Anthropic’s $2 Trillion IPO: Does 31x Revenue Make Sense?
Anthropic’s $2 Trillion IPO: Does 31x Revenue Make Sense?
This piece builds on a conversation I had with Neil Borate on the Thefynprint podcast on 8 October, where we spent most of our time on the Anthropic IPO before turning to personal AI agents and the Global AI portfolio. If you would rather listen than read, you can watch the full episode above. What follows takes the IPO part of that conversation further, with the numbers set out in sequence and the questions an Indian investor should be asking before the listing.
The first pure-play listing
In the last week of September, the draft prospectus that Anthropic had filed confidentially with the SEC on 1 June found its way into the hands of investors and then the press, and within days the two numbers everyone repeated were a run-rate of roughly $65 billion and an expected valuation of close to $2 trillion. Put one over the other and you get about 31 times revenue, which is the figure that set off most of the arguments on social media, and it was the first question Neil put to me when we sat down to record.
Before getting to that multiple, it is worth being clear about why this listing matters at all. Investors have been able to buy AI for years through Alphabet, Microsoft, Amazon or Meta, but in each of those cases the AI business sits inside a much larger company with search, cloud, retail or advertising revenue around it. A handful of model companies have listed in Hong Kong this year, and they are interesting in their own right, but Anthropic would be the first American frontier lab to come to the public market, and the first time a retail investor anywhere could own a model company whose entire revenue comes from the model. That is why the leaked filing has been read so closely, and why the multiple has drawn so much heat.
Does 31x revenue make sense?
A multiple on its own tells you very little about a company growing at this pace, so the more useful exercise is to walk through how Anthropic got here. The company was founded in January 2021 and only began earning meaningful revenue in 2023, when it brought in a few tens of millions of dollars. In 2024, according to the leaked filing, revenue was $386 million. In 2025 it was $4.6 billion, roughly twelve times the year before. Then, in the first half of 2026, the numbers stopped looking like a startup’s: $4.7 billion in the first quarter and $11.5 billion in the second, with the annualised run-rate moving from about $9 billion at the start of the year to $47 billion by May and around $65 billion by late July.
Anthropic’s reported annual revenue, 2024 to 2025
Seen in that sequence, a 31x multiple is a snapshot of a figure that has been moving every few weeks, and it was probably already stale by the time our episode went out. If Anthropic reaches the roughly $100 billion run-rate some investors expect by the time it lists, which is reportedly targeted for the weeks before Thanksgiving in late November, the same $2 trillion valuation works out to about 20 times revenue. Against the history of venture-backed companies growing anywhere near ten times a year, that is a demanding price but not an absurd one.
Neil’s objection on the podcast was the right one, and it is the one most readers of Thefynprint will share: investors value companies on earnings, and a revenue multiple says nothing about how much of that $65 billion will ever turn into profit. The honest answer is that a P/E ratio is not yet a useful tool here. In 2025 Anthropic made an operating loss of about $8 billion on its $4.6 billion of revenue, because it was spending heavily on the compute it would need for the following years. The headline net loss of about $42 billion that circulated after the leak is mostly an accounting charge on convertible financing, which grew in value as the company’s valuation rose, so it tells you more about the price of earlier funding rounds than about the business itself. The more important line in the filing is that Anthropic turned an operating profit in the second quarter of 2026, on that $11.5 billion of revenue. One profitable quarter does not make a durable business, and the public S-1 and the next two or three quarterly reports will tell us far more, but the direction of travel has changed, and that is what a revenue multiple is ultimately betting on.
Revenue, run-rate and profit are three different measures
The $518 billion bet on compute
The number in the filing that deserves the most attention is not the valuation but the commitments. Anthropic has signed up for at least $518 billion of compute over roughly the next ten years, spread across six partners: about $111 billion with Google, $110 billion with Amazon, around $31 billion with Microsoft, up to $84.5 billion of Nvidia-based capacity with xAI (now part of SpaceX), and large equipment arrangements tied to Broadcom, with AMD supplying chips while also agreeing to buy up to $5 billion of Anthropic stock. Averaged over a decade, that is around $50 billion of compute a year, which is roughly what Anthropic’s second-quarter revenue annualizes to today.
What buying compute means: $518 billion of future commitments
Anyone who has run a business will recognise the shape of this. A frontier lab cannot rent compute the way a startup rents office space; it has to reserve data-center capacity, and the chips inside it, years in advance, because the hyperscalers will only build if someone has committed to pay. About 80% of Anthropic’s commitments are non-cancellable or must be paid regardless of usage, which means that if demand slows, Anthropic still pays Google, Amazon and the rest. Its customers, meanwhile, are largely enterprises, many of which can scale back or switch providers without breaking a contract. That is a fixed cost base sitting on top of a more flexible revenue base, and the whole structure depends on revenue continuing to grow faster than the obligations, which so far it has.
It also helps to ask, as Neil did, who is actually paying Anthropic. The answer is mostly companies rather than consumers: almost every large enterprise, in the US and well beyond it, now buys models from Anthropic, OpenAI or both, and between them the two labs are running at something like $110 billion of annualised revenue by my estimate. Open-weight models from China and the US are taking a growing share, and Google earns its part through Gemini, but for now perhaps 70 to 80% of model revenue still flows to those two companies. Most of the rest of the AI industry, from chip designers to data-center landlords to power utilities, is in effect a supplier to them, which is why the commitments in this filing matter well beyond Anthropic’s own shareholders. For Anthropic’s equity they are the biggest single risk in the prospectus; for the companies on the other side of those contracts they are close to guaranteed revenue for the rest of the decade.
Eighty pages of risk
The detail that has surprised most first-time readers of the filing is how much of it is about danger. Roughly 80 of its 261 pages are risk factors, and among them is the possibility that advanced AI could pose a catastrophic or existential risk to humanity. Pranav raised this in the audience questions at the end of our episode, asking how an investor is supposed to price something like that.
Part of the answer is legal and part of it is structural. Anthropic is a public benefit corporation, which means its directors answer to a stated public mission as well as to shareholders, and its leadership has talked about the risks of AI in public for years; leaving those risks out of a securities filing would have created far more legal exposure than putting them in. The more practical part of the answer is that the risk shows up in the business long before it shows up as catastrophe. A growing share of compute and research is going into alignment, monitoring and safety work, which raises costs, and a lab that is not confident a model is safe has every commercial reason not to release it, because the first company to ship something genuinely harmful would be the first to suffer for it. Pacing of that kind tends to happen on its own as models become more capable, and if incidents increase, I would expect the labs themselves to slow their releases before anyone forces them to.
Three AI safety concepts behind the risk section
The summer gave investors a preview of how this plays out. A run of incidents in which AI agents behaved outside their developers’ intent, a wave of public warnings from researchers at both Anthropic and OpenAI, and Dario Amodei’s call to “pace the frontier” were followed by a voluntary accord signed at the White House at the end of September, and an FTC inquiry into both labs. There are several readings of the push for regulation, and they are not mutually exclusive: the labs want some protection from liability, the US government sees AI as a race with China and does not want to slow it, the chip and cloud companies have hundreds of billions of capex riding on continued growth, and sceptics argue that rules written by the leaders would raise compliance costs for everyone behind them, which is a form of regulatory capture. An investor does not need to settle that debate to see the practical consequence, which is higher costs, slower release cycles and more scrutiny, all of which sit inside the risk section of the prospectus rather than outside it.
Regulatory capture, a term from the discussion
Abhas asked the related political question, whether the US midterms, or the presidential election two years out, could derail the AI companies. My view is that in the short term politics can cause setbacks, a state can restrict data centers or a city can hold hearings, but over fifteen or twenty years of American technology history the companies have kept advancing whichever party held power, and the checks and balances of a federal system make it very hard for anyone to stop a technology outright. I would treat the politics as volatility around the long-term trend rather than a threat to the trend itself.
How to approach a listing like this
The way I framed it for Neil is that Anthropic should be treated as what it still is, a startup that happens to be very large, and an investment in it should be sized as a bet rather than as a core holding. The earlier one owns a company growing like this the better the returns tend to be, but nobody knows yet how the model layer will settle: whether open-weight models will take a much larger share, whether Google or another lab will close the gap, or whether the two leaders will keep most of the revenue as they do today. A small allocation, a close watch on the next few quarters, and a willingness to adjust as the evidence comes in is a reasonable way to hold that uncertainty.
For an Indian investor there are a few practical points worth keeping in mind. IPO allocations rarely reach resident Indians through LRS platforms, so for most readers the realistic entry is on the secondary market after listing, which means paying whatever price the first days of trading set; large, heavily anticipated listings often open well above their offer price, and waiting for the first two public quarters and the end of the initial lock-up is a legitimate choice rather than a timid one. The mechanics are the same as for any US stock, within the $250,000 annual LRS limit, with TCS above ₹10 lakh, long-term capital gains after 24 months, and US estate-tax exposure on a single US-listed share holding above $60,000. Many investors also own a slice of this story already without realising it: Amazon and Alphabet hold large stakes in Anthropic and sit inside the Nasdaq-100, and the companies selling it chips and capacity sit inside the broad semiconductor and data-center funds.
That last point is the one I would emphasise most. The model layer is the hardest place in the AI stack to pick a winner, because it is still changing quarter to quarter, and that is why the Thefynprint Global AI portfolio holds the top two layers, applications and models, through a broad index and tilts its core towards the physical layers beneath them: chips, data centers and energy. Those are the companies on the receiving end of commitments like Anthropic’s, and they are profitable, revenue-generating businesses today. The portfolio has not been immune to the volatility of this year, having peaked at around 37% in June and fallen to roughly 15% during the late-summer pullback, but as of 7 October it stood about 29% higher in rupee terms since its launch on 5 February, at a time when the Nifty and broader emerging markets have struggled and US bond yields have moved above 5%. The earnings across that supply chain have been growing at 20 to 30% a year, and outside the private labs themselves, the valuations still look reasonable against that growth.
The five layers of AI behind the portfolio framework
The days since we recorded show just how volatile this theme can be. Between 7 October and the close on Friday 16 October, a span of seven trading sessions, the portfolio gave back roughly 13 percentage points, from about 29% to about 16% in rupee terms, which is a fall of close to 10% in its value in little more than a week. The semiconductor ETF at its core dropped about 14% over those sessions and the Nasdaq-100 about 6%, while a firmer rupee, which moved from around 96.4 to 94.5 against the dollar, deepened the decline for an Indian investor. None of the earnings or contractual commitments underneath the portfolio changed in those seven sessions, and that gap between how fast prices move and how slowly fundamentals do is the reason position sizing and holding period matter far more than timing in a theme like AI.
Portfolio gains and volatility since launch, as stated on 8 October
The long view
In 2008 very few investors wanted to own cloud infrastructure, because the revenue looked tiny next to the capital being spent; by 2020 the platforms were there, and the capex of a decade earlier looked cheap. Anthropic’s prospectus reads like an early chapter of a similar story, with one difference that earlier cycles never had, which is that the builders themselves are publicly asking how fast the technology should be allowed to go. That question will be argued for years, and while it is argued, the compute keeps getting built and the revenue keeps compounding. Whether one buys the IPO, waits for a few public quarters, or owns the theme through the layers underneath, the more important decision is to own an industry being born with position sizes that can survive being wrong about any single company. The thesis is just getting started.
If you want the full background, watch my conversation with Neil on the Anthropic IPO on the podcast, where we also discuss Meta’s Muse, why Amazon blocked AI shopping agents, and how the portfolio is built across the five layers, and see the Global AI Portfolio factsheet for its construction and latest performance. To talk to the research team about whether this fits alongside your existing domestic and global holdings, you can book a consultation call through Thefynprint.
This article is for education only and is not investment advice. Figures from Anthropic’s filing are taken from a draft prospectus reported in the press and may change in the public S-1; portfolio returns are for a model portfolio and past performance does not guarantee future results.