Claude sold through Amazon, Google and Microsoft.7
Geography is by the customer's billing address, separated by US and International addresses.8 The International share has hovered around 30% rather than steadily rising: 32% in 2024, 27% in 2025, and 29% in the first nine months of 2026. The 2025 dip coincided with a surge in US-billed cloud marketplace and top-customer spend, and the share ticked back up to 31% in the third quarter of 2026 as both of those eased.
More than 1,900 customers spend $1M or more a year. The rest of Anthropic's revenue comes from about 40 million smaller customers.
It gives no revenue by country, industry or product, so there's no breakdown between direct API use, Claude Code and claude.ai, and nothing on how much is coding.9
Estimating revenue by usage
The filing says nothing here, so everything below is reported by someone else or estimated by me. Most of the spending figures come from Menlo Ventures' 2026 enterprise survey. Menlo is an Anthropic investor.10
Coding is the largest use. Menlo puts enterprise spending on AI applications at $77.3B a year at the September 2026 run-rate. Coding is $34.0B of that, 44%, and Claude Code and Codex together are $24.0B. My estimate is that coding is roughly 40-50% of Anthropic's revenue.11
Finance is reportedly second. Anthropic told Reuters in May 2026 that 40% of its top 50 customers are financial institutions and that finance is its second-largest industry by enterprise revenue, after technology. It gave no dollar figure.12 The filing notes that the 10 largest banks in the Fortune 500 are among its customers above $100,000 a year.
Menlo's figures for healthcare ($3.7B, 77% of it from providers), customer service ($2.7B), and legal ($1.8B) are spending on applications from all vendors. How much of it reaches a model lab varies widely.13 Harvey, the largest legal AI vendor, was reportedly paying more for models in June 2026 than it earned, and has since moved work to a cheaper model of its own.14 The filing counts 9 of the 10 largest pharma companies and 8 of the 10 largest healthcare companies as customers.
For education and government I found no revenue figure of any kind. Coursework fell from 19% of claude.ai conversations in November 2025 to 12% in February 2026, per the Economic Index's March 2026 report. This trend is likely an artifact of the academic calendar; coursework-related usage was 16.5% globally in May 2026. About 500 US federal, state and local organizations use Claude, by the company's count. Civilian agencies can buy it, and defense work is closed to it for now.15
Spending is concentrated in a few firms. Ramp's AI Index tracks what US businesses pay AI vendors on its cards.16 The median paying firm spent $16 per employee per month on AI in October 2026. The top 10% spent $801 and the top 1% spent $10,199. About 1% of firms generate about 80% of spending on OpenAI and Anthropic.
So: today's revenue is about 71% billed to US addresses and roughly 40–50% comes from coding, and most of it comes from a small number of firms that spend heavily.
What Claude is used for
Anthropic's Economic Index is its public dataset of what people ask Claude to do. I use the release of 26 June 2026, which covers April and May. A classifier assigns each conversation to a task in O*NET, the US government's catalog of occupations and the tasks that make them up.17 Each task belongs to an occupation, so each conversation ends up counted under one occupation. Using this data, I try to estimate Claude's adoption for particular occupations and industries. Cross-referencing that with total wages in each industry gives an estimate of how much revenue Anthropic generates from each industry (and how much more revenue it can hope to generate).
One caveat: a conversation is counted under the occupation whose work it looks like, no matter who is asking. For example, a patient asking Claude to explain a blood test counts as healthcare work, though no healthcare worker was involved.
A raw usage share doesn't say much on its own. An occupation can have a large share because Claude is used heavily in it, or simply because the occupation accounts for a lot of work. So I compare each occupation group's share of Claude usage with its share of the estimated US wage bill.
Wages only pay for work, though, and a lot of claude.ai use isn't work. The Economic Index also labels each conversation as work, personal or coursework.18 In the US, 41% of claude.ai conversations are work, 50% are personal and 9% are coursework. On the API, 84% are work.
The split varies a lot between occupation groups.
I multiply each group's share of conversations by the fraction of them labeled work, then rescale so the groups add up to 100%.
I estimate the US wage bill from the May 2025 BLS occupational wage survey. Multiplying 155.5M jobs by the mean annual wage of $69,770 gives $10.85T.19
In other words, the ratio is a group's share of work usage divided by its share of the wage bill. If each group's share of work usage matched its share of the wage bill, every group would sit at 1.0x. Above 1.0x, Claude is used more per dollar of wages than average, and below 1.0x it's used less. A large group far below 1.0x is a lot of paid work that Claude barely touches yet.
The table has the 12 of 22 occupation groups where most of the work happens at a desk or partly at one.20 The claude.ai (US) column uses May data for US Claude chat and Cowork on Free, Pro and Max. The API column uses May data for global first-party API usage and excludes Claude Code.21 Both count work conversations only.
| Occupation group | Est. US wage bill | US claude.ai ratio | Global API ratio |
|---|---|---|---|
| Management | $1,617B | 0.59x | 0.49x |
| Healthcare practitioners | $1,068B | 0.07x | 0.06x |
| Business and financial operations | $1,003B | 0.81x | 0.57x |
| Office and administrative support | $915B | 1.32x | 1.64x |
| Sales | $738B | 0.83x | 0.54x |
| Computer and mathematical | $632B | 4.73x | 5.49x |
| Education and library | $615B | 1.27x | 1.05x |
| Architecture and engineering | $280B | 1.55x | 1.85x |
| Legal | $182B | 0.78x | 0.25x |
| Community and social service | $169B | 0.37x | 0.47x |
| Arts, design and media | $163B | 10.86x | 8.57x |
| Life, physical and social science | $139B | 3.25x | 3.59x |
Writing, software, and science sit at 3x to 11x; per dollar of wages, Claude is used that much more than average. The three largest groups in the table, management, healthcare practitioners and business operations, are all below 1.0x. Together, these groups account for $3.7T in wages, or about a third of the estimated US wage bill. The ten groups I left out are physical work, and they sit between 0.03x and 0.33x on claude.ai, apart from personal care at about 0.7x.22
The gaps are wider at the level of single occupations. The release gives work shares for single occupations only globally and not for the US, so this table counts all conversations.
| Occupation | Est. US wage bill | US claude.ai ratio |
|---|---|---|
| General and operations managers | $473B | 0.04x |
| Registered nurses | $343B | 0.04x |
| Pre-K to 12 teachers | $308B | 0.07x |
| Accountants and auditors | $137B | 0.09x |
| Customer service representatives | $121B | 0.07x |
The 25 largest US occupations by wage bill make up 34% of the estimated US wage bill. But these same occupations only make up 9% of claude.ai usage and 6.5% of API usage. At the other end, writers and authors account for about 0.04% of the estimated US wage bill, while tasks associated with that occupation account for 4.1% of US claude.ai conversations.23
I built an interactive explorer of the same release, which goes down to single tasks.
Wages as an upper limit
If a model does work that people are paid to do, the most anyone would pay for it is roughly what that work costs now. So the wage bill of the occupations whose tasks models can already do is a rough upper limit on revenue from doing or assisting that work.24 A vendor captures a small fraction of it. The buyer keeps most of the saving, and competing vendors bid the price down.
The wage bill itself has barely moved so far. The August 2026 update of Stanford's "Canaries" paper finds employment of 22-to-25-year-olds in highly exposed occupations 19% below its counterfactual, the level the authors estimate it would otherwise have reached, through less hiring (rather than layoffs).25 A separate Anthropic paper (Massenkoff and McCrory, March 2026) finds no systematic rise in unemployment for workers in exposed occupations. I found no study showing falling wages for people already in those jobs. So, I believe it should be sound to treat the wage bill as a usable yardstick for the next few years.
Revenue as a share of wages today
I split the $81B run-rate across the 22 occupation groups. Anthropic doesn't disclose revenue by product or by use, so we make some assumptions to estimate revenue split by occupation groups.
- Claude Code is 25% of revenue, and all of it goes to computer and mathematical occupations.26
- The rest of consumption revenue, 54% of the total, is split by each group's share of work conversations on the API.
- Subscription revenue, 21% of the total, is split by each group's share of claude.ai conversations.27
- 70.6% of each group's revenue is from the US, the share in the filing for the first nine months of 2026.
I divide each group's estimated US revenue by its US wage bill to get what I call intensity.
Intensity describes how much Anthropic takes in for each dollar the US already pays people in an occupation.
Software, by which I mean the whole computer and mathematical group, comes out to 4.28%.28 So, for every $100 the US pays in software wages, Anthropic takes in about $4.28. Arts and media is 3.45%.29 Science is 1.42%, and everything else is under 1%, down to 0.06% for healthcare practitioners. Across all occupations, the US share of the run-rate is $57.2B (70.6% of $81B), which is 0.53% of the $10.85T estimated US wage bill, or about 53 cents per $100 of wages. Personal use is in that revenue and has no wage behind it, so the share for paid work is even lower.
Anthropic's revenue ceiling based on adoption
Software is the occupation where adoption has gone furthest, so I use it as the benchmark.30 Suppose every desk occupation reached software's intensity today, 4.28%. Mixed occupations, where part of the work is physical or in person, reach half of that. Physical work stays at zero.31
Here \( w_g \) is 1 for desk groups, 0.5 for mixed groups and 0 for the rest, and software's intensity is 4.28%.
This model leaves $6.2T of US wages exposed to AI. This is the upper limit if we replaced all desk work with AI (and priced it like we do humans).
Anthropic (probably) won't automate and capture all wage labor. So, we model Anthropic's ceiling for revenue as other occupations adopting Claude at software's intensity today. With these assumptions, the ceiling is $267B a year in the US and $378B globally.
So the paid work that Claude is already used for is a large enough pool to cover the commitments. If, and only if, there is wide adoption of Claude for non-coding tasks.
Reaching that $378B ceiling by the end of 2030 while holding software at today's intensity implies 64% annual growth in non-coding revenue, to about 8x today's level. In that scenario, revenue from office and admin would have to reach 7.5x today's level, and sales 7x. Business and financial operations would need to reach 18x, management 23x, legal 31x and healthcare practitioners 38x. Quite dramatic, but possible, given that adoption in these occupations is currently very low.
Software adoption will continue to grow
I expect coding revenues to keep growing, just more slowly than everything else. Most developers already use AI tools, so the growth has to come from each developer spending more.
Let's assume coding revenue grows 15% a year, from $38.3B today to about $64B in 2030. I would argue this is fairly conservative, as evidence from OpenAI and Ramp suggests that some firms are finding lots of benefits to spending more on AI per employee.
That defines the base case I'll use for the rest of the post (at the stricter 50% compute target):
| Piece | $B |
|---|---|
| 2030 revenue target (at 50% compute) | 240 |
| Less: coding in 2030 (growing 15% a year from $38.3B) | −64 |
| Less: non-coding revenue today | −43 |
| Non-coding growth needed (the $133B gap) | 133 |
Under this assumption, the other occupation groups don't need to reach software's current intensity to cover the 2030 commitments—and against the total wage bill, the required capture is a small sliver of the market. The top bar sets Anthropic's US revenue today, what it needs in 2030, and the 4.3% ceiling against the $6.2T of exposed US wages; the bars below zoom into that slice by occupation group (global revenue).
Management, business and financial operations, and office and admin bring in about $15B today (my estimate). To cover commitments by 2030, they would need to bring in about $103B, an intensity of 2.0% to 2.2%. At software's intensity they would bring in about $215B. So Anthropic's future revenue could rest on whether managers, analysts and administrators come to spend like developers.
Overall, covering the 2030 commitments in this scenario takes 3.4x to 4.1x today's non-coding revenue ($146B to $176B, or 39% to 47% growth a year). At the $240B target—where each non-coding group closes the same share (about 44%) of its gap to the ceiling—revenues from the largest under-penetrated occupations need to reach these multiples:
- Office and admin: 3.9x (to $28.7B, down from 7.5x)
- Sales: 3.7x (to $11.7B, down from 7x)
- Business and financial operations: 8.7x (to $28.7B, down from 18x)
- Management: 11x (to $45.7B, down from 23x)
- Legal: 14x (to $5.1B, down from 31x)
- Healthcare practitioners: 17x (to $14.8B, down from 38x)
The remaining desk and mixed groups need smaller multiples: social service about 3.5x, engineering 3.3x, education 2.3x, science 1.9x, and arts and media 1.1x.
What if Anthropic creates new demand?
The previous section makes two strong assumptions: that models gain no capabilities beyond what Claude can do today, and that demand is fixed. Neither is true.
Models will get better
Anthropic already has a more capable model than the one most customers can buy. It released Mythos in April 2026 as a gated preview for select organizations, and Fable 5 later as a generally available model with safety restrictions.32 The filing ties capability to revenue directly: as it delivers more capable models, "customers have expanded their usage into new use cases and increased their spending accordingly."
In our wage model, a new capability does two things. It adds occupations to the $6.2T of exposed wages. And it raises the intensity an occupation can reach. Software's 4.28% is where spending is today, and as models continue to develop capabilities across the software development lifecycle, I expect this spending to increase.
Cheaper work reaches new buyers and creates new markets
The wage bill only counts work that someone already pays for. A model can also do work that nobody bought because it cost too much.
Take a first review of a 30-page contract. One to two hours of a US lawyer's time costs $362 to $724 at Clio's average rate. The same review done by an AI agent could be a few dollars. At a few dollars, someone who'd never hire a human lawyer to read a lease might pay an AI to do it.
There is existing, latent demand for legal work. The Legal Services Corporation found that low-income Americans get no help or not enough help for 92% of their substantial civil legal problems. I'd guess the same is true of tutoring, tax questions, and second opinions on a medical bill.
None of that spending is in the estimated ceiling because nobody earns a wage for it today.33 So as models improve and prices fall, the market can grow larger than it is now.
A toy model with two sources of growth
Suppose the growth comes from two places.
- Existing work. Employers spend a share of what they pay in wages on Claude, as in the previous section.34
- New use. Businesses and consumers pay for something they didn't buy before.
Start from that same $133B non-coding gap (call it \( G \)): a share \( s \) comes from existing work and the rest from new use.
New use has two kinds of buyer. A business pays for tokens as it uses them, and a large customer spends millions a year. A consumer pays for a subscription at about $20 a month. Today 79% of Anthropic's revenue is consumption and 21% is subscriptions, which is $64B and $17B of the run-rate.35 I assume new use splits the same way.
Here \( B \) is the number of businesses and \( b \) what each spends a year, and \( C \) and \( c \) are the same for consumers. For \( b \) I use $5M, the average for the 11,900 customers that spend more than $100,000 a year.36 For \( c \) I use $240, which is $20 a month.
| Growth from existing work | Annual non-coding growth | New business spend | New customers | New consumer spend | New subscribers |
|---|---|---|---|---|---|
| 100% | 47% | $0 | 0 | $0 | 0 |
| 75% | 38% | $26B | 5,300 | $7B | 29M |
| 50% | 29% | $53B | 10,500 | $14B | 58M |
| 25% | 17% | $79B | 15,800 | $21B | 87M |
| 0% | 0% | $105B | 21,000 | $28B | 116M |
In the first row, existing work accounts for all of the $133B gap, so non-coding revenue needs to grow 47% a year. If half comes from new use, that falls to 29% a year.
On today's revenue split, that's 10,500 more customers at $5M a year and 58M more subscribers at $20 a month. Anthropic has about 11,900 customers of that size now, and its $17B of subscriptions is 71M people at $20 a month. So half the growth from new use means nearly doubling both.
Recap
- Anthropic needs to roughly triple its revenue in three years. It seems quite possible (I daresay even probable) that it will.
- Much of this growth will come from expanding non-coding usage.
- Increasing capabilities will enable new avenues for revenue growth. Novel markets will be created by new capabilities, and increased subscription adoption seems to follow better models.
I think we live in a world where improved model capabilities will translate into growing markets. Several new capabilities come to mind on the consumer side: improved memory, increased trust and security around sensitive domains, better long-horizon planning, and a better ability to ask users questions and help them clarify what they want. Personal AI use cases such as optimizing finances, handling medical and logistical issues, and collaborating on hobby projects seem blocked by these limitations.
And on the enterprise and science side, we are already seeing AI help discover new drugs and materials and accelerate the legal processes around both. Tens of billions of dollars of value can be created by accelerating these R&D processes.
There is a whirlwind of sentiments and uncertainty around Anthropic's IPO and the financials of AI. It doesn't help that the companies themselves lack transparency, in part due to genuine concerns around privacy and security, but also due to competitive dynamics that incentivize secrecy. Despite data being in the dark, we can do our best, triangulating reality, part by part.
Appendix: Breakdown of the commitments
Anthropic's $900.9B of committed spend is made of five pieces.
| Component | $B |
|---|---|
| Contractual obligations at 30 September 2026, mostly cloud hosting purchases | 320.0 |
| Short-term commitments at 30 September (cancelable, or under one year) | 17.4 |
| Leases signed but not yet commenced | 451.2 |
| Commitments entered after 30 September, mostly data center leases and cloud hosting | 105.7 |
| Short-term agreements entered after 30 September | 6.6 |
| Total | 900.9 |
Half of the total is leases that haven't started.37 They start between 2026 and 2028 and run for about 10 years on average.
The filing names these counterparties.
| Counterparty | Service | Commitment | Period | Cancellation |
|---|---|---|---|---|
| Broadcom | Leases of TPU hardware (Google's AI chips) | $161.2B of rent, before interest | 5-year leases from 2026 | Non-cancelable |
| Cloud hosting | at least $121.3B | April 2026 to July 2033 | Take-or-pay | |
| SpaceXAI (formerly xAI) | Nvidia GPU capacity | up to $120.5B | into 2029 | Either side can cancel on 90 days' notice from early 2027 |
| Amazon | Cloud hosting | at least $110.0B | May 2026 to April 2036 | Take-or-pay |
| Microsoft | Cloud hosting | $47.2B to $49.2B | November 2026 to June 2033 | Minimum regardless of usage. Non-cancelable |
| Nscale | Nvidia GPU capacity | up to $44.6B | into 2033 | Fees start on delivery. Depends on Nscale raising financing |
| Lambda | GPU capacity | $41.3B | into 2034 | $34.9B of it can currently be terminated by either side |
These add up to about $646B. The Google and Amazon contracts are take-or-pay: Anthropic pays the minimum whether or not it uses the capacity.
Not all of the $900.9B is fixed. The two short-term lines, $24.0B together, are cancelable or run under a year. A further $186.2B is subject to termination rights or to conditions such as the counterparty securing financing, facilities or approvals. That leaves roughly $690B that I understand as Anthropic being on the hook for.
Some spending isn't counted at all. The total leaves out arrangements with no fixed dollar amount, specifically:
- additional Google TPUs that the filing estimates at $125B to $150B for delivery in 2028 to 2031,
- and an AMD agreement expected to exceed $20B.
How the annual schedule is estimated
The filing's contractual obligations table as of 30 September 2026 gives exact annual minimums for $171.6B of the $451.6B due before the end of 2030 (with the remaining $148.4B of that table due after 2030). The other $280.0B due by 2030 comes from items with no annual table: $24.0B of short-term commitments under one year ($17.4B at 30 September plus $6.6B entered after), uncommenced leases ($451.2B total, commencing 2026–2028 on a ~10-year weighted-average term, including Broadcom's $161.2B five-year TPU leases), and $105.7B of post-September deals.
| Period | S-1 contractual obligations (exact) | Estimated other commitments | Total |
|---|---|---|---|
| Q4 2026 | $5.2B | $9.4B | $14.6B |
| 2027 | $29.2B | $50.8B | $80.0B |
| 2028 | $44.4B | $67.6B | $112.0B |
| 2029 | $47.8B | $77.2B | $125.0B |
| 2030 | $45.1B | $74.9B | $120.0B |
| Total through 2030 | $171.6B | $280.0B | $451.6B |
Two filing totals pin down the shape of that $280.0B residual:
- Q4 2026 and 2027 ($70.7B next-12-month anchor). Putting $14.6B in Q4 2026 ($5.2B from the table plus $9.4B from short-term commitments and Q4 commencements, stepping up from $9.9B of actual compute cost in Q3 2026) and $80.0B in 2027 fits the $70.7B October 2026–September 2027 total smoothly: $56.1B in Q1–Q3 2027 ($18.7B a quarter) and $23.9B in Q4 2027, bridging into $28.0B a quarter in 2028.
- 2028 to 2030 ($357.0B remaining, or $119B a year on average). Because uncommenced leases all start by 2028 (flattening out thereafter), SpaceXAI's up-to-$120.5B contract runs into 2029, and the disclosed contractual obligations table itself peaks in 2029 ($47.8B vs. $45.1B in 2030), I split 2028–2030 as $112B, $125B and $120B.
Actual 2030 spend could still be higher once the uncounted 2028–2031 deals ($150B–$175B) and any new contracts are added.
Sources
- Filings. Anthropic, PBC, Form S-1 (26 October 2026) for the first half of 2026, Form S-1/A No. 1 (2 November 2026) for January to September, and Form S-1/A No. 3 (17 November 2026) for everything else, including the restated marketplace share, customer counts and concentration. Anthropic's IPO pricing announcement for the confirmed IPO price.
- Economic Index. Anthropic, Economic Index report: Cadences, data release 2026-06-26 (CC-BY), for usage by occupation; the March 2026 report for coursework shares.
- Revenue by use. Menlo Ventures, 2026: The State of Enterprise AI (3 November 2026) and 2026: The State of AI in Healthcare (14 November 2026), for application spending and token prices; Bloomberg's reporting on Harvey's margins, via The Next Web; Ramp's AI Index, data through October 2026, for spending per firm; Reuters, via The Star (5 May 2026), for financial institutions among Anthropic's top customers; TechRepublic on Claude for Government for the count of government organizations; Taft on the D.C. Circuit ruling for the supply-chain-risk designation.
- Wages and jobs. BLS, Occupational Employment and Wage Statistics, May 2025, for employment and wages by occupation; Brynjolfsson, Chandar and Chen, "Canaries" (August 2026), for employment in exposed occupations.
- Legal need and cheap services. Clio for lawyers' hourly rates; the Legal Services Corporation's Justice Gap report (2022).
-
Anthropic priced the IPO at $83 a share across 980M shares. Anthropic is selling 956.4M of those shares ($79.4B gross; the other 23.6M are from existing stockholders) alongside a $26.0B concurrent private placement to Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC, for $105.4B gross before underwriting fees. Underwriters also have a 30-day option to buy up to 147M additional shares ($12.2B). The $319B total nets out a $10B overlap between Google's loan facility and the private placement. ↩
-
The filing's contractual obligations table covers $171.6B through 2030 (peaking in 2029 at $47.8B). Fitting the remaining $280.0B to the next-12-month ($70.7B) and end-2030 ($451.6B) totals gives this schedule. See the appendix for the full breakdown. ↩
-
For comparison, Reuters reported in August 2026 that Anthropic's internal forecast for 2028 revenue was $190B to $200B. Reported sell-side estimates (from brokerage analysts) for 2027 are about $135B to $137B. ↩
-
Anthropic defines it as consumption fees for the trailing 28 days times 13, plus monthly subscription fees times 12. It includes customers later cut off for policy violations, whom the filing puts at 5% to 7% of nine-month revenue. Third-quarter revenue times four is $67.8B. The filing states the $81B and the $10B increase. ↩
-
Third-quarter revenue was 1.47 times the second quarter's (16,956 / 11,555), which is 13.6% a month compounded. October: 81 / 71 = 1.14. The filing says price cuts and contractual discounts took about $600M off third-quarter growth. ↩
-
Anthropic books marketplace revenue gross and pays the cloud provider a fee. ↩
-
A multinational billed in the US counts as US revenue wherever its staff use Claude. Outside the US, companies often bill through Ireland, the Netherlands or Singapore, so those countries probably have more revenue than their usage implies. ↩
-
In place of industry revenue it gives counts, which I quote under "Estimating revenue by usage". On coding it gives one comparison: Claude Enterprise users who did not use Claude Code spent under $75 a month on average in the third quarter, and the top 5% of Claude Enterprise users who used Claude Code spent about 20 times that. ↩
-
Menlo's figures are annualized run-rates for September 2026, built from a survey of nearly 400 US enterprise decision-makers. They are estimates of what enterprises spend, not vendor revenue. Menlo holds Anthropic shares, so I treat its figures on Anthropic's position with more caution than its market totals. ↩
-
My guess is that finance is 8% to 12% of enterprise revenue, or $4B to $7B at run-rate. Much of it is probably coding agents for bank technologists, so it overlaps with the coding figure. ↩
-
Menlo counts $10.1B for all industry-specific applications together, against $34.0B for coding. The legal figure fits vendor revenue: Harvey says it passed $400M of annual recurring revenue in September 2026 and Legora $200M in October. The healthcare figure is from Menlo's separate healthcare report. ↩
-
Bloomberg, via The Next Web, reported that Harvey's gross margin fell from about 50% in January 2026 to about -50% by June, after its agents drove token use up twentyfold, and turned positive after it launched a model built on open weights in August. ↩
-
The filing lists litigation over the US Department of War's supply-chain-risk designation as a risk. A district court had ruled largely for Anthropic. On 25 September 2026 a federal appeals court upheld the designation 2-1, and Anthropic has asked the full court to rehear the case. The designation covers the military and its contractors, not civilian agencies. ↩
-
Ramp sees card and bill payments from its own customers, which lean toward technology-forward US firms. In October 2026, 56.8% of them paid for some AI product. Large contracts paid by invoice or through a cloud marketplace are mostly outside this data. ↩
-
O*NET is the Occupational Information Network, maintained for the US Department of Labor. It describes about 900 occupations and lists the tasks each one involves. The Economic Index maps a conversation to the task it most resembles. ↩
-
Anthropic's classifier assigns a label to each sampled conversation, and the Economic Index release publishes the shares for each occupation group. A label describes what the conversation looks like and not who is asking. ↩
-
I use $10.85T as an illustrative estimate because the survey breaks wages down by occupation. The BLS leaves out overtime pay and nonproduction bonuses. The BEA puts US wages and salaries at $13.0T in 2025. That accounted for about 42% of US GDP. BLS survey methods · BEA wages and salaries · BEA GDP ↩
-
A SOC major group is the top level of the Standard Occupational Classification, the 22 two-digit families such as "Management" or "Legal" that the BLS reports wages by. I shortened some names. "Healthcare practitioners" is "Healthcare Practitioners and Technical", which includes nurses, physicians and pharmacists but not aides. ↩
-
The Economic Index gives API usage only as a global total, with no breakdown by country. So the API column divides each group's share of global API usage by its share of US wages, where the claude.ai column divides US usage by US wages. The API ratios are accurate only if API use outside the US has the same occupational mix as inside it. About 29% of Anthropic's revenue is billed outside the US, so a different mix abroad could meaningfully change these ratios. ↩
-
Food service is 0.03x, construction 0.05x, transportation 0.07x, cleaning and maintenance 0.15x, and production 0.33x. Personal care and service is 0.68x on claude.ai and 1.41x on the API, which I read as the task taxonomy more than as adoption. ↩
-
Shares for single occupations are noisy because they follow the task taxonomy. Searching for information lands in "Librarians", which gets 5.3% of US claude.ai conversations on a $10B wage bill. Personal-finance questions land in "Credit Counselors", which gets 1.9% on $1.6B, a ratio of 132x. The group totals are more trustworthy. ↩
-
Employer contributions to benefits and social insurance bring total US employee compensation to $15.7T in 2025, about 51% of GDP. I use wages here because the BLS publishes them by occupation. BEA compensation figures ↩
-
Brynjolfsson, Chandar and Chen, using ADP payroll data to June 2026. The gap was 15% in the July 2025 version. The paper is descriptive, finds no comparable gap for experienced workers, and says there is no evidence of economy-wide displacement. ↩
-
The 25% is my assumption. ↩
-
For subscriptions I use the global claude.ai shares, with personal and coursework conversations included, because subscribers pay for those too. For the API I use work conversations only. ↩
-
The group includes IT support, data scientists and statisticians. Its revenue is $20.3B of Claude Code, $14.0B of other API use and $4.1B of subscriptions. ↩
-
Arts and media gets $8.0B in my revenue split, close to its $9.9B ceiling. I doubt that employers of writers and designers pay Anthropic 3.45% of their wage bill. Much of it is other people asking Claude to write. ↩
-
Google's 2025 DORA report has about 90% of developers using AI tools at work, and Stack Overflow's 2026 survey has 58% using them daily. Both surveys cover AI tools generally. ↩
-
Desk groups: management, business and financial operations, computer and mathematical, engineering, science, legal, arts and media, office and admin. Mixed: community and social service, education, healthcare practitioners, sales. The other ten groups have zero weight, including construction and transportation. ↩
-
The filing says Mythos went first to select organizations for cybersecurity work through Project Glasswing, "rather than making it broadly available at launch", and lists delayed and limited releases as a risk to revenue. ↩
-
Unmet need isn't the same as demand at a price. Only 46% of the people in the LSC survey who didn't seek help gave cost as a reason. ↩
-
The wage bill grows too. If US wages rise 4% a year for the 50 months to the end of 2030, the $378B ceiling rises to about $445B with no change in adoption. ↩
-
Consumption and subscriptions aren't quite businesses and consumers. Subscriptions include seats that companies buy on Team and Enterprise plans, and some consumption is individuals on the API. The filing doesn't split revenue by type of buyer, so I use this as the nearest thing. ↩
-
The filing says more than 11,900 customers spend over $100,000 a year and make up 74% of revenue. 74% of $81B is $60B, or $5.0M each. The division is mine, and the top five customers alone are about a quarter of revenue, so the typical large customer spends much less than the average. ↩
-
A lease "not yet commenced" is signed, but the building or hardware has not been handed over. Under US accounting rules it goes on the balance sheet only then, so for now it appears in the notes. ↩