On November 10 Anthropic told the SEC what it thinks it’s worth. $70 to $78 a share. 825 million shares. It wants to list on Nasdaq as “ANT.” Bloomberg ran the fully diluted math and got about $1.4 trillion.
I may be made a fool, but the idea that Anthropic could be worth $1.4 trillion strikes me as absurd.
It isn’t a bad company. It’s a very good one, and I’m a paying customer. But look at what you’d be buying: a lead over the free models that you measure in months, sitting on infrastructure the company mostly rents. Don’t take my word for either half. Anthropic’s own prospectus says both.
Where the $1.4 trillion comes from
The headlines blur this one, so get the number straight.
Here’s the funny part. The filing never says $1.4 trillion. It’s an amended registration statement, Amendment No. 2 to the S-1, and all it sets is a price range: “between $70.00 and $78.00.” The company’s selling 801,351,354 shares. Existing holders are selling another 23,648,646. And off to the side, seven companies are set to buy about $26 billion of non-voting stock in a private placement. Hang on to that one. It’s the best part of the whole filing.
Bloomberg did the math this way: “At the top of the price range, Anthropic would have a market value of about $1 trillion, based on the shares outstanding listed in the filing. Accounting for employee stock options, restricted shares and stock reserved for the company’s charitable donation matching program, Anthropic would have a fully diluted value of about $1.4 trillion.”
I checked the first half myself. After the offering and the private placement there’ll be 13,086,672,591 shares in the two classes that actually carry economic rights. At $78 that’s about $1.02 trillion. At $70 it’s about $916 billion. Bloomberg’s $1.4 trillion adds the employee stock options, the restricted shares, and the stock reserved for the company’s charitable donation matching program. That’s how you’d count it if you expect those shares to get issued.
So it’s a trillion on the shares outstanding after the offering and $1.4 trillion counting the ones still coming. Take the smaller number if you like. It’s still a trillion dollars, and the argument doesn’t budge.
What a trillion dollars buys
So what does a trillion dollars get you? Here’s the business, straight out of the filing.
Revenue for the first nine months of 2026: $33.2 billion. The same nine months of 2025: $2.5 billion. That’s thirteen times in a year, and it’s a jaw-dropper. Add back the fourth quarter of 2025 and trailing twelve-month revenue through September 30 comes to about $35.3 billion.
Anthropic also reports a number it calls Run-Rate Revenue. Take the last 28 days of usage fees, multiply by 13, then add a month of subscriptions times 12. That number was “approximately $81 billion as of October 31, 2026, up $10 billion from September 30, 2026.” And the company’s own risk factors tell you what it’s worth: “Run-Rate Revenue is not a substitute for revenue prepared in accordance with GAAP or a forecast of revenue.” That’s the company telling you not to treat its own metric as revenue.
Now run the multiples on Bloomberg’s $1.4 trillion:
| Measure | Figure | $1.4T divided by it |
|---|---|---|
| Revenue, trailing twelve months to Sept. 30, 2026 | about $35.3 billion | about 40x |
| Run-Rate Revenue, Oct. 31, 2026 | about $81 billion | about 17x |
Forty times trailing revenue. Seventeen times one month’s activity, annualized. That’s the ask. And it’s a multiple of revenue, before anybody gets to profit.
So where does the money go? The biggest operating expense for those nine months is “Compute and infrastructure”: $18.8 billion on $33.2 billion of revenue. Fifty-seven cents of every customer dollar goes to compute before anybody draws a salary.
Remember that 57 cents. It’s the whole ballgame.
Four months is a release schedule
Epoch AI keeps score on how far the best open-weight models trail the best closed ones. Here’s its latest overall number, published May 29, 2026:
“Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months in the Epoch Capabilities Index (ECI), our aggregate measure of model capability.”
That’s the “four months” I had in my head. Turns out it’s right.
I got something wrong in July, though. In The Model Will Be Free. The Electricity Won’t., I gave you Epoch’s three-month figure as the current one. It wasn’t. Three months was Epoch’s October 2025 number, covering January 2023 through October 2025. The May number is newer, and Epoch says flat out that the gap it found “is slightly larger” than October’s. So between those two readings the gap didn’t shrink. It got a little wider. My mistake.
Epoch also says its four months may flatter the open models. Two factors “mean that our estimate may tend to understate the true gap”: open models tend to do worse than closed models on private benchmarks, plausibly because they’re tuned harder to the public ones, and the leading closed labs don’t always release their best models. Use a stricter test, where the open model has to score strictly higher than the closed model it’s chasing, and the gap “would grow to six months”. On cybersecurity work, Epoch’s October 16 measurement puts the best open model “roughly where the closed frontier stood 6–7 months ago.”
So take the worst number on the page. Seven months. You can lose a seven-month lead between one annual contract and the next. Nobody’s held pricing power for long on a head start that short, and a trillion dollars is what you pay for pricing power that lasts.
Here’s the part that gets me. Anthropic agrees. This is from the risk factors in the same filing:
“If the ability to develop highly capable models becomes more broadly available, our technological differentiation could erode and competition could increasingly depend on access to large-scale compute rather than on model development capability. In such an environment, access to sufficient compute on economical terms could become the primary constraint on our ability to train and serve competitive models . . . .”
That’s my July essay, written by Anthropic’s lawyers and filed with the SEC. Sure, risk factors are written to cover every bad thing under the sun, so don’t read it as a forecast. Read it as the company telling you, in writing, how it could lose.
The same section names the leak, too. Competitors “have attempted in the past and will likely continue to attempt to replicate, distill, extract, or otherwise derive the functionality or capabilities of our models”. The product can be used to train its own replacement. I’ve written about that leaky moat before. Now it’s in a prospectus.
The moat is the building. Anthropic rents the building.
So if the model edge melts down into a fight over compute, the next question asks itself. Who owns the compute?
I can see the answer from my own house. I run open-weight models on hardware at home, and I pay Anthropic $200 a month. The difference between what I get downstairs and what I get for $200 isn’t mostly the model. The best open models are four months back. The real gap is the warehouse full of accelerators on the other end of the line, and the power plant behind it.
So who owns that warehouse?
Last November Anthropic announced “a $50 billion investment in American computing infrastructure, building data centers with Fluidstack in Texas and New York, with more sites to come.” It said the facilities “are custom built for Anthropic”. Beyond saying it picked Fluidstack for “rapid delivery of gigawatts of power”, the announcement goes no further. Not a word about who owns the land, the buildings, or the power. Not a word about who runs them.
The prospectus fills in the blanks. Here’s how Anthropic describes its own data center strategy:
“[W]e also procure specialized accelerators directly and operate this hardware in dedicated data center facilities leased and operated on our behalf, to our specification.”
The chips? They come through “master equipment leases,” for “deployment in dedicated data center facilities operated by others on our behalf.” And the buildings:
“In certain of our data center infrastructure arrangements, third-party partners procure the underlying real estate, power, and data center facilities, while we lease capacity in those facilities and contract with specialized third-party operators for ongoing facility management. For example, we have entered into a colocation and services agreement with Fluidstack USA I Inc. (‘Fluidstack’), under which Fluidstack provides data center development and operations services to us.”
Leased chips. Leased capacity. A partner that lines up the real estate and the power. A third party that runs the place. Under the facilities heading, the filing boils it down to one line: “We also lease data centers in the United States”. Now, the filing doesn’t tie those paragraphs to the $50 billion by name. But Fluidstack is the partner the announcement named, and the filing says what Fluidstack does: development and operations. Anthropic leases the capacity. That’s a tenant.
Then there’s the bill, and it’s a whopper. As of and after September 30, Anthropic has signed purchase commitments and leases with “an aggregate committed spend of approximately $900.9 billion,” “substantially all of which is related to compute and infrastructure.” About $451.6 billion lands before the end of 2030. The rest runs out to 2045. Here are the terms on some of it:
“[C]ertain of our dedicated power and data center infrastructure arrangements have initial terms of 20 years with annual escalators and include take-or-pay compute payment obligations and uncapped fuel cost pass-throughs that apply for the duration of the term regardless of actual utilization.”
Read that the way you’d read it for a client. Twenty years. The rent goes up every year. You pay whether you use it or not. Fuel costs pass straight through, no cap. Any lawyer would call that a tenant’s lease, and a sweet one for the landlord.
Now put two numbers side by side. Anthropic has promised about $900.9 billion to the companies that supply the scarce layer. On shares outstanding, the whole company’s worth about $1 trillion at the top of the range. Those two numbers sit a lot closer together than anybody pitching the stock would like.
The landlords are buying the tenant’s stock
Back to that private placement. Here’s who’s buying the $26 billion, per the filing: Google, $10 billion. NVIDIA, $10 billion. AMD, $2.5 billion. Broadcom, Micron, and Microsoft, $1 billion each. TSMC, $500 million.
That’s two of the clouds, the three chip companies whose hardware the filing names as Anthropic’s compute platforms or suppliers, and the memory and chipmaking outfits upstream of them. The supply chain’s buying in.
Amazon isn’t on the list. It’s already in deep. By my arithmetic on the filing’s share counts, Amazon will hold about 14.7% of the economic shares after the offering and Google about 13.7%. Amazon also holds convertible notes for another 1,563,617,444 shares that can’t convert yet without pushing it past an ownership cap. And through the first nine months of 2026, the filing says 44% of Anthropic’s revenue came “from sales to customers through CSP marketplaces”, meaning the cloud providers’ storefronts.
So the companies that sell Anthropic its compute also own big chunks of it, sell a big share of its product, and collect its rent. Smart. That’s a hedge with a bow on it. If Anthropic wins, they own the stock. If the models commoditize, they’ve still got the data centers, the chips, and the storefront. I’d love to be on their side of that deal. Paying a trillion dollars for the tenant is a different trade.
The best case for the number
Now the other side, and some of it’s darn strong.
The growth is real, and it’s enormous. $2.5 billion to $33.2 billion, nine months against nine months. Customers paying $100,000 or more a year went from “more than 950 to more than 11,900” between September 2025 and September 2026. Run-rate jumped $10 billion in one month. I’ve never seen numbers like that from a company this size.
It’s nearly breaking even while it grows. The operating loss for the nine months was $1.3 billion, down from $5.2 billion a year earlier. The company reports a non-GAAP “Adjusted Operating Income” of $4.1 billion for the period. (The GAAP net loss is a lot bigger, about $104 billion, but nearly all of that is a non-cash change in the fair value of convertible notes and similar instruments. It’s not money out the door, and I’m not going to pretend it is.)
The product is more than a model. Enterprises don’t buy weights. They buy a vendor that’s on all three big clouds, signs the contracts, passes the security reviews, and ships the coding tools their engineers already live in. Nobody ever got a sales call from an open-weight model. Sure, switching models in a well-built system takes one line of configuration, like I said in July. Most enterprise systems aren’t that well built, and inertia’s worth a lot of money.
And it’s reaching for more of the stack. The filing says “[a]n increasing portion” of its compute spending “relates to building out our own compute and infrastructure and vertically integrating across the compute supply chain”, and that Anthropic has “also executed data center leases directly and expect[s] to do so increasingly over time.” If that turns into owning land and power, the moat starts to belong to Anthropic.
Here’s why none of that gets me to $1.4 trillion.
The growth is real. The company itself says it won’t last at this pace: “We expect our revenue growth rate to decline over time, and it may decline significantly and more rapidly than we anticipate as our revenue base increases and as the market matures.” The product’s real, too, but it sits on a model edge the company admits could erode, and on compute it doesn’t own. And “vertically integrating”, as the filing describes it so far, runs through leases: leased chips, leased facilities, facilities operated on Anthropic’s behalf. That still isn’t owning the building. A longer lease locks in the rent. It doesn’t make you the landlord.
Boil it down and $1.4 trillion is a bet that three things hold at once: a four-month lead, a 57-cent compute bill that doesn’t eat the margin, and landlords who keep giving good terms to a tenant they also compete with. The filing itself describes Amazon, Google, and Microsoft as companies that “simultaneously serve as our investors, customers, distribution partners, and competitors”. Before I paid forty times revenue, I’d want one of those three to be a sure thing. Not one of them is.
Here’s how I’d know I blew this one
I may be made a fool. Here’s what it would take.
- The gap blows out. If two years from now the best open-weight models are more than a year behind on Epoch’s measure, the lead’s a moat and I called it wrong. And the last move already went against me: three months to four, October 2025 to May 2026. Two more readings like that and I’d have to take it seriously.
- Anthropic becomes its own landlord. If the “Anthropic Direct” buildout turns into owned land, owned buildings, and owned power, with fewer partners in the middle, then it owns the scarce layer and my renting argument goes out the window.
- Compute gets cheap. If the industry overbuilds and there’s a glut, the landlords lose their pricing power, Anthropic’s 57 cents shrinks, and the value moves to whoever owns the customer. That’s Anthropic’s strongest suit. A glut hurts the owners and helps the tenant.
- The run-rate keeps doing what it did in October. If Anthropic adds $10 billion of run-rate every month for another year, it’s past $200 billion, and $1.4 trillion starts looking like a bargain. I don’t think that happens. I’ve been wrong before.
What this means for your firm
None of this tells you whether to buy the stock. I’m a lawyer, not your broker, and this isn’t investment advice. It’s about the subscription you’re already paying for.
Pay for the best tool while it’s the best deal. I do. The frontier still earns its keep on the narrow band I described in July: hard reasoning, new arguments, a long and difficult record.
Just don’t build your practice on the idea that any one vendor’s lead lasts. A four-month edge is a reason to subscribe. It’s not a reason to marry the vendor. If the price goes up, the terms change, or somebody else’s model passes it, you ought to be able to switch in an afternoon. That means your prompts, your templates, your document store, and your evaluation set live in plain files you control, and the model sits behind one setting you can change.
And keep the exit open by walking through it now and then. Run something locally, even if it’s only document sorting and date extraction, so you know what the free stuff can do on your own work. I wrote up what that costs and the stack I’d install, and Apple has since put a price tag on the local option. Nobody’s saying quit paying for the frontier. Just know what you’d do the day you have to.
That’s my setup. I pay the $200 and I run open weights downstairs. If Anthropic’s still four months ahead next year, I’ll keep paying. If it isn’t, I’ll quit. Either way, I’m not the one who signed a twenty-year take-or-pay deal.
Sources
- Anthropic, PBC, Amendment No. 2 to Form S-1 Registration Statement (filed Nov. 10, 2026; SEC CIK 0002130357), retrieved from EDGAR Nov. 10, 2026 — cover page (825,000,000 shares; 801,351,354 by the company and 23,648,646 by selling stockholders; $70.00–$78.00 estimated range; Nasdaq symbol “ANT”); “The Offering” (7,427,991,791 Class A and 5,658,680,800 Class N shares outstanding after the offering and concurrent private placement; concurrent private placement purchasers and amounts); summary consolidated statements of operations (revenue of $4,585 million for 2025, $2,545 million and $33,242 million for the nine months ended Sept. 30, 2025 and 2026; compute and infrastructure expense of $18,808 million; operating loss of $1,260 million; net loss of $104,039 million and the change in fair value of convertible notes, derivative liabilities, and SAFEs of $(94,324) million); definition of Run-Rate Revenue and the $81 billion figure as of Oct. 31, 2026; the risk factor stating that Run-Rate Revenue “is not a substitute” for GAAP revenue; the competition risk factors on distillation, eroding differentiation, and compute access; the strategic-partner risk factor describing Amazon, Google, and Microsoft as “investors, customers, distribution partners, and competitors”; “Compute Investments” ($900.9 billion aggregate committed spend; $451.6 billion before Dec. 31, 2030; vertical integration language; Anthropic Direct); “Our Data Center Strategy” (facilities “leased and operated on our behalf”; master equipment leases; the Fluidstack colocation and services agreement); the 20-year take-or-pay risk factor; the Facilities section (“We also lease data centers in the United States”); share of revenue through CSP marketplaces (44% for the nine months ended Sept. 30, 2026); customers with $100,000 or more of TTM revenue; Amazon and Google Class N holdings and the remaining Amazon convertible note shares; the revenue-growth-rate risk statement
- Bloomberg News, Anthropic targets $1.4 tn valuation in IPO, eyes $90.4 billion fundraise, as republished by Business Standard (Nov. 10, 2026) — the “about $1 trillion” market value on shares outstanding and the “about $1.4 trillion” fully diluted value
- Anthropic, Anthropic invests $50 billion in American AI infrastructure (Nov. 12, 2025) — the $50 billion announcement, Fluidstack, Texas and New York, “custom built for Anthropic”
- Anthropic, Claude pricing (retrieved Nov. 10, 2026) — consumer plan pricing, including the $200-per-month tier
- Epoch AI, Jack Edwards and Luke Emberson, Open models lag state-of-the-art closed models by 4 months (May 29, 2026) — four-month average lag Jan. 1–May 28, 2026; 8-point ECI gap; “slightly larger” than the October 2025 finding; six months under the strict criterion; the two factors that may understate the gap
- Epoch AI, Luke Emberson, Open-weight models lag state-of-the-art by around 3 months on average (Oct. 30, 2025) — the earlier three-month measurement
- Epoch AI, Jaeho Lee and Alexander Barry, Open-weight models trail closed models in cyber capabilities by 6–7 months (Oct. 16, 2026) — the cyber-specific lag
- Arithmetic by the author from the S-1/A figures above: economic shares after the offering (13,086,672,591) times $70 and $78; trailing twelve-month revenue ($4,585M + $33,242M − $2,545M = $35,282M); $1.4 trillion divided by $35.3 billion and by $81 billion; $18,808M ÷ $33,242M; Amazon (1,920,511,218) and Google (1,789,598,269) Class N shares after the offering as a share of 13,086,672,591
- Companion essays in this section: Running Models on Your Own Hardware (June 8, 2026), The Model Will Be Free. The Electricity Won’t. (July 27, 2026), Run It Yourself, Then Let the Client Choose (July 27, 2026), Protectionism in a Safety Costume (Aug. 20, 2026), Apple Just Priced the Local Thesis (Sept. 24, 2026), and How Do You Know It Works?
Commentary on technology and the business of law. Not legal advice, and not investment advice: nothing here is a recommendation to buy, sell, or hold any security. Figures are as stated in Anthropic’s preliminary prospectus of November 10, 2026, which is subject to completion and may change before the offering prices. The opinions and stated failure conditions are the author’s. No client information appears in this article.