What you have to believe for a $1.5 - $2 trillion Anthropic

By Lewis Grafton, StreetCred · 5 November 2026 · Educational analysis, not investment advice

Anthropic is heading for one of the largest IPOs on record, with talk of a valuation between $1.5 trillion and $2 trillion.¹ We built a discounted cash flow model from its amended S-1 to work out what you would have to believe at those prices. It is not a price target. The answer depends on a few judgements about 2036, so we set them out plainly and ran five scenarios.

The cases run from $17 to $166 a share. Our base case is $82.24 ($1,416B of equity) and the weighted value is $77.42 ($1,341B). The full model is free, and you can put your own numbers in it: the link is at the end of this post.

The five scenarios

Almost all the value in Anthropic rests on three questions. How big does spending on AI models get? Does the most capable tier keep a premium price, or does AI become a commodity? And does Anthropic hold its place among the leaders? Each case below is a different answer to those questions, and each is a full run of one model.

Base: AI becomes a large line item and Anthropic stays a leader (45%)

By 2036, spending on AI models reaches about 6% of what knowledge workers are paid in the major economies, up from about 1.1% today. In the US that is about $735 a month per worker. Agents take on real parts of jobs, and customers pay for that as a fraction of a salary. The best models keep a premium. Anthropic gives up some of its lead in usage-based spending, from about 53% to 45%, but still holds about 44% of the whole market. Revenue passes $195B in 2028 and reaches $738B in 2036, with an operating margin near 26%.

We think this is the most likely single path, but it is not neutral. Our research put the central estimate for the market at 5% of pay, with a range of 3% to 9%, so 6% is a view above the middle.² Treat Base as a bullish-leaning house view.

Volume leader: the market is much bigger and Anthropic keeps its share (20%)

Here AI spending reaches 9% of pay by 2036, the top of our research range, or about $1,011 a month per US worker, with more of the market outside the US. Anthropic holds the same 44% it has in Base, close to the 42% we estimate for the end of 2026. Revenue reaches $1,116B in 2036. Value is $122.09 a share, or $2,144B.

We do not assume Anthropic gains share in this case. Winning share in a market this much bigger, against OpenAI and Google, did not look realistic to us. Holding 44% is already demanding, given that Google has about 21% of enterprise API spend and can sell through Workspace and its cloud.³ You would want to see spend per worker rising quickly, more customers above $1M, and agents spreading well beyond software engineering.

Moonshot: Volume leader, plus a real business in life sciences (5%)

This case asks what happens if Anthropic becomes a major player outside its core market. Global drug sales grow to about $3T by 2036 and Anthropic takes 10%: $300B of revenue at a 35% margin, valued at 20 times operating profit and against $100B of capital.⁴ That adds $43.72 a share to Volume leader, for $165.81.

It is the least supported case, which is why it gets a small weight. Only 7.9% of drugs entering Phase I reached approval in 2011 to 2020, and our research counts only $0.3B to $12B of 2036 cash revenue from AI drug programmes, so $300B is about 25 times the high.⁵ It needs AI to shorten drug development a great deal, or Anthropic to buy late-stage assets.

Share loss: the market grows, but Anthropic’s share halves (15%)

AI spending still reaches 6% of pay, but rivals take the work. Anthropic’s share of the market falls to 23%. This is the case history argues for most directly: OpenAI’s share of enterprise API spend fell from 50% to 27% in two years while Anthropic’s rose from 12% to 40%.³ Enterprises use several vendors and follow capability. Stickiness protects today’s workloads for a year or two, but not new spending, which is most of the market while it grows. Revenue reaches $392B in 2036, and Anthropic is still paying for capacity it committed to at a larger scale. Value is $34.67 a share. You would want to see rival models leading on the benchmarks enterprises use, and top customers diversifying.

Commoditisation: the frontier loses its price (15%)

Good-enough models get cheap and most work moves to them. The market reaches only 3% of pay, the bottom of our research range, and prices sit near cost. Followers need less training spend, but Anthropic has already committed to the capacity: operating profit stays negative until 2032, with a low of minus 22.7% in 2029. Revenue reaches $247B in 2036 and value is $16.70 a share. You would want to see price cuts spreading beyond the roughly $600M of quarterly revenue the filing already attributes to them, and open-weight models matching frontier quality for most tasks.⁶

Side by side

Why these weights

The weights are 25% on the upside cases, 45% on Base and 30% on the downside, and they are the most subjective input in the headline. We weight Share loss and Commoditisation equally because we cannot rank them. No weighted case has demand stalling outright, which is itself a judgement that demand keeps growing.

We also ran four cases with no weight. Premium Frontier ($181.12) combines a 9% market, a bigger frontier premium and a high share, the most the core model can produce. Demand plateau ($14.19) stops the market at 1.5% of pay, a stress test of the compute commitments: liquidity falls $7.6B short before new money. The plan case ($87.14) is what $1.5T needs. Combined downside ($6.72) puts commoditised pricing and lost share together; moving 5 points of weight to it lowers the headline by $0.50.

What this does and does not mean

If you accept the Base assumptions, the model gives a value near the price talk: Base is 5.6% below $1.5T and the weighted value 10.6% below it, on one share count where $1.5T is $87.14 a share. It does not show that $1.5T is right. That gap is smaller than the effect of any one of several inputs. Thirty per cent of the weight sits on cases worth under $35 and 25% on cases worth over $120, so no single case is expected to land on the weighted value.

History adds caution. Among companies starting with sales above $50B, growth above 20% a year over five years happened in 1.4% of observations, and over ten years in none.⁷ Base grows 49% a year from 2026 to 2031 and 29.5% a year over ten years. The sample is old and Anthropic is a different kind of company, but it is the right place to start.

Our approach

The general approach

We started from what the filing states, then asked what would have to be true, not what the stock should trade at. Rather than one forecast with a range around it, we use one model and change its assumptions for each case, so differences between cases come from the assumptions and nothing else.

Revenue is built from the demand side, from jobs and what is spent per job, so that every case makes a claim you can check: how much would companies have to pay for AI, per worker, for this to be true? Costs are built from what Anthropic has already signed. Compute commitments of about $895B mean that a weak revenue case does not get a matching cost cut, and we think that is the most important feature of this company’s economics.

The valuation approach

We value the equity at 31 December 2026 with a discounted cash flow, before any money raised in the offering, so a primary raise does not inflate the answer. Five choices shape the result:

• A flat discount rate of 9.69%. The scenarios and the compute floor already carry early-year risk in the cash flows, and a higher early rate would count it twice.
• No probability of failure. For the same reason, the weak cases carry that risk.
• Commitments are operating costs, never also debt. Doing both counts them twice.
• Stock pay is a real cost. Because future grants are expensed, we do not also add the plan reserve to the share count.
• Growth stops adding value after 20 years. From 2047, new capital earns only the cost of equity. Two thirds of Base’s value still comes after 2036.

The weighted value is a probability-weighted average of the five cases. It summarises the range, and is not a forecast of any one outcome.

The business, from the filing

Revenue. $41M in 2023, $386M in 2024, $4,585M in 2025 and $33,242M in the first nine months of 2026. By quarter it rose from $376M in Q1 2025 to $4,731M in Q1 2026, $11,555M in Q2 and $16,956M in Q3, up 46.7% on Q2 and 1,127% on a year earlier.⁸

Consumption versus subscriptions. For the nine months, consumption revenue (API and metered use) was $26,306M, or 79.1%, and subscriptions $6,936M, or 20.9%.⁹ The model treats them separately because consumption pays per use, while a subscriber can use more compute than the fee covers.

Geography. By billing address, 70.6% of nine-month revenue was US and 29.4% international. Billing address is not where work happens: a multinational can be billed from the US with staff abroad.

Customers. More than 11,900 customers spent $100,000 or more over the trailing twelve months, up from more than 950 a year earlier, and they were 74% of revenue. More than 1,900 spent $1M or more. The top five were about 25% of nine-month revenue. Over half the Fortune 500 are customers, but their Claude spend averaged under 0.1% of their total expenses in Q3. Net revenue retention was about 900%, which mostly reflects expansion from a tiny base. Many contracts let customers cut usage or leave with limited or no notice.¹⁰

Channels. Cloud marketplaces carried 11%, 32%, 47% and 44% of revenue in 2023, 2024, 2025 and the nine months of 2026. Anthropic books that revenue gross, and the fees sit in sales and marketing.¹¹

Compute. Q3 compute and infrastructure cost was $9,863M, 58.2% of revenue against 48% in Q2. About $5.0B was short-term compute bought when demand outran contracted capacity, which costs significantly more. Cost of revenue, which is serving alone, was 14.3% of Q3 revenue; the rest of compute is training, research and internal use. The margin after all compute was 41.8% in Q3.¹²

Commitments. Committed spend is about $895.4B, substantially all compute and infrastructure. About $447.5B falls before the end of 2030, most of the rest in 2031 to 2045, and $69.8B in the next twelve months. Of the total, $213.8B can be terminated or depends on conditions before it starts. Named deals include Google (at least about $121B), Amazon (at least $110B), Microsoft (at least about $45B), Broadcom-supplied chip leases (about $161B of rent), SpaceXAI (up to $120.5B, cancellable on 90 days’ notice after early 2027), Nscale (up to $44.6B) and Lambda ($41.3B). Some contracts are take-or-pay. Contracted future revenue is only $13.8B.¹³

Losses. The operating loss was $8,060M in 2025 and $1,260M in the nine months of 2026. The Q3 operating loss of $567M came after a $1.96B non-cash donation-matching charge. The nine-month net loss was $104.0B, of which $94.3B was non-cash fair-value changes on convertible notes, derivatives and SAFEs. Operating cash flow was positive at $6.6B.¹⁴

Cash. At 30 September, cash and short-term investments were $131.8B. Reported liquidity is $146.6B including a credit facility. No interest-bearing debt is drawn, and the $6.2B of Amazon convertible notes convert into shares.¹⁵

Policy-violating revenue. The company estimates that 5% to 7% of nine-month revenue came from customers it found breaching usage policy, and expects stronger enforcement to reduce revenue and growth rates. We take a 3% drag on Q4 revenue.¹⁶

How the model works

One engine runs every case. A scenario changes inputs, and each case is a full re-run.

Revenue: slots × spend per slot × Anthropic’s capture

A “slot” is a job in an occupation where AI can take on tasks: managers, professionals, technicians, clerical workers. The core markets have 264.7 million slots and a $17.5T wage bill in 2025, of which the US is 73.7 million slots and $6.8T.¹⁷ Slots grow 0.5% a year and wages 3.5%, taking the core wage bill to about $27T in 2036.

Each US slot sits on a ladder: no paid AI; a standard seat at about $20 a month; a power seat at about $150; and an agent-heavy pay-per-use worker, priced as a percentage of that slot’s wage. The top rung carries the growth, because an agent doing part of a job is worth a fraction of the salary. In Base by 2036, 20% of US slots are on a standard seat, 30% on a power seat, 45% on agents and 5% on none, with agents priced at 19.21% of wage, about $1,525 a month. Average market spend is then about $735 a month per US slot, up from about $167 at the end of 2026, and about $1.6T across the core.

Anthropic’s revenue is the market times its capture, built in pieces: its share of subscription spend (24% now, 25% in Base 2036), the split of consumption between a frontier tier and cheaper tiers (frontier is 45% now, 40% in Base), and its share of each (frontier 65% now, 60% in Base; cheaper tiers 42.5% now, 35%). The starting shares are low-confidence, since nobody publishes spend by tier.

International revenue is not built independently. It is US revenue scaled by an assumed ratio, moving from 26% of core revenue at the end of 2026 to 40% in 2036. So the market-to-wages figure shows what the revenue implies, and does not confirm it.

Calibration to the Q4 2026 run-rate

FY2026 revenue is nine months of actuals plus a Q4 estimate: Q3’s $16.956B grown 35%, less the 3% enforcement drag, gives $22.2B, so FY2026 is $55.4B and the Q4 exit run-rate is $88.8B. The ladder’s starting mix is solved to reproduce Anthropic’s revenue per slot at that exit. The filing gives no Q4 guidance, so this is our estimate.

The spend curve and the 2028 fit

Every case moves from the 2026 exit to its 2036 endpoint on one curve, steep early and flattening. One parameter sets the speed: the midpoint year, by which half the move is done. For Base it is 2029.17. We did not derive that from evidence. We fitted it so 2028 revenue lands near $195B, the middle of a reported company projection of $190B to $200B.¹⁸ The other cases do not pass through it.

Compute: the larger of demand and the contracted floor

Each year’s compute cost is the larger of demand (serving plus a training budget) and the floor, the cost of capacity already committed. If revenue is too low to use that capacity, the company still pays for it. The floor comes from a deal-by-deal schedule: $79.7B in FY2027, peaking at $133.4B in FY2030 and running off to $22.3B in FY2036. We count 80% of contingent commitments, all committed items with no fixed price, none of the SpaceXAI amounts either side can cancel, and the full firm residual. In Base the floor binds in two of ten years and costs $1.66 a share. In the weak cases it binds for four to six years, which is why they lose so much value.

Serving scales with usage: by 2036, 12% of frontier-tier revenue, 37% of cheaper-tier revenue and 26% of subscription revenue, or 23.8% of Base revenue. Training starts as fixed dollars, $18.7B annualised in 2026, doubling in FY2027 to $37.3B and growing 60% in FY2028 to $59.7B, then moves to a share of revenue, 20% by FY2036. Together they leave a margin after all compute of 56%, against 42% to 43% today.

Fees, other costs and stock pay

Channel fees are a load on consumption revenue: 15.8% at the start, 11.5% by 2036 in Base. They probably include partner-delivered compute and are not a pure distribution cut. Other costs in 2036 are 18% of Base revenue, including stock-based compensation at 6%. The 2036 cost stack in Base is serving 23.8%, training 20.0%, sales and marketing including fees 12.5% and other costs 18.0%, leaving an operating margin after stock pay of 25.8%.

Tax and free cash flow

Long-run tax is 23%, after using $7.5B of federal losses and about $1.0B of R&D credits. Equity awards are deductible when they settle, worth about $36.6B of equity. That depends on each case’s own share price, so the model iterates until each case’s settlement price equals its own value. Free cash flow is operating profit after tax less capital spending (2% of revenue) and working capital (2% of the change in revenue), plus depreciation. Capital spending is light because chip leases sit in compute cost.

The discount rate

The 9.69% is a 10-year Treasury yield of 4.94% plus a beta of 1.0 times a risk premium of 4.75%.¹⁹ Debt is nil, so cost of equity is also cost of capital. The beta of 1.0 is our judgement. Bottom-up peer estimates were about 1.3 to 1.4, so treat it as lenient. Restoring 12.07% for FY2027 to FY2029, fading to 9.69% by FY2036, takes Base to $73.92. A flat 10.19% gives $76.85.

Terminal value

Terminal value is 67% of Base enterprise value. From FY2037 to FY2046, growth fades from 9.0% to 3.5% over five years and then holds, with the margin held. Net reinvestment is the change in revenue divided by a sales-to-capital ratio of 3. From FY2047, growth is 3.5% and new capital earns only the cost of equity. The present value of FY2027 to FY2036 cash flow is $418.3B, FY2037 to FY2046 $453.7B and everything after $403.6B, giving enterprise value of $1,275.6B. Terminal value at the end of 2036 is $2,161.7B, 11.4 times FY2036 operating profit.

Free cash flow falls 30% in FY2047, from $216.4B to $151.7B. That is the convergence assumption, not a forecast. The sales-to-capital rule also implies a return of about 60% after tax on new capital in FY2037 to FY2046.²⁰ We chose it because growth to FY2036 is already paid for in the income statement, so the older growth-over-return formula charged twice. It is also one of the model’s generous choices.

Equity bridge and dilution

Enterprise value plus cash, short-term investments and restricted cash ($134.4B), Q4 2026 cash flow ($0.7B) and dated option proceeds and other items ($4.9B) gives Base equity of $1,415.6B.

The S-1/A leaves its fully diluted table blank, so we built the count. The claims on equity are common 2,199M, preferred as converted 8,559M, Amazon notes as shares 1,840M, options and restricted stock units 2,824M plus 125M granted since, donation-matching shares issued (28M) and expected (453M), and the initial Public Benefit issuance 847M. That is 16,876M. The Public Benefit mechanism adds 2% for each 50% rise in the share price, so Base, with one such rise, uses 17,213M. $1,415.6B over 17,213M is $82.24.

Key assumptions

What moves value

Demand and share dominate. The top rows are the size of the market and Anthropic’s slice, both judgements about 2036. Training is the cost to watch: one point of training share is worth about $2.60 a share, and it is the line a competitor could force up. Treatments are small: dropping the compute floor adds $1.66, capitalising the Broadcom leases as debt lowers value $0.56, and removing future award deductions lowers it $2.13.

Our first terminal method charged reinvestment as growth over return on capital. That produced a cliff: net reinvestment jumped from $4.9B in FY2036 to $62B in FY2037 while the company was still growing. Moving to the current method added $6.63 to Base ($75.61 to $82.24) with no change in operating view. We think it is the fairer treatment, and it also raised the value.

How to read $1.5T

The plan case

The plan case holds Base through 2028, then raises the agent price to 20.7% of wage so the market reaches 6.4% of core wages. Revenue is $423B in 2031 and $793B in 2036 at the same margin, and the value is $87.14 a share. So $1.5T needs about 7.5% more 2036 revenue than Base, with no change in share or margin. At Base margins, a separate calculation says $1.5T needs about $431B of 2031 revenue, 44.8% a year from the Q3 run-rate of $67.8B. For $2T it is about $554B and 52.2%.

Two routes beyond Base

Value above Base comes from a premium-and-share route or a bigger-market route. The order of adding them changes each step’s size.

Either route alone clears $1.5T: premium and share gives $2,095B of equity and the bigger market alone $2,144B. The price talk needs only a part of either.

Base rates

Base needs Anthropic to keep growing above 10% a year through FY2035, after passing $195B in 2028. The only large companies we can name doing something close are Amazon and Nvidia, each riding an unusually large platform shift.²² Cloud history warns the other way: AWS growth fell sharply as it passed $80B of revenue, before re-accelerating. We cannot rule out either path.

Compared with Damodaran

Aswath Damodaran valued Anthropic at $1.475T on 28 October and $1.45T on 3 November.²³ It is the closest public comparison and lands near our Base by different routes.

Restated on our basis, his 28 October spreadsheet is worth about $1,497B, about 6% above our Base: remove the $100B of IPO cash, add the $19.3B of short-term investments he omits, and treat the Amazon notes ($108.6B) as shares. The differences run both ways. He deducts commitments and applies a failure haircut, which lower value, and holds a 15% return on capital and 4.9% terminal growth, which raise it.

His readers are more sceptical than either of us. He describes their collective value as closer to $1T than $1.5T, and the biggest single objection was that his 9.74% cost of capital was far too low. Ours is lower still. At 10.7%, Base would be $72.06.

Where this could be wrong

How the model changed

The model changed a lot, mostly in one direction. Base moved from about $43 a share in an early version to $58.28, then $75.41, and now $82.24. At $58.28 Base was 33% below $1.5T; now it is 5.6%. We did not use the price talk to choose assumptions, and a separate rebuild found nothing in Base solved to $1.5T. Even so, almost every judgement call raised the value, and you should discount for that pattern. The exception came last: taking a share gain out of Volume leader cut the weighted value from $82.85 to $77.42, without changing Base.

• The discount rate began near 11%, moved to a rate fading from about 12% to 9.69%, and settled on a flat 9.69%. Each step after the first raised the value.
• Market size moved from 4.1% of core wages to 5.0%, then 6.0%.
• Three cost lines sit at the favourable end of the research ranges. When they were set, undoing any one lowered Base by $3.75 to $4.59. Training was raised to 20%.
• A share count that double counted restricted stock units and omitted the Public Benefit shares was corrected to 17,213M from about 18,500M.
• Commoditisation was half of Base, so a downside case improved whenever Base did. It is now a fixed 3%.

The weakest assumptions

1. The 2036 market and Anthropic’s share. They drive more value than anything else, and neither is independently confirmed.
2. Terminal reinvestment. A ratio of 3 implies a 60% return on new capital for ten years, then a large step-down. $403.6B of Base’s present value sits after that step.
3. International revenue is scaled from the US, and billing addresses blur where work happens. Using 37.2% rather than 40% lowers Base by $3.42.
4. The 2028 fit. Only Base passes through the reported plan. Volume leader and Premium Frontier reach $239B and $255B.
5. Cost lines are set separately though serving, fees and training are probably linked.
6. Moonshot. $300B of drug revenue is about 25 times the research high.
7. A beta of 1.0 and zero failure probability are both generous.

What is not modelled

• Google’s distribution. Volume leader and Base assume Anthropic holds its share against the incumbents.
• Regulation. Damodaran applies a 5-point margin haircut. We apply none.
• Governance. A 10% discount would give $74.01.
• Supply and obsolescence. Hardware obsolescence and capacity shortfalls are not separate risks.

What would change our mind

FY2027 revenue separates the cases: about $104B to $106B for Plateau and Commoditisation, $120B for Share loss, $128B for Base, $142B for Volume leader and $146B for Premium Frontier. Beyond that, watch Q4 2026 revenue against our $22.2B; the margin after all compute recovering from 41.8%; international revenue rising from 29.4% toward 35% or more by 2028; price reductions growing faster than $600M a quarter; and any disclosure of tier mix or independent share data. In future filings, look for the offering price and share count, Run-Rate Revenue, the fully diluted table, total commitments, the marketplace share of revenue and net revenue retention.

Get the model

The full model is a free Excel download. Start on the Your case tab: change the four yellow cells for AI spend per worker, market share, margin and discount rate, and the valuation updates. Then see what you have to believe.

Download the Anthropic valuation model (.xlsx)

Want to build models like this yourself? Our DCF guide covers the method, and StreetCred’s practice cases score your Excel models cell by cell.

Notes and sources

1. Press reports, not confirmed by the company. About $2T in August 2026 (New York Times) and about $1.5T in October 2026 (Reuters, citing investor-day attendees), both read through AI Weekly’s summaries. The S-1/A leaves the price and deal size blank. August summary October summary S-1/A ↩
2. Our evidence notes, which read outside work including Gartner, the Atlanta Fed, Ramp, the US Census Bureau, the BEA, Acemoglu and the Penn Wharton Budget Model. The 3%, 5% and 9% are our synthesis, not a figure any one source publishes. Atlanta Fed Ramp AI Index Penn Wharton ↩
3. Menlo Ventures, The State of Generative AI in the Enterprise (December 2025). Menlo is an Anthropic investor, and the figures cover enterprise API spend only. Menlo Ventures ↩
4. IQVIA, cited in the S-1/A, p.201. The molecular biology lab is on p.201 and the Coefficient Bio acquisition on p.202. S-1/A ↩
5. BIO, Informa Pharma Intelligence and QLS, Clinical Development Success Rates 2011 to 2020 (2021). The $0.3B to $12B of cash revenue from AI drug programmes is our own research estimate. BIO ↩
6. S-1/A, p.161, on quarterly revenue affected by price reductions. S-1/A ↩
7. Michael Mauboussin and Dan Callahan, The Base Rate Book: Sales Growth (Credit Suisse, 2015). S&P 1500, 1994 to 2014. Base Rate Book ↩
8. Anthropic, PBC, Form S-1/A, Amendment No. 1, filed 2 November 2026, with financials to 30 September 2026. Price, deal size, Run-Rate Revenue and the fully diluted table are blank. Growth of 46.7% on Q2 is our arithmetic. S-1/A ↩
9. S-1/A, pp.10 and 160 (revenue); Note 2 (consumption, subscriptions and geography). S-1/A ↩
10. S-1/A, pp.10, 43, 57, 136 and 144. The Fortune 500 figure covers those with $100,000 or more of trailing revenue, in Q3. Reading the 900% as expansion from a tiny base is our interpretation. S-1/A ↩
11. S-1/A, p.58. The link between fees, distribution and delivery cost is on p.147. S-1/A ↩
12. S-1/A, pp.135, 137 and 138. The all-compute margin is our arithmetic. S-1/A ↩
13. S-1/A, pp.6, 33, 77, 139, 173 and F-26. The floor schedule built from these is our own. S-1/A ↩
14. S-1/A, pp.138 and 163, and the financial statements. S-1/A ↩
15. S-1/A, p.139. S-1/A ↩
16. S-1/A, p.146. The 3% drag is our estimate. S-1/A ↩
17. US Bureau of Labor Statistics, Occupational Employment and Wage Statistics (May 2025), and ILOSTAT, with proxies for some countries. Plus or minus 15% by country. BLS ILOSTAT ↩
18. Reuters, “Anthropic IPO valuation hinges on $190-200 billion 2028 revenue forecast, sources say” (August 2026), citing two people familiar with the company’s finances. Read through a syndicated copy; not confirmed by the company and not filing guidance. Syndicated copy ↩
19. US Treasury 10-year yield, 30 October 2026. The 4.75% premium is about the midpoint of Kroll’s recommended 5.0% and Damodaran’s implied premium, restated to our risk-free rate, of 4.55%. US Treasury Kroll, via BVWire ↩
20. The ratio of 3 is Damodaran’s, from a model that capitalises R&D, and our research put a defensible return on new capital at 12% to 30%. Damodaran, October 2026 ↩
21. Peer shares are inferred from Alphabet and Microsoft annual reports, which do not disclose core markets alone. ↩
22. Named from memory, not from a systematic screen. ↩
23. Aswath Damodaran’s posts of 28 October and 3 November 2026. Year-5 revenue, terminal growth and the 15% return come from his published spreadsheet. His years are not our fiscal years. October post November post ↩

How we made this. One discounted cash flow model of Anthropic, PBC at 31 December 2026, built in Excel. All figures are US dollars, before offering proceeds, per diluted share on our share count, and every model output is our estimate. The analysis used Claude, Anthropic’s AI model, as a drafting and modelling aid. Anthropic is the subject of the analysis. We used only public filings and the sources above and have no inside information.

Disclaimer. This is educational analysis, not investment advice, and not a recommendation about any security. The S-1/A is a preliminary filing and the offering price and size are not set. The base case and the weights are our judgements and may be wrong. Press-reported items have not been verified by us. All model figures will change as new filings arrive.