The $900 Billion Prepayment

Anthropic may become one of the largest companies ever built. That still does not make it worth $2 trillion today.

Subscribe
MarketsAIStartups
Listen to this post
AI Summary Anthropic is preparing for an IPO that could value the company at more than $2 trillion, but by July 2026 it had reached an annualized revenue run rate of approximately $65 billion, making it potentially a $100 billion company rather than the $4.6 billion in 2025 actual revenue. …
  • Anthropic is preparing for an IPO that could value the company at more than $2 trillion, but by July 2026 it had reached an annualized revenue run rate of approximately $65 billion, making it potentially a $100 billion company rather than the $4.6 billion in 2025 actual revenue.
  • Unlike traditional software companies, Anthropic must continuously reinvest enormous capital to maintain its competitive position because frontier AI models depreciate intellectually as competitors advance and capabilities that command premium pricing today become baseline expectations tomorrow, with the company listing approximately $518 billion in future cloud, computing and infrastructure obligations.
  • Using a model that assumes Anthropic succeeds spectacularly—growing to more than $600 billion in annual revenue by 2036 with a 27% free cash flow margin—the author calculates a reasonable current equity valuation of approximately $1.1 trillion, not the $2 trillion IPO price.
  • At $2 trillion, investors would be paying roughly $900 billion beyond what extraordinary execution justifies, effectively prepaying for success that requires Anthropic to become one of the largest businesses ever created while retaining strong pricing power even as the marginal cost of intelligence collapses and competitors try to commoditize its core product.

Anthropic is preparing for an IPO that could value the five-year-old artificial intelligence company at more than $2 trillion.1 The temptation is to attack that number with the company's historical financials.

In 2025, Anthropic generated just under $4.6 billion in revenue, spent $7.33 billion on compute and infrastructure, recorded an $8.06 billion operating loss, and reported a headline net loss approaching $42 billion.1 Against those numbers, $2 trillion looks almost absurd.

But that is the wrong analysis.

Roughly $34 billion of the $42 billion net loss was an accounting charge tied primarily to financing instruments whose value rose along with Anthropic's valuation. It was not $34 billion disappearing into GPU clusters.1

More importantly, Anthropic is growing so quickly that its 2025 revenue is already largely irrelevant to a forward valuation. By July 2026, the company had reached an annualized revenue run rate of approximately $65 billion. It has projected revenue approaching $200 billion in 2028, and Reuters Breakingviews calculates that a $100 billion run rate could be within reach during 2026.23

Important

Anthropic is not a $4.6 billion company being valued at $2 trillion. It is potentially a $100 billion company arriving extraordinarily quickly.

That makes the valuation considerably more interesting. It does not make it correct.

The question isn't whether Anthropic can become enormous. It is how much of that enormous future an investor should be required to pay for today.

Start With the Number Underneath the Number

At a $2 trillion valuation, Anthropic's multiple depends entirely on which revenue number you choose.

Reference point Revenue Implied $2T multiple
2025 actual $4.6B 435×
July 2026 annualized run rate $65B 30.8×
Potential 2026 exit run rate $100B 20×
Anthropic 2028 projection ~$200B 10×

The first number is sensational but largely useless. The last one is far more instructive: investors considering Anthropic at $2 trillion are effectively being asked to pay about ten times revenue that management does not expect to reach until 2028. That is not unprecedented for an exceptional software company.

But Anthropic is not economically structured like conventional software, and that distinction matters enormously.

Claude Is Not an Annuity

In evaluating SpaceX, I argued that a satellite constellation should not be valued as though today's deployed infrastructure simply stays in orbit indefinitely, collecting recurring revenue.7 Satellites expire and must be replaced. Capital expenditure is not incidental to that business; it is part of the engine required to keep the revenue alive.

Frontier artificial intelligence has a surprisingly similar problem.

Claude Opus 5.5 may be enormously valuable today, but Anthropic cannot train Claude once and collect software rents from it for the next fifteen years. Claude 5.5 will become Claude 6, then 7, then 8. Competitors will advance. Architectures, hardware and training methods will change. Context windows will expand and inference economics will collapse. Capabilities that command premium pricing today will become baseline expectations tomorrow.

Anthropic IPO Valuation

The frontier model is therefore an economically depreciating asset, even though that depreciation never appears on a balance sheet the way a satellite's or a semiconductor fab's does. Maintaining the frontier requires continuous reinvestment. Anthropic spent $7.33 billion on compute and infrastructure in 2025 alone, more than half of its $12.65 billion in total operating expenses.1

The scale of what comes next is extraordinary.

Anthropic's prospectus lists approximately $518 billion in future cloud, computing and infrastructure obligations.1 Reuters Breakingviews separately identified roughly $100 billion of Amazon cloud commitments, a reported $200 billion Google infrastructure and chip arrangement, and another $45 billion involving Nscale, among others.3

These are not the economics of a traditional SaaS business. They are closer to a hybrid of software platform, semiconductor company, hyperscaler and research laboratory.

A frontier model may look like software to the customer while behaving like industrial infrastructure to the owner.

That changes the multiple it deserves.

Revenue Growth Is Not the Same Thing as Economic Rent

There is a remarkably bullish argument for Anthropic. The price of intelligence will collapse. Inference efficiency will improve dramatically, hardware will become more capable, models will become more efficient, and the amount of intelligence the economy consumes could grow even faster. That is entirely plausible.

But those economics can resolve in two very different ways. Imagine that delivering a unit of intelligence costs Anthropic 30 cents today and eventually costs three cents. If Anthropic keeps charging a dollar, margins become spectacular. Competitive markets, however, rarely let producers keep every productivity gain. The price might instead move:

flowchart LR
    A["$1.00"] --> B["$0.30"]
    B --> C["$0.10"]
    C --> D["$0.05"]

The consumer captures most of the efficiency gain. At a nickel against a three-cent cost, Anthropic still earns a margin, just a thin one. Consumption can still explode, revenue can still become enormous, and the infrastructure can remain fully utilized, yet the producer earns far less economic rent than the original valuation assumed.

This is one of the recurring errors in technology valuation: demand is confused with pricing power.8 Compute demand can grow by orders of magnitude without guaranteeing that the company selling intelligence earns monopoly-like margins on it. Chinese models, open-weight models, Google, OpenAI, Meta, xAI and future entrants all press on the same economics.

Important

Anthropic does not merely need intelligence consumption to explode. It needs enough differentiation to keep falling production costs from being competed away.

The Customer Base Is Not Yet an Annuity Either

The same problem exists on the revenue side. Anthropic disclosed that nearly one-quarter of its 2025 revenue came from just two customers, and warned that many large customers are not bound by long-term commitments and can reduce or terminate spending.1

During hypergrowth, this can be almost invisible.

When revenue expands from $5 billion to $25 billion to $65 billion to $100 billion, churn is overwhelmed by new demand. At $200 billion or $300 billion, the mathematics change. The valuation eventually depends less on how quickly companies experiment with Claude and more on the mundane characteristics that have always determined enterprise value:

  • retention
  • switching costs
  • pricing power
  • gross margins
  • free cash flow
  • durability

Hypergrowth can postpone those questions. It cannot eliminate them.9

What I Think Anthropic Is Actually Worth

So rather than valuing Anthropic on trailing revenue, I built a model around an outcome in which Anthropic succeeds — not modestly, but spectacularly.

Year Revenue Free cash flow margin
2027 $125B −5%
2028 $195B 4%
2029 $260B 10%
2030 $325B 16%
2031 $385B 20%
2033 $490B 25%
2036 $610B 27%

In this base case, Anthropic eventually generates more than $600 billion in annual revenue. For comparison, Alphabet generated approximately $403 billion in 2025. Its operating margin was roughly 32%, but after $91 billion of capital expenditures, its free cash flow margin was closer to 18%.4 This is hardly a bearish model. It assumes Anthropic becomes larger by revenue than Alphabet is today, and that it converts revenue to free cash flow far more efficiently than Alphabet does, at 27%, despite the permanent need to fund model development and enormous inference infrastructure.

Using a 12% discount rate, 3% terminal growth and approximately $100 billion of cash-like value, those assumptions produce an equity valuation of roughly $1.07 trillion. At 11%, the same operating assumptions produce approximately $1.24 trillion; at 12.5%, roughly $1.0 trillion.6

The discount rate matters, especially because more than half of that $1.07 trillion comes from cash flows beyond 2036. On September 28, the U.S. 10-year Treasury yield was approximately 5.25%.5 When investors can earn more than 5% on government debt, assigning an 8% or 9% discount rate to an unprofitable frontier AI laboratory facing enormous infrastructure commitments, technological obsolescence, customer concentration, and intense competition would require extremely generous assumptions.10

That produces a reasonable current valuation range of roughly:

Scenario Estimated equity value
Commoditization / slower growth $400–600B
Exceptional execution (base case) $800B–$1.3T
Near-flawless execution $1.4–1.7T
Near-perfect bull case $2T+

My midpoint is approximately $1.1 trillion — not terribly far from Anthropic's $965 billion private valuation in May 2026.2 That private valuation already embeds extraordinary success. What it does not do is make shareholders pay for that success twice, which is roughly what a $2 trillion price asks.

So What Does $2 Trillion Require?

Don't ask whether $2 trillion sounds ridiculous. Ask what must become true for it to be reasonable. To generate roughly $2 trillion of present value without an unusually low discount rate, Anthropic needs something resembling:

Metric Approximate bull case
2028 revenue ~$220B
2030 revenue ~$400B
2036 revenue $800B+
Mature FCF margin 30–32%
Discount rate ~10–10.5%

Under assumptions around that level, the mathematics can reach $2 trillion. So $2 trillion is not impossible. But consider what investors have now assumed. Anthropic must become one of the largest businesses ever created. It must retain strong pricing power as the marginal cost of intelligence collapses, and translate a compute-intensive model into free cash flow margins approaching those of the world's best mature technology companies.

It must keep growing after revenue reaches hundreds of billions of dollars, manage hundreds of billions in infrastructure obligations, and maintain a technological lead while some of the richest companies in history try to commoditize its core product.

That is no longer merely a bet that Anthropic succeeds. It is a bet that almost everything important goes right.

At $1 trillion, investors are paying for extraordinary execution. At $2 trillion, they are beginning to prepay for the extraordinary execution after that.

The $900 Billion Prepayment

The useful comparison is not $4.6 billion of revenue against a $2 trillion valuation. That makes a better headline than analysis. The real comparison is roughly $1.1 trillion of value under a model that already assumes extraordinary success, against roughly $2 trillion of market capitalization. The difference — about $900 billion — is optionality the prospective IPO investor is being asked to buy before it has materialized.

The same mistake appears whenever extraordinary technology companies reach extraordinary valuations. The underlying business can be real. The growth can be real. The technology can be transformational, and the founders can execute brilliantly.

And the stock can still be overpriced.

Those ideas are not contradictory. The more impressive the company becomes, the easier it is to forget the distinction.

Intelligence Has a Depreciation Schedule

The most important feature of Anthropic's economics may be one that accounting statements barely capture. Buildings, satellites, semiconductor fabs and trucks all depreciate, and we know how to model them because their useful lives are visible. Frontier intelligence depreciates too. It simply depreciates intellectually rather than physically.

A model that represents billions of dollars of training, research and infrastructure today can become commercially ordinary remarkably quickly. And unlike a traditional software product, staying competitive may require another enormous capital cycle almost immediately. Anthropic could become one of the most important companies in the world while running one of the fastest replacement cycles ever created.

That does not make Anthropic a bad business. It makes valuation discipline more important. The value of Anthropic is not determined by how much intelligence Claude can produce. It is determined by how much durable free cash flow Anthropic can retain after continually buying the right to remain at the frontier.

On the evidence available today, that business is worth approximately $1.1 trillion. That is already an extraordinary valuation for an extraordinary company.

The remaining $900 billion still has to be earned.

Footnotes

  1. Exclusive: Anthropic's IPO Prospectus Shows Sweeping AI Vision, Surging Costs — Reuters, September 28, 2026 — Every 2025 figure in this piece comes from the prospectus as reported here: revenue up twelvefold to just under $4.6 billion; an operating loss above $8 billion excluding writedowns of liabilities mostly tied to previous fundraising; $7.33 billion of compute and infrastructure spending, roughly three times the prior year; $20.28 billion of cash and short-term investments at year end; nearly a quarter of revenue from two customers, with large customers not bound by long-term commitments; and the qualification this piece leans on, that roughly $34 billion of the $42 billion headline net loss was an accounting charge tied to financing instruments convertible into shares rather than cash spent on compute. Worth stating plainly: the prospectus is a draft registration statement submitted confidentially on June 1, 2026, and no public Form S-1 has appeared on EDGAR. These figures are Reuters reporting a document it was shown, not a filing any reader can open and check. Reuters describes the $518 billion of cloud, computing and infrastructure obligations as falling in the coming year rather than across an open-ended future. Linked through CNBC's carriage of the Reuters exclusive. https://www.cnbc.com/2026/09/28/anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-reuters.html ↩
  2. Anthropic's Path From AI Startup to Industry-Defining IPO — Reuters, September 28, 2026 — The growth figures the forward valuation actually rests on: $65 billion raised in May 2026 at an estimated $965 billion valuation, an annualized revenue run rate of approximately $65 billion by July 2026, and a projection of as much as $200 billion of revenue in 2028. The May valuation is the anchor this piece's midpoint is measured against. Reuters also reports the listing is likely to be pushed past the November midterm elections. Linked through a syndication carrying the full Reuters wire piece. https://kfgo.com/2026/09/28/anthropics-path-from-ai-startup-to-industry-defining-ipo/ ↩
  3. Anthropic's IPO Is a Crucial AI Gut-Check — Reuters Breakingviews, September 23, 2026 — The source for the $100 billion run rate being within reach during 2026, and for the composition of the infrastructure commitments beyond the prospectus total: roughly $100 billion with Amazon, a reported $200 billion Google infrastructure and chip arrangement, and about $45 billion involving Nscale, among others. The counterparty concentration matters as much as the total. https://www.breakingviews.com/columns/considered-view/anthropics-ipo-is-crucial-ai-gut-check-2026-09-23/ ↩
  4. Alphabet Inc., Form 10-K for fiscal year 2025 — The comparison that makes the base case a bull case rather than a bear one. Alphabet reported $402.836 billion of 2025 revenue and $129.039 billion of operating income, an operating margin of about 32%. Operating cash flow of $164.71 billion less $91.45 billion of capital expenditure left $73.27 billion of free cash flow — a margin near 18%. The model in this piece assumes Anthropic grows past that revenue and converts at 27%, half again as efficiently, while funding a frontier research program. Note the asymmetry in what can be checked: Alphabet's figures are in a filed annual report, while Anthropic's are in a document the public cannot yet read. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm ↩
  5. U.S. 10-year Treasury yield, September 28, 2026 — The benchmark 10-year Treasury yield was approximately 5.24% late in the September 28 session, after moving above 5.25% intraday and reaching levels not seen since June 2007. Reuters reported that the 10-year had already reached a 19-year high of 5.2251% during the preceding week and remained above 5% as the bond selloff continued. At that level, the risk-free rate becomes a material constraint on long-duration equity valuations: projected future cash flows must clear a substantially higher discount-rate hurdle before generating comparable present values. ↩
  6. Valuation methodology and model assumptions — The valuation ranges are the author’s scenario estimates, not Anthropic guidance. The base-case DCF assumes roughly $610B in 2036 revenue, a 27% FCF margin, 12% discount rate, 3% terminal growth, and $100B of cash-like value. Given Anthropic’s substantial infrastructure commitments and uncertainty around financing, dilution, and long-run capital needs, these figures should be read as scenario analysis, not forecasts. ↩
  7. A Constellation Is Not an Annuity — The argument this paragraph is referring back to. Fiber is dug once and collects rent for thirty years; a satellite constellation has to rebuild its entire revenue-generating base every five, which makes capital expenditure part of the engine rather than a line beneath it. The claim here is that a frontier model sits closer to the constellation than to the fiber. ↩
  8. Tokens Are the New Clicks — The same confusion examined from the demand side rather than the valuation side: how much of agentic token growth is real consumption and how much is coordination overhead, and what that does to a capex thesis built on the volume. Rising units are not by themselves evidence of a business that captures the value of those units. ↩
  9. Solvent, But Not on Schedule — The timing version of this problem, written before the prospectus existed: the system can afford the losses but perhaps not the timeline, and the compression that reset private software marks was already climbing toward the trillion-dollar listings. This is one of those listings. ↩
  10. The Treasury Market May Not Be Breaking — It May Be Searching for a New Price — Why the 5.25% in the sentence above should be read as a level rather than a spike. That piece argues the clearing yield is structurally higher because government borrowing, AI infrastructure and equities are all bidding for the same savings — which means the discount rate applied here is not a temporary inconvenience a valuation can wait out. The same AI buildout appears on both sides: as the thing being valued, and as one of the claimants pushing up the rate it is discounted at. ↩
Back to the Journal