From FR Advisory.
The author spent twenty years running commercial organisations inside software and private equity portfolios. He no longer works in the technology industry, holds no position in any company named in this series, has no clients in it, and is selling nothing to anyone in it.
Updated 19 November 2026, after the close. The first version of this piece was written the night before Anthropic listed, using press-reported pricing. Two figures in it were wrong and are corrected here, with the detail in the Sources note.
Start with the day, because the day did something more useful than the week.
Anthropic's final prospectus priced 980 million Class A shares at 83 dollars. The offering raised 81.34 billion dollars, with a concurrent private placement of 26 billion in a separate share class at the same price. That is 107.34 billion dollars of capital raised in a single transaction, the largest initial public offering on record.
It opened at 111.50 dollars, 34% above the offer. Within minutes it touched 115.10, up 39%, and the valuation briefly passed two trillion dollars.
It closed at 97.17.
Up 17.07% on the day, which every headline reported as a triumph. But measured against its high the stock lost 15.6%, and against the open 12.9%. Of the 32.10 dollars of first day gain it had in hand at the high, 56% was gone by four o'clock. It traded as low as 92.72. Volume was around 840 million shares, which is 86% of the entire offering changing hands in four and a half hours.
Nvidia reported its third quarter two days earlier. Revenue of 115.8 billion dollars, up 103% year on year and 20% sequentially. Data centre revenue of 107.6 billion, up 110%, now 93% of the company. Guidance of 130 billion for the next quarter, and that guidance assumes no data centre compute revenue from China at all.
And on the day the largest listing in history arrived, the Dow fell 2.19%, the S&P 500 fell 2.02% and the Nasdaq Composite fell 2.25%. Nebius fell 7.8%, CoreWeave 5.0%, Nvidia 1.7%.
Hold those two things together. The technology is working, the revenue is real, the capital arrived in record size. And the trade it was supposed to validate went down.
What everyone got wrong by Tuesday lunchtime
Before the argument, the correction, because almost every outlet is still running the wrong number.
The headline from the Anthropic filing was sixteen point three billion of revenue against eighty point seven billion of losses.
Ninety per cent of that loss, 72.4 billion dollars of it, is the change in fair value of convertible notes, derivative liabilities and simple agreements for future equity. The filing states what drove it: the fair value of the convertible notes rose as the enterprise valuation increased.
The loss got bigger because the company got more valuable.
The operating loss was 693 million dollars. On 16.286 billion of revenue in six months, against 1.163 billion in the same period a year earlier. Revenue growing fourteen times year on year, at a business running 4.3% below operating breakeven.
Anyone arguing this cycle is collapsing on the strength of an eighty billion dollar loss has built the case on an accounting artefact of success. The bears took the wrong number and the bulls let them, because it suited both.
What was supposed to happen
Here is the comparison nobody runs, and it is the one that locates us.
The case made to every board between 2023 and 2025 was not that model companies would become enormous. It was that artificial intelligence would show up inside ordinary companies, as margin. Lower cost to serve, higher output per head, faster cycles, landing in the operating line within two to three years.
That was the plan. Mark it.
MIT's work put 95% of enterprise AI pilots at no measurable return. IBM's 2026 study of two thousand chief executives found 76% of organisations have appointed a Chief AI Officer against 25% of the workforce using the tools regularly. In the United States this year, artificial intelligence was named in 121,401 announced job cuts, which is the cost side arriving ahead of the benefit side.
Against all of that, spending accelerated. UBS put total hyperscaler capital expenditure at 492 billion dollars in 2025, about 1.009 trillion in 2026 and around 1.447 trillion in 2027, with Amazon, Alphabet and Microsoft spending roughly 102% of their cloud revenue on capital expenditure this year.
The plan was that returns would arrive by now and spending would normalise. What happened is that returns did not arrive and spending doubled.
That is not a bubble in the ordinary sense. A bubble is a crowd paying too much for something real. This is a crowd that stopped asking when the thing arrives, and doubled the stake while waiting.
The line nobody quoted
One number in the Anthropic filing has gone unmentioned everywhere, and it is the most important line in the document for anyone who sells software for a living.
Sales, marketing and partnerships. Three thousand one hundred and sixty eight million dollars in the first half of 2026. The same line a year earlier was one hundred and twenty million.
Twenty six times. Revenue grew fourteen times. The cost of selling grew twenty six.
For scale, compute and infrastructure in the same period was 8.945 billion, and personnel was 3.152 billion including 2.2 billion of stock based compensation. The go to market line is now in the same weight class as the people line.
The most capable technology ever commercialised, at the company that builds it, with inbound demand most businesses would consider fantasy, did not sell itself. It required a distribution machine built at enormous speed and enormous cost.
Two years of boardroom conversation has rested on the assumption that sufficiently good technology compresses the cost of going to market. The company at the frontier has published the receipt and the receipt says the opposite. Capability went up. The cost of selling went up faster.
If that is true at the frontier, there is no version of it that is false in your business.
Identical economics. Opposite verdicts.
Now the argument, and until yesterday it had a hole in it that I want to name before I make it.
Three weeks ago HubSpot's chief executive moved the company's support agent from a dollar per conversation to fifty cents per resolution. She published the rate card. She let customers convert contracted seat spend into credits mid-contract without waiting for a renewal. That is, almost exactly, what every buyer survey says buyers want and what this framework argues vendors must do.
The shares fell 19% on the announcement. They fell 19% again on the quarter that beat revenue, operating income and earnings, the worst single day in the company's twelve years on the New York Stock Exchange. Net revenue retention slipped to 102%, and the chief financial officer attributed part of that to customers swapping seats for credits, which is the mechanism working exactly as designed. At 11 November the stock sat at $196.99, down 50.91% on the year, though the contemporaneous reporting attributes most of that to sector pressure rather than the pricing move. Eight days after those results the chief executive left and the shares rose roughly six per cent.
A week earlier, ServiceNow guided fourth quarter subscription revenue growth of 21.5% in constant currency against contracted backlog growth of 18.5%. Revenue outrunning backlog is the arithmetic signature of converting committed seats into variable consumption. The shares fell 27% in a day, the worst in the company's history, on a quarter that beat and raised.
Both of those verdicts were produced by trading. By thousands of participants, over days, discovering a price.
Yesterday I could not say the same about Anthropic, and I wrote the comparison anyway. Its valuation was a number negotiated between underwriters and a book of institutions. That is a different instrument from a share price. One is set, the other is discovered, and an argument about what the market concludes cannot use a price the market had not yet produced.
Today it has produced one.
No seats. No contracted backlog. No renewal date. No committed term. Revenue any customer can reduce to zero this afternoon by changing one line in a configuration file. The top five customers are about a quarter of revenue.
Roughly 1.76 trillion dollars on a fully diluted basis at the close, against about 1.5 trillion at the offer price on the same basis. Fifty four times annualised first half revenue, up from forty six at the offer. On shares outstanding rather than fully diluted it is about 1.27 trillion and thirty nine times, and anyone quoting a valuation for this company without saying which basis they are using is not telling you anything.
Three companies. The same economics. Variable revenue, no contractual floor, value that has to be re-earned every month.
Two were punished for acquiring those properties. One was awarded the largest initial public offering in history for being born with them.
The variable is not the business model. It is the story the business model is allowed to tell.
Incumbents are marked against what they are giving up. Every quarter of a transition reads as decay, because the instruments were built to measure the thing being left behind. Entrants are marked against what they might become, and there is no instrument for that, so the number is whatever the story supports.
The four hours
And then the part I did not expect, which is the most useful thing that happened all week.
Yesterday's version of this piece ended with the line: when the story needs a number, there will not be one.
The story needed a number at 11:48 in the morning, New York time.
The market opened Anthropic 34% above its offer, ran it to 39%, and then spent four hours taking 56% of that gain back. Not on news. Not on a downgrade. Not on any disclosure, because there was none to make. It went up on the story and came down on the absence of anything to check the story against, and it did that in a single session with 86% of the offering turning over.
That is the repricing mechanism I argued would take quarters, arriving on day one.
The low of the day, 92.72, is the more interesting number than the close. At that price the stock had given back three quarters of its opening gain. Somewhere between 115.10 and 92.72, a market that had just paid a record price went looking for a reason and did not find one, because for a consumption business at this scale there is nothing standard to look at. No contracted backlog. No net revenue retention anyone has agreed how to calculate for a token meter. No disclosure of value realised against spend.
Seventeen per cent up on the day is a successful listing and I am not going to pretend otherwise. But the shape of the day is the finding. The price was not discovered against evidence. It was discovered against sentiment, twice, in opposite directions, inside four hours.
Why the instruments produce this result
This is not stupidity, and treating it as stupidity is why most commentary on it is useless.
Net revenue retention was designed for a world with three outcomes: expand, flat, churn. It cannot represent a customer who spent the same money through a different meter, and it certainly cannot represent one who will spend more through that meter next year. Conversion and contraction look identical.
Contracted backlog measures committed future contract value. Consumption does not book that way. A company genuinely succeeding at the transition shows softening backlog almost by construction, because it is exchanging multi-year commitments for variable usage. Decline and success produce the same line.
So for an incumbent mid-crossing, both headline instruments return a false negative. Reliably. Every quarter. For as long as the crossing takes.
For a company that never had seats, neither instrument applies and nothing has been built to replace them. There is no accepted measure of consumption durability. No standard disclosure for value realised against spend. No agreed way to tell a customer whose usage plateaued because they finished rolling out from one whose usage plateaued because they are leaving.
An absence of instruments does not produce caution. It produces whatever price the story can support, and then it produces yesterday afternoon.
What this means for you
If you run an incumbent. The market will punish you for crossing, on schedule, whether you cross well or badly. HubSpot crossed well. That is the finding. So the question is not whether crossing is correct, it is whether you can survive being measured on the old instruments for the four to eight quarters it takes, and whether your board understands that before the first bad print rather than after it.
If you are early or pre-revenue. You are valued on possibility, and possibility is the only asset in this cycle with no floor under it. Yesterday showed how fast that floor gets tested. The companies that come through will be the ones that built the measurement of realised value before anybody demanded it, because when the narrative premium goes, evidence is the only defence.
If you sit on a board or an investment committee. One question at the next meeting. What number, that we currently publish, would tell us the difference between a customer finishing their rollout and a customer quietly leaving? If nobody can answer, you do not have a forecast. You have a trend line and a hope.
If you carry a quota. None of this has changed what you are paid for, and that is the whole of Season Three. Pricing moved at thousands of companies this year. Metering moved at hundreds. Compensation moved at almost none. You are being asked to sell a consumption product on a commission plan designed for a signature.
How this plays out
I would rather be wrong in public with a mechanism than vaguely right without one, so here it is, and one phase of it has now partly arrived.
Phase one. The funding changes shape. The spending is no longer coming out of cash flow. Amazon's trailing twelve month free cash flow to September was negative 42.4 billion dollars. Alphabet's long term debt went from 46.5 billion in December 2025 to 125.1 billion by September 2026. Microsoft's free cash flow is positive but down 27% year on year. FactSet's analysis has incremental debt rising from 9% of capital expenditure in the 2024 financial year to around 32% by mid-2026. The Kansas City Fed estimates data centre special purpose vehicles are financed around 90% with debt. The buildout has moved from being paid for by profits to being paid for by credit, and that changes who gets hurt and how fast.
Phase two. The circle gets noticed. The 26 billion dollar private placement alongside the listing is filed, and the purchasers are reported to be AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Those are its suppliers and its cloud partners. The IMF counted roughly 300 billion of hyperscaler equity and 650 billion of debt placed into weaker AI counterparties, against about 1.4 trillion of future revenue commitments. Supplier revenue funded by supplier capital is not demand. It is the same money doing two laps and being counted twice.
Yesterday gave this phase a data point. The largest capital raise in the history of public markets completed, and the sector it belongs to fell more than two per cent on the day. A record listing did not create confidence in the AI trade. It absorbed 107 billion dollars of capital and the trade went down. In fairness, the Federal Reserve minutes and hotter inflation data were in the same session, and AMD's nine per cent fall followed a disclosed short position rather than pure sentiment. The debut did not cause the selloff. It simply failed to stop it, which for the largest IPO ever executed is the more telling result.
Phase three. The first deceleration. Consumption revenue has no contracted floor. No backlog to point at, no renewal cohort to lean on, no four quarter cushion in which to fix the story. The first time a frontier lab prints a quarter where growth slows for ordinary reasons, there will be no instrument capable of separating a plateau from a decline.
I thought that was a story for 2027. Yesterday afternoon was a four hour rehearsal of it, with no bad news required. The price went up on narrative and came down on the absence of anything to verify it. A quarter of slowing growth would do the same thing with a reason attached, and it would not stop at one company.
Nvidia's own margin is the clock. Non-GAAP gross margin was 74.9% this quarter, guided to between 71 and 72 next. Over three points of compression while revenue doubles. Even the shovel business is getting less profitable per dollar sold.
What would have to be true for me to be wrong
No hedging, so here are the conditions.
If enterprise AI return shows up in operating margin rather than in pilots, at scale, inside twelve months, I am wrong and the valuations were simply early. If a frontier lab publishes a durable consumption measure that survives a slow quarter and the market accepts it, I am wrong about the instruments. If hyperscaler capital expenditure gets funded back out of cash flow rather than credit as revenue catches up, I am wrong about the transmission into the wider economy.
And one more, added today. If Anthropic holds above its offer price through its first reported quarter without a standard consumption disclosure, then narrative is more durable than I am arguing and the absence of instruments matters less than I think.
Those are falsifiable, they are public, and they resolve inside eighteen months.
The position
The technology is working. The capital is flowing, in record size. The revenue at the model layer is real and growing at a rate almost nothing in corporate history has matched.
The business model never moved.
It is still priced, measured, reported and compensated on recurring revenue principles designed for software sold by the seat to a human being, in a world where neither the seat nor the human is the unit of value any more. Pricing pages were rewritten at thousands of companies this year. Metering changed at hundreds. Compensation, revenue architecture and the instruments handed to the market, almost nowhere.
That gap is not a technology risk. It is a measurement failure with a trillion dollars of credit leaning against it.
For three weeks the market demonstrated it cannot tell a crossing from a collapse. Yesterday it demonstrated something worse. Handed a company with no crossing to misread, no legacy to mourn and nothing but possibility to price, it marked the thing up 39% and then took back 56% of the gain in four hours, because there was nothing in the filing anyone has agreed how to read.
The story needed a number. There was not one.
Brace yourself.
The Agentic Commercial Model. agenticcommercialmodel.com
Sources
Corrections to the first version of this piece
- The first version, published on the evening of 18 November, stated that Anthropic "began trading on Nasdaq under the ticker ANT". It had not; it listed the following day. Corrected within the hour of being identified
- The first version stated the company "raised at least 68 billion, more than 94 billion including a 26 billion private placement", from press reporting. The filed figures are an offering of 81.34 billion dollars and a concurrent private placement of 26.0 billion, so 107.34 billion in total. The press figure used was materially low
- The first version gave the pricing terms as fact on the strength of press reporting, before any final prospectus was filed. Those terms are now filed and are cited to the filing below
Primary filings
- Anthropic, PBC Form 424(b)(4) final prospectus and pricing free writing prospectus, filed 19 November 2026: 980,000,000 shares of Class A common stock at a price to public of $83.00, comprising 956,351,354 shares offered by the company and 23,648,646 by selling stockholders; total offering size excluding the option $81,340,000,000; price to underwriters $82.43878 per base share; underwriters' option for up to 147,000,000 additional primary shares; concurrent private placement of $26.0bn of Class N common stock, being 313,253,007 shares at the price to public; listed on the Nasdaq Global Select Market under "ANT". The pricing document does not name the private placement purchasers
- Anthropic, PBC Form S-1, publicly filed 26 October 2026: revenue $16,286m for the six months ended 30 June 2026 against $1,163m in the prior-year period; operating loss $(693)m; net loss $(80,748)m against $(12,969)m; change in fair value of convertible notes, derivative liabilities and SAFEs $(72,388)m, described in the filing as primarily driven by the increase in fair value of the convertible notes as enterprise valuation increased; other expense, net $(8,757)m; compute and infrastructure $8,945m; sales, marketing and partnerships $3,168m against $120m; personnel $3,152m including approximately $2.2bn of stock-based compensation; other operating and administrative $1,714m; total liabilities $117,467m at 30 June 2026; purchase and lease commitments substantially all related to compute and infrastructure, with substantially all of the remainder falling between 2031 and 2045. A widely circulated $321bn figure for future lease obligations is not used here, because the dollar amounts in that disclosure were not legible in the filing text reviewed
- NVIDIA Corporation Form 8-K, third quarter fiscal 2027, quarter ended 25 October 2026, released 17 November 2026: revenue $115.8bn, up 103% year on year and 20% sequentially; Data Center revenue $107.6bn, up 110% year on year; GAAP diluted EPS $2.70, non-GAAP $2.68; GAAP gross margin 74.8%, non-GAAP 74.9%; fourth quarter revenue guidance $130.0bn plus or minus 3%, assuming no Data Center compute revenue from China; fourth quarter gross margin guidance 71.0% to 72.0%
First-day trading, 19 November 2026
- Opening price $111.50, first trade at approximately 11:48 ET; intraday high $115.10; intraday low $92.72; close $97.17, up $14.17 or 17.07% on the offer price; volume approximately 840 million shares, about 86% of the 980 million shares sold. Reuters reported the valuation briefly passed $2tn shortly after the open
- Derived here from those figures: the stock fell 15.6% from its intraday high to the close, and gave back 56% of the $32.10 first-day gain it held at the high
- Valuation at the close approximately $1.27tn on shares outstanding per Nasdaq's listing page, and approximately $1.76tn on a fully diluted basis. Multiples in this piece annualise first-half revenue of $16.286bn to $32.57bn, which understates forward revenue at a company growing fourteen times year on year: 46x at the offer price and 54x at the close on a fully diluted basis, 39x at the close on shares outstanding
- Index closes: Dow Jones Industrial Average 49,097.06, down 1,101.11 points or 2.19%; S&P 500 7,175.83, down 147.80 or 2.02%; Nasdaq Composite 23,879.20, down 550.69 or 2.25%. Nebius down 7.8%, CoreWeave down 5.0%, Nvidia down 1.7%, AMD down 9.2% following a disclosed short position by a Hunterbrook-affiliated fund. Reuters attributed the session's weakness to Anthropic's debut failing to revive AI enthusiasm alongside Federal Reserve minutes and hotter inflation data, so the selloff is not attributed here to the listing alone
Third-party estimates, attributed and not independently verified
- UBS, reported 22 August 2026: total hyperscaler capital expenditure of $492bn in 2025, approximately $1.009tn in 2026, $1.447tn in 2027 and $1.619tn in 2028; Amazon, Alphabet and Microsoft spending approximately 102% of cloud revenue on capital expenditure in 2026
- FactSet, 23 July 2026: incremental debt rising from 9% of capital expenditure in FY2024 to approximately 32% on a trailing basis by mid-2026
- Federal Reserve Bank of Kansas City, 5 October 2026: approximately 90% debt financing for data centre special purpose vehicles and approximately 85% for neoclouds
- International Monetary Fund, October 2026: approximately $300bn of hyperscaler equity and $650bn of debt placed into weaker AI counterparties, against approximately $1.4tn of future revenue commitments
- Free cash flow and debt figures for Amazon, Alphabet, Microsoft and Meta are company filings as compiled by A.L. Capital Advisory, version 3.0, 2 November 2026, not independently re-checked against each original filing
- The UBS, FactSet, Kansas City Fed and IMF figures are relayed via that same compilation, which states its figures should be checked against primary sources before use. They are presented here as attributed estimates, not verified fact
Carried from earlier issues of this series, with the caveats stated there
- MIT: 95% of enterprise AI pilots showing no measurable return
- IBM Institute for Business Value with Oxford Economics, 2026 CEO Study, 2,000 chief executives across 33 geographies: 76% with a Chief AI Officer against 25% of the workforce using AI regularly
- Challenger, Gray & Christmas, October 2026: artificial intelligence named in 121,401 announced US job cuts year to date, about 20% of the total, with total announced cuts of 617,189 down 44% year on year
- HubSpot: Customer Agent repriced 14 April 2026 from $1 per conversation to $0.50 per resolution; shares down 19% on 8 May and 19% at the close on 6 August, having been down 20.2% intraday, the latter the largest single-day fall in twelve years on the NYSE, on a quarter beating revenue, operating income and adjusted EPS; net revenue retention 102%, attributed in part by the CFO to customers swapping seats for credits; down approximately 51% year to date at 11 November 2026, a decline the contemporaneous reporting attributes substantially to broad sector pressure rather than the pricing transition alone; Yamini Rangan stepped down 12 November 2026 with shares rising approximately 6%
- ServiceNow: Q4 guidance of 21.5% constant-currency subscription revenue growth against 18.5% cRPO growth; shares down 27% on 29 October 2026, the company's worst single day