Editor's note, 22 September 2026. The price bands in this piece were built on the $47 billion run rate Anthropic disclosed in May. The business has since moved far past that figure, and we have revised our view. Read the revision, We Were Wrong on Anthropic, here: [LINK]. We have left the original bands visible below so readers can see exactly what changed and why.
What we said, and why it no longer holds
In August we published price dependent bands on Anthropic. Buy below roughly $700 billion of implied market capitalisation. Hold to $1.1 trillion. Trim to $1.3 trillion. Avoid above that. At the then rumoured $1 trillion plus listing, we rated it a hold or pass for most public investors.
Those bands were anchored to a $47 billion annualised run rate, the figure Anthropic disclosed alongside its $65 billion Series H in May.
That number is now four months old and the business has moved a long way past it.
Anthropic’s run rate went from $9B at the end of 2025 to $65B confirmed in July, with more than $100B expected by year end. Our August bands were built on the $47B figure from May.
Bloomberg reported on 17 August that the run rate reached $65 billion by the end of July, with second quarter revenue above $11.5 billion, more than double the first quarter’s $4.73 billion and over fourteen times the same quarter a year earlier. On 18 September the New York Times reported, via Bloomberg and Axios, that the figure is expected to exceed $100 billion this year. The Financial Times has reported investors modelling $100 to $120 billion by year end.
That is about seven times the level of a year earlier, and more than $100 billion is expected by year end.
We want to be precise about the nature of our error, because it matters for how you should read our future work. The framework did not fail. We ran a five year view of what the equity would be worth and worked backwards to an entry price, which is the same method we applied to OpenAI this week and would apply again. What failed was the vintage of the input. We built on the most recent company disclosed figure available, and in a business compounding at this rate, four months was long enough for that figure to stop describing the company.
The lesson is not that we should have forecast higher. It is that in assets where the denominator can double inside a quarter, a valuation band needs a stated expiry and a trigger for rerunning it. Ours had neither. It does now.
The revision
In August we said buy below $700B and avoid above $1.3T. Rebuilt on current numbers, it’s buy below $1.15T, hold to $1.6T, trim to $2.05T. The rumoured $2T listing moves from deep in avoid to the top of trim.
We have rerun the same model with Anthropic’s current reported profile. Three inputs changed materially.
The revenue base moved from $47 billion to $65 billion confirmed in July, with more than $100 billion expected by year end. Gross margin came in far higher than we assumed, with the FT reporting margins above 80% before revenue sharing and training costs, against the roughly 60% we had modelled. And the profitability date moved from a projection to something closer to a fact, with the FT reporting a second consecutive quarter of adjusted operating profit, excluding stock based compensation.
That last point deserves care. Adjusted operating profit excluding stock compensation is not GAAP net income, and the exclusion can be large at a company issuing equity at this scale. It is still a materially different picture from the one we modelled in August.
The revised bands: buy below roughly $1.15 trillion, hold to $1.6 trillion, trim to $2.05 trillion, avoid above that.
The rumoured $2 trillion listing now sits at the top of trim rather than $700 billion inside avoid. That is a large move and we are not going to dress it up as a refinement.
Where this lands in five years
Our base case is $470B of revenue by 2031, worth about $1.3T today. At base case margins, a $2T listing needs roughly $720B of revenue to earn 15% a year.
The bands come from a view of 2031, not from a multiple on today’s revenue.
In our base case, Anthropic exits 2026 near the reported $100 billion run rate and compounds to roughly $470 billion of revenue by 2031 as enterprise adoption broadens and coding agents mature. At a 25% net margin and 28 times earnings, the equity is worth around $3.3 trillion in 2031. Assume today’s shareholders keep 80% of that after future dilution, discount it back five years at 15% a year, and it supports paying roughly $1.3 trillion today. Our buy line of $1.15 trillion sits about 12% below that, the same margin of safety we applied to OpenAI.
The bull case has revenue near $850 billion at a 30% margin and 30 times earnings, worth about $7.7 trillion in 2031, which would justify paying around $3 trillion today. That case requires Anthropic to hold the enterprise lead through at least two more model cycles, and requires margins to survive the open weight price floor.
The bear case flattens to $200 billion by 2031 as competition commoditises the workloads that currently pay well. At a 15% margin and 22 times earnings the equity is worth about $660 billion, which supports paying only around $265 billion today.
The spread between those outcomes is more than ten times, which is why we publish a band rather than a target. One number is worth holding onto: at base case margins, a $2 trillion listing needs roughly $720 billion of revenue in 2031 to earn 15% a year, against our base case of $470 billion.
The three arguments against us
The bull case is well covered elsewhere. Here are the serious objections, including the ones circulating publicly that deserve better treatment than they usually get.
The audit theory. A widely shared argument holds that the recent shift in tone from Anthropic, including the profitability signals and the safety rhetoric, is downstream of the PCAOB audit required for an S-1. The reasoning is that a rigorous audit forces honest treatment of credits, discounts and revenue recognition, and that what looks like strategy is actually disclosure discipline arriving on a deadline.
We cannot verify this and neither can anyone posting it. But the underlying point is correct and worth holding onto: reported run rates and adjusted profit figures have never been through an audit, and the public S-1 will be the first document that tests them. If recognised revenue comes in materially below what the run rate implies, our revision was premature.
The passing the bag critique. The related argument is that a $2 trillion listing exists to move ownership from venture funds into public markets and retirement accounts at the top of a cycle. This is unfalsifiable as stated, and it is also not baseless. The Series H at $965 billion was led by Altimeter, Dragoneer, Greenoaks and Sequoia, crossover style investors who typically buy late private rounds expecting a higher public print, rather than investors taking a long term private position. The honest response is not to dismiss it but to price it: a first print at $2 trillion leaves very little room for error, and the float will be small enough to make early trading violent.
The safety contradiction. On 12 September Dario Amodei published an essay calling for the industry to slow the rate at which capabilities advance. Sam Altman and Elon Musk backed it the same day and AI stocks fell on the Monday. Reuters then reported that Anthropic is weighing a new frontier model to respond to OpenAI’s Astra, ahead of its own roadshow.
Two readings compete. The charitable one is that a company preparing to put frontier risk into a prospectus as a legal risk factor has strong reason to state its position publicly and early. The cynical one is that an incumbent with the current lead benefits from a slower race, and that safety language is a competitive instrument.
We do not think the evidence settles it, and we would caution against the confident version of either. What matters for the equity is narrower: if Anthropic genuinely paces capability releases while a competitor does not, that is a revenue cost, and it will show up in the enterprise share data within two quarters. If it ships a counter model before the roadshow, the pacing argument was rhetorical. Either way, the answer arrives soon and it is observable.
The comparison that decides both listings
On every reported figure, Anthropic looks better and cheaper than OpenAI: a higher confirmed run rate, a lower multiple at about 15x, and gross margins above 80% before revenue share. None of it is audited yet, and the prospectus is where that gets tested.
We published on OpenAI this week and concluded that at $1.2 to $1.5 trillion it is priced above what the evidence supports. The relative position is the sharpest part of that argument.
On reported figures, Anthropic has a higher run rate, a lower multiple on it, dramatically better gross margins, adjusted profitability already arriving rather than guided for 2030, and a listing date three to twelve months earlier. Every one of those figures is unaudited and every one favours Anthropic.
That is the problem OpenAI’s bankers have to solve. A public Anthropic will set the comparable multiple for frontier labs, and it will do so before OpenAI reaches the market.
It is also the reason we are not more bullish than we are. A company with these reported numbers listing at $2 trillion is asking for roughly 20 times run rate at a moment when its own founder is arguing publicly for slower development, its competitor has just taken back developer share in two months, and nothing in the financial record has been audited.
What we would need to see
We are at the top of trim rather than in avoid, and we would rather be precise about what moves us either way.
Upward, toward a genuine buy at listing: recognised revenue in the public S-1 that tracks reasonably close to the run rate implied figures, gross margin holding above 70% after revenue sharing is disclosed properly, and evidence that enterprise retention survives OpenAI’s price competition following Astra.
Downward: a material gap between recognised revenue and run rate in the prospectus, disclosure that adjusted profitability depends heavily on excluded stock compensation, a further slip of the listing into live midterm election volatility, or enterprise share data showing the Astra recovery continuing into a second quarter.
The single most useful number to watch before the roadshow is not the valuation. It is whether the reported $15 billion revolving credit facility closes. A company that can walk away from a soft market prices differently from one that cannot.
The correction, stated plainly
We were too conservative on Anthropic in August, by a wide margin, and the cause was using a four month old revenue figure in a business that was doubling inside a quarter. The framework was sound and we applied it to bad inputs.
We are revising to buy below $1.15 trillion, hold to $1.6 trillion, trim to $2.05 trillion, and avoid above that. The rumoured $2 trillion listing sits at the edge of what we would pay, not far beyond it.
The next revision will come from the public S-1, which will be the first audited look at any frontier lab and will either validate this or embarrass us again. We would rather publish that when it happens than quietly update a post.
Source register
A means issuer primary disclosure. B means identified reporting on private information. C means commentary, used within its stated limits.
Our original call
Rovenin, Anthropic pre IPO research. 21 Aug 2026. https://www.rovenin.com/research/anthropic-ipo-research-the-first-audited-look-at-frontier-lab-economics
Revenue and run rate
AIwire, Anthropic $65B Series H at $965B post money, $47B run rate. 29 May 2026. A syndicated. https://www.hpcwire.com/aiwire/2026/05/29/anthropic-raises-65b-in-series-h-funding-at-965b-post-money-valuation/
Bloomberg, run rate surpasses $65B. 17 Aug 2026. B. https://www.bloomberg.com/news/articles/2026-08-17/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo
Axios, run rate and IPO timing. 17 Aug 2026. B. https://www.axios.com/2026/08/17/anthropic-revenue-run-rate-ipo-openai
Bloomberg citing the NYT, annualised revenue to top $100B in 2026. 18 Sep 2026. B. https://www.bloomberg.com/news/articles/2026-09-18/anthropic-s-annualized-revenue-to-top-100-billion-in-2026-nyt
Axios, $100B pace and November trading. 18 Sep 2026. B. https://www.axios.com/2026/09/18/anthropic-100-billion-revenue
Margins and profitability
Bloomberg citing the FT, second adjusted operating profit, gross margin above 80% before revenue share and training. 13 Sep 2026. B. https://www.bloomberg.com/news/articles/2026-09-13/anthropic-sees-adjusted-operating-profit-this-quarter-ft-says
Deep Quarry (Olga Usvyatsky), the adjusted measure excludes stock based compensation. Sep 2026. C.
IPO process
Bloomberg, $15B revolving credit facility led by Morgan Stanley with Goldman, JPMorgan and Citi, the same banks reported to lead the IPO. 3 Sep 2026. B. https://www.bloomberg.com/news/articles/2026-09-03/anthropic-nears-finalizing-15-billion-pre-ipo-credit-facility
BeBeez, roadshow moved to mid October, Series H leads (Altimeter, Dragoneer, Greenoaks, Sequoia), valuation talk up to $2T, Nvidia anchor talks. 18 Sep 2026. B. https://bebeez.eu/2026/09/18/anthropic-postpones-its-nasdaq-ipo-until-mid-october-and-attacts-a-revolving-facility-of-15-billion-us-dollars-nvidia-holding-talks-for-pouring-10-billion-as-anchor-investor/
Competition and the safety debate
Techstrong citing Reuters, Anthropic weighing a new model in response to GPT-6 Astra. 19 Sep 2026. B. https://techstrong.ai/articles/anthropic-weighs-counterstrike-against-openais-gpt-6-astra-ahead-of-2-trillion-ipo-report/
Reuters via The Star, reaction to the 12 September industry slowdown pledge. 15 Sep 2026. B. https://www.thestar.com.my/business/2026/09/15/ark039s-wood-says-global-ai-leaders039-pledge-is-positive-for-technology-development
Arguments circulating publicly, not independently verifiable The audit theory and the passing the bag critique are summarised from widely shared commentary. We present them as arguments, not facts, and do not attribute them to named individuals.
Model Five year scenarios, 80% ownership retention and the 15% required return are Rovenin assumptions, not company guidance.