Executive Summary

What the Journal carried on September 19, 2026 was a date being moved. Anthropic's listing had slipped from October to November, and the reason given was to put third-quarter results in front of investors first. Waiting another month to walk out holding a results sheet sounds like a company whose price rests on results. Except that the calculation producing this price tag sits in neither the third quarter nor the fourth. Which year it does sit in is the subject of this article, and the cleanest place to start is by separating what is settled from what is still being discussed.

First separation: there is no public registration statement for this offering. The company filed confidentially on June 1, 2026, and that document moves only through the review process. No exchange, no ticker, no share count, and no lead underwriter has been officially confirmed. The valuation, the raise, and the timing discussed below all rest on press reports and on what sources told reporters, not on anything read in a filing. And the September 19 facts were confirmed through an article that re-reports the Journal's story. We could not open the original and check it line by line, so this article treats those facts at the grade of secondhand reporting.

When that document opens, though, is fixed by rule. Confidential filing comes with a paired condition. A company may revise its draft while trading comments with regulators, but it must file a revised statement publicly at least 15 days before it starts marketing the offering to institutional investors. That is the moment audited financials, risk factors, major holders, and the use of proceeds come out. So the items marked below as "not verifiable right now" are not sealed forever. They open together, two weeks before the roadshow, and several sentences here will have to be read again that day.

On top of all that sit the two figures under discussion: the roughly $2 trillion valuation and the raise of up to $100 billion. Neither is settled; both are proposals being weighed. A sense of scale makes the later sections easier, though. No public offering has ever raised $100 billion at once, and $2 trillion would put a company among the world's largest by market capitalization from its first day of trading. What is being discussed is attaching that figure to a company less than six years past its first institutional round.

The number did not appear from nowhere. The valuation was $623 million at the Series A in May 2021, $380 billion at the Series G in February 2026 alongside a $30 billion raise, and $965 billion at the Series H on May 28, which raised $65 billion. Four months after that, the figure under discussion has doubled again. Through all of that movement in the private market, not one public document has existed against which to check the number. The listing is the first time that check happens in the open.

Private prices and public prices are produced differently. The first is what a small set of investors agreed on in a single round, with no public document behind the agreement. The second is what thousands of buyers and sellers re-set every day. Nothing guarantees that a price survives the crossing from one to the other, and the recent cases in Section 5 show what the crossing can look like. Roughly $2 trillion, then, is closer to a starting line than a destination. That is also why this article traces the formula rather than arguing the number up or down. The formula can be checked now; the number gets settled later, by the market.

Laying the past four months of coverage out in order shows where confirmed facts, sourced statements, and computed figures started mixing together.

Money has been moving outside the company too. Bloomberg reported on September 3 that Anthropic was finalizing an expansion of its pre-IPO revolving credit facility to $15 billion, with Morgan Stanley leading the arrangement — six times the $2.5 billion facility of 2025. On the anchor-investor side, a Reuters exclusive on September 11 reported that Nvidia was weighing up to $10 billion, but that item has not been confirmed against a primary source and is carried here only as reported. The underwriters named in coverage — Morgan Stanley, Goldman Sachs, and JPMorgan — sit at the same grade.

On ownership, the grades of certainty split. Google's stake is 14%, confirmed in a court filing, with a contractual cap of 15% and no voting rights and no board observer seat. Amazon's stake, by contrast, is estimated by the press at somewhere between 15% and 21%, with no filing or court document behind it. Written side by side in one paragraph, both read as confirmed; one came out of a document and one is an estimate. This article does not give them equal weight.

Last, the September 19 article set out four things investors will examine during the roadshow: revenue growth, compute costs, customer concentration, and the spending required to train and operate advanced AI models. That list is the article's own framing rather than Reuters' wording, which is worth stating. The third item returns in Section 6 with its direction reversed. If the people putting money into a vendor ask how concentrated that vendor's customers are, the people paying money to the vendor should be asking how concentrated their own dependence is. The timing could also slip further. A delay past the November midterms has been floated, though two sources who spoke to Reuters did not expect the elections to matter much to the listing itself.

The most common way to price a company that has not accumulated profits is to divide its enterprise value by its revenue. The result is called a revenue multiple, and a whole set of instincts rides on it: ten is cheap, fifty is expensive. The division has the same hole in it every time, though. There is one numerator and several possible denominators. Put a different year's revenue underneath and the same price tag yields a completely different multiple. For Anthropic, the spread is more than threefold.

Start with which denominator bankers and investors are actually using. No guessing is needed here. Reuters printed the method in its August 14 exclusive.

The forecast in question is Anthropic's internal projection of $190–200 billion in 2028 revenue — a long-range number the company built for itself, unaudited by anyone outside it. Reuters attached a caveat to the method as well, and that sentence is the single most important line in this article.

Those two sentences draw a line. Dividing by revenue is standard practice; dividing by revenue two years out is not. Arguing that the method is wrong is not this article's job. The reporting already separated the standard part from the less typical part, so the line can simply be carried over. Reuters also noted why this company gets that horizon at all: it has said its run-rate revenue grew more than tenfold in each of the three years through early 2026. Past speed is what licenses the future denominator.

Now put different denominators under that one price tag. In the table below, the column worth studying is not the multiple but the two to its left. Without naming which metric and which year sit underneath, the number on the right means nothing.

The first row is the yardstick we just saw. A GraniteShares analysis published on August 31 did the division itself: up to $2 trillion works out to roughly ten times projected 2028 revenue and about thirty-one times the current run rate. The ~31x in the third row is that figure, and Pebblous ran the same division once three days ago in a piece on how Chinese AI companies report revenue. Here it is cited, not recomputed. The second row puts the projected year-end run rate underneath, and the $110 billion in it is itself not a company disclosure but an investor projection relayed by the Journal. The point is that none of the three is wrong. The same numerator simply met denominators from different years. The analysis that produced the numbers attached its own caveat too. The three comparables' multiples come from LSEG data cited by Reuters. The Anthropic figures are derived from reported valuation and revenue, and the analysis offered them as illustration, not forecast.

A fourth reading from the same period shows what happens when the denominator goes unstated. In an August 13 story, one investor said a company growing 800% a year would conservatively command thirty times revenue, which by that arithmetic gets you to $3 trillion. The figure matches no row in the table above, because the remark leaves out what the thirty was multiplied by. Against the projected year-end run rate it gives $3 trillion; against the 2028 forecast it gives $5.7 trillion; against the July run rate it lands short of $2 trillion. One "thirty times," three different companies.

Moving the denominator from the past to the future changes more than the number, though. It changes who carries the risk. A multiple built on revenue already collected has a settled denominator, which leaves one judgment: what that revenue is worth paying for. A multiple built on revenue two years out has a denominator that has not happened. If the forecast lands 30% short, a ten turns into a fourteen; if it lands at half, a twenty. The multiple doubles without anyone touching the price tag. The party absorbing that gap is the investor buying at the price, and the party that produced the forecast is the company. When Reuters wrote that looking two years ahead is less typical, the sentence was pointing not only at the length of the horizon but at where the risk had moved.

Why not simply check whether the multiple is high or low against other companies, then. Reuters named three comparables and printed each one's multiple from LSEG data. The three figures below are each company's enterprise value divided by its own projected 2026 revenue. None of them is Anthropic's valuation divided by Palantir's revenue.

The most valuable thing in this table is not a number in it but the cell that cannot be filled in. All three companies use 2026 as the denominator. Measuring Anthropic with the same yardstick would take Anthropic's full-year 2026 revenue, and the company has not produced a figure for that year. The third quarter had not even closed as this article went out. So "Anthropic is cheaper than Palantir" and "it is expensive against its peers" are both sentences that do not currently parse — not because the comparison comes out badly, but because one side has no figure measured for the same year. Pushing the listing to November to show third-quarter results first reads as aimed at exactly that empty cell.

The denominator carries one more thing besides a year: the date the value was observed. The 41.6x and the 53x were struck on August 14, 2026, and the valuations behind them had already been moving for months before that. That, too, is not conjecture. The same story says so.

The story recorded, in other words, that the yardstick companies were themselves in motion. So a multiple handed to you raises three questions. What was divided by what, which year is the denominator from, and when was the value observed. Drop any one of the three and the number only looks like a magnitude; it cannot be compared. The next section takes up the first question, because the word revenue sitting in that denominator does not mean one thing.

All three rows in the previous table put revenue in the denominator, and all three were counting something different. The first row is what the company expects to collect two years from now. The two below it are annualized run rates. A run rate stretches the revenue pace at one moment out across twelve months; it is not money that arrived over those twelve months. One piece of coverage on this company spelled out the same definition: the metric reflects the revenue pace at a point in time, not revenue already collected over a year. Given that definition, a steeply rising curve makes the run rate print far ahead of actual revenue.

Start with the trajectory of that run rate. It was $9 billion at the end of 2025, $47 billion in May 2026, and past $65 billion by the end of July. The company stated three of those itself — the $9 billion, the $47 billion, and the $65 billion — while the February, March, and April readings in between come from press reports. As of the same July, OpenAI's annualized revenue was reported at roughly $40 billion. Read that comparison as one of revenue scale only; the two companies' valuations and fundraising are not set against each other here. One more caveat belongs here, and it was attached not by us but by the analysis that printed both run rates side by side: the two companies may not measure revenue the same way, so run-rate comparisons should be read with care. Why such a caveat is needed comes later in this section.

The same analysis carries one more sentence. Ahead of the listing, the company tightened its own definition of run rate. That passage is easy to skim past, and it overlaps exactly with the problem set out above. It means the earlier values on the curve may not have been computed under the same rule as the later ones. Which items were removed, and from when, has not been disclosed, so this article records the fact and leaves the trajectory as reported. Still, reading a growth curve by its shape alone — without asking whether the yardstick that drew it changed midway — leaves no way to tell growth from a change of definition. This is an old problem in data quality, and here it turned up in a company's headline metric.

Now the money that actually arrived. Revenue in Q2 2025 was $787 million. In Q1 2026 it was $4.73 billion, and in Q2 2026 it passed $11.5 billion, ahead of the at-least-$10.9-billion projection Reuters reported on August 14. Hold that single quarter up against its counterpart a year before and the increase is more than fourteenfold. These are preliminary, pre-audit figures, though, and no full-year ratio can be computed from them. Q1 2025 revenue is unconfirmed, so no half-year-over-half-year ratio exists. The two 2026 quarters add up to about $16.2 billion.

The profit side moved in the same quarter too. Reuters put it this way.

First quarterly operating profit here means on an adjusted basis, and preliminary before audit. Even so, the familiar summary — $2 trillion for a loss-making company — does not hold at this moment. Saying anything about profit and loss requires naming the period and the metric together. On an annual basis, exactly one loss figure could be traced back to its source: roughly $5.6 billion for 2024. That one, too, is a tertiary source relaying a document Bloomberg reviewed, so it does not rate highly. The far larger recent-year loss figures in circulation are not used here, because the original could not be found. Numbers that leave opposite impressions of the same company's finances circulate at once for the same reason: the period and the metric are not attached to them.

Overlaying the two metrics in one figure shows where the confusion comes from. In the diagram below, the line on top and the bars underneath use different vertical scales, because they measure two things that cannot share an axis.

Reducing both charts to the same unit makes the distance concrete. A $65 billion run rate at the end of July means money arriving at roughly $5.4 billion a month. Divide the $11.5 billion that actually came in over the three months of Q2 by three and you get roughly $3.8 billion a month. Both are simple conversions we did, not figures the company published. The two differ not because either is inflated but because they were measured at different moments: an average pace across April through June against an instantaneous pace at the end of July, and on a steep curve the later one is naturally larger. But both leave the building wearing the same word — revenue.

For the same reason, setting the roughly $16.2 billion that actually arrived in the first half beside the $110 billion year-end run-rate projection and calling the gap nearly sevenfold means nothing. The first is money collected over six months; the second stretches the pace at some point in December across twelve months, so if that pace holds, the period it points at is mostly the following year. Divide one by the other and you do get an answer. You just cannot say what the answer means. That is why every multiple in this article carries the name of the metric and the year underneath it.

Pebblous took up this property of splitting metrics once already, three days ago, in Chinese companies' filings. There, two revenue figures one company published for the same period stood 6.5x apart. So the explanation is not repeated here. What this article adds is two further layers: the quarterly results we just saw, and the boundary problem that comes next.

It is not only the length of the window that differs. Where a company draws the boundary of what counts as its revenue differs too, and that boundary presses hard on this company for a reason. As Anthropic itself has said, Claude is the first frontier model sold across all three of the largest clouds, with Amazon Web Services as its primary cloud and training partner. When a product sells through several channels, deciding which channel's money counts as your revenue is the same as deciding how large your revenue is. Anthropic is understood to book amounts sold through clouds gross, recording the cloud providers' share as a selling expense. OpenAI is reported to book the equivalent Azure revenue net. US accounting rules settle this split under ASC 606 as principal versus agent: control the good before it reaches the customer and bear the fulfillment risk and you are the principal, so you record gross; take only a brokerage fee and you are the agent, so you record net. Both treatments are permitted under the standard. Neither side is breaking a rule.

The boundary has already produced a fight. On April 13, 2026, a four-page internal memo by OpenAI's chief revenue officer was reported, arguing that Anthropic's $30 billion run rate at the time was inflated by roughly $8 billion by its accounting treatment — about $22 billion if restated net. Anthropic's position is that it is the principal. This article does not rule on who is right. The coverage could not be checked against a primary document, and more to the point, the outcome is not what interests us. What interests us is that competing definitions are running inside one word. When the definition splits, margin comparisons wobble with it, and analysts noted that a reversal during the listing review could force past revenue to be restated.

This boundary is not a side story to the division in the previous section. Every multiple in Section 2 used revenue as the denominator. Whether that revenue is booked gross or net changes the size of the denominator produced by the same business, and a smaller denominator means a larger multiple against the same price tag. An accounting choice looks like a technicality inside the financial statements; in practice it sets the size of the multiple quoted outside them. The listing review is also where that definition gets fixed in public. The September 19 article listed, among the things the public filings will reveal, whether the cloud providers are a large share of this company's costs or a large share of its revenue. Putting cost and revenue in one sentence was not sloppiness. It is an accurate description of the structure, because the same counterparty sits on both sides. Once the registration statement appears, the table in Section 2 will have to be read against whatever definition it sets.

Growth itself supports two readings at once. From one side, the 2028 forecast assumes deceleration. If the run rate really grew more than tenfold in each of three straight years as the company has said, extending that pace to 2028 would put the forecast far higher. At $190–200 billion the curve is much flatter than that. From the other side, calling it deceleration is premature. Q2 beat the projection published a month earlier, and the absolute dollars added each quarter are the largest on record. Several observers have argued that the growth rate is slowing, but this article could not verify those observations well enough to put numbers on them. Carrying both readings is the accurate thing to do for now.

So where is the $190 billion of 2028 supposed to come from. The company's revenue runs along three tracks: subscriptions, API usage by developers, and enterprise contracts. Most of it originates in organizational work rather than personal hobbies — software development, research, customer support. Which means the 2028 revenue in that denominator is the sum of how much of their own work various organizations hand to this tool. One of those organizations may be the company reading this.

The clearest statement of that structure came from an investor. Here is the sentence from Sequoia's Alfred Lin in the May Series H announcement.

That is the investment thesis compressed into one line, because what holds the valuation up is not a model's test scores but the context and the procedures held inside the organizations that buy it. The half usually quoted is the second one, but the causation lives in the first. The act of handing complex work over is itself the learning path. The side doing the learning gains the value; the side doing the teaching is the organization that handed the work over. This is an investor's assertion, though, not a measurement. An assertion and a measurement do not carry the same weight here.

Some figures do come closer to measurement. Claude Code, the coding agent, went from a $1 billion run rate in November 2025 to $2.5 billion in February 2026, and the company has said more than half of that revenue comes from enterprise use. Customers spending over $1 million a year passed 500 in February 2026 and went beyond 1,000 by April. Customers above $100,000 a year grew sevenfold in a year. Eight of the Fortune 10 are customers. The product line is moving outward from developers, too: on top of a family that started as a coding tool sits a product aimed at general knowledge work. The direction in which the money grows is unmistakably inward, into organizations.

What actually happens inside those organizations, the company studied and published itself. The research came out in June 2026 and aggregated roughly 400,000 sessions from 235,000 users between October 2025 and April 2026, with personal information stripped. The most-quoted sentence concerns the division of labor: humans make most of the decisions about what to do, and Claude makes most of the decisions about how. The study attached numbers to that split as well. Sorting the decisions in a session into planning and execution and counting who made each, humans made about 70% of the planning decisions and only about 20% of the execution decisions.

What the sessions were for shifted over those seven months too. Sessions spent fixing broken code fell from 33% to 19%, and the work around the code took that space. Standing software up and running it went from 14% to 21%, and documentation and data analysis went from about 10% to about 20% — that second pair is the one that doubled. The value of the work rose alongside it. The researchers' estimate of what each session's work would fetch at freelance-market rates rose 27% on average over the same seven months, with building, operating, and fixing all up by more than 30%. The study cautioned that these figures should be read as comparisons between two points in time rather than as dollars.

One problem follows from that planning-versus-execution split. A plan a person made survives in meeting notes and approval documents; where does handed-over execution survive. If the record of which files were opened, which values were changed, and why that approach was chosen exists only in logs outside the organization, then the deeper the adoption goes, the less the organization can explain its own work from its own records. That is also why the third of the three layers for measuring adoption depth is the review rate. The share a human looks back over is quality control, but before that it is the procedure that keeps the record on your own side. The problem returns in Section 6 as a line in a contract.

The study's limits are stated by the company itself. Whether work that looked successful inside a session was actually used or discarded outside it was not observed. The classification also relies on a model reading conversation transcripts. And one more limit matters most for this article's purposes. The research covers interactive use only, and the non-interactive use it leaves out accounts for a substantial share of total activity. Only exchanges typed by a person at a screen were counted. What enterprises wire into pipelines and run automatically is not in there, so the real depth inside organizations could be either greater or shallower than this research shows.

The company has published two further documents, and because they are different in kind they have to be read apart. One is a forward-looking report setting out eight expected trends in agentic coding for 2026. It says up front that it describes what the company is seeing among customers today rather than asserting anything about tomorrow. It contains one widely quoted gap: developers use AI for roughly 60% of their work, while the share they say they would fully delegate stays between 0% and 20%. The report states that this figure came from the company's Societal Impacts research rather than from the report itself. The point is the size of the gulf between using and delegating — and that a vendor recorded that gulf in its own document.

The other document is a survey. Conducted with the research firm Material in late 2025 among more than 500 US technical leaders, it holds the figures closest to what this article calls adoption depth. More than nine in ten responding organizations use AI in coding, and 86% are past experimentation and running coding agents on code that reaches production. That rate splits into 91% at enterprises and 83% at smaller companies. And 42% say they would trust these agents to lead development work under human oversight. Better to keep that distinct: it is a statement of trust, not a record of work already handed over. The timing matters too — late 2025, roughly nine months back from now.

Put those numbers in one place and adoption depth resolves into three layers. How far the agent is connected, how much it does on its own, and how much a human looks back over. The three layers measure different things, and not every layer has a measurement for it.

Here the article stops for a moment, and marks precisely where it is stopping. That depth increases the model's workload is genuinely measured. In the same research, a newcomer's single instruction produced an average of five actions and about 600 words of output, while an instruction from someone who knows the domain produced twelve actions and 3,200 words. Controlling for task type and value, month, occupation, and model family, each step up in expertise added 9% more actions and 13% more output. Decision rights move the same way: when the human holds most execution decisions Claude takes about eight actions per turn, and when Claude holds the planning as well that rises to sixteen. The company's economic index put a figure on input context as well — a stable relationship across many tasks in which a 1% increase in input context length brings a 0.38% increase in output quality and length. Output length is the quantity that gets billed.

So what is missing is not measurement but an independent check connecting those measurements to the vendor's revenue. Every figure just cited came from the vendor's own usage records, and the vendor also designed the measurement — bundling quality and length into a single variable, for instance. We found no outside source establishing how much a contract grows when one more system is connected. Which makes "the speed at which agents move into enterprise data produces the 2028 forecast" an assertion whose direction is corroborated from several sides, not a causal link verified down to the numbers. Leave that distinction unwritten and the assertion hardens into fact.

There is another frame for reading the forecast, worth recording. The analysis that calculated the multiples earlier reported that the company presents its 2028 projection as a dated construction schedule of contracted compute. Framed that way, progress can be checked from outside — and if construction runs late or demand falls short, that shows up in the results immediately as well. Adoption depth and the construction schedule are the two legs holding the same forecast up, and the second leg reappears on the cost side in Section 5.

What blocks the depth is something the same survey asked respondents directly. The most cited obstacle was integration with existing systems at 46%, followed by implementation costs at 43%, data access and quality at 42%, and change management at 39%. Among smaller companies, human problems such as staff resistance and training stood out more. The report's own conclusion is blunter: the biggest barrier to getting value from AI is neither model capability nor cost. It is organizational readiness, and within that, data accessibility and the work of assembling context. The company's economic index points at the same place — enterprises whose data is fragmented across systems cannot unlock complex use cases. The data modernization cost of retrieving context may be the main bottleneck for adoption. Enterprise and personal use are also not alike. The same report also analyzed 3.5 million anonymized conversations. In business API traffic, 77% took the form of automation, handing over a task whole; in personal use that share sat at around 50%. Conversations that delegate a task entirely rose from 27% to 39% over eight months.

That this lands on something Pebblous has been saying for years is worth stating plainly. The diagnosis that what limits depth is data rather than models is not our reading; it is the vendor's own conclusion. What we can add, then, is not the diagnosis but what follows from it. If the diagnosis holds, the party the valuation rests on is not the vendor but the organizations holding the data, and what those organizations should check is written out in Section 6.

Before that, there is more to look at. Pebblous has traced the concrete shape of that interior several times. One piece examined how coding agents actually read internal documentation; another covered a case where dozens of sales-system functions moved inside the model. From the opposite direction, a third catalogued six data defects that stall agents in pilot. What the three share is a picture of how that diagnosis looks on the ground. The place things jammed was never the model. It was the records on the side doing the connecting.

So what comes out of it is not a prediction. It is a way to locate where you stand. For the 2028 denominator to hold, the organizations now running pilots have to descend these three layers one at a time. If your company is one of them, you are a line in that forecast. Being a line is not in itself a bad thing. The question is whether your terms are written on the same line.

If the 2028 forecast breaks, where does it break first. The easy guess is the point where a rival model gets smarter. But companies cut AI spending not by switching vendors but by dropping to a cheaper tier. That downgrade is already showing up in data published month by month.

The most frequently cited source is the monthly index published by Ramp, the corporate card and spend management company. One caveat has to travel with the metric. What Ramp publishes is the share of businesses that actually paid a given vendor — not revenue share and not market share. Since one company can pay several vendors at once the categories are not mutually exclusive, and a single $20-a-month seat counts the same as an eight-figure contract. Strip that caveat away and the same table tells a completely different story.

April is the month the order flipped. What came after it is the more interesting part, though. Four months past the crossover, the gap still sits near four percentage points. That is not one vendor pushing the other out. It looks like most companies running both and deciding, month by month, which one to spend more with. In that arrangement, the side that raises prices is punished first.

Look at the rate of climb rather than the gap and the table reads differently. In July, Anthropic added 1.1 percentage points in a month; in August it added 0.34. OpenAI added 0.23 and then 0.09 over the same two months. Ramp described this as new corporate AI adoption still rising but decelerating. For the 2028 forecast, that matters more directly than which vendor is ahead. If fewer new companies arrive, the remaining growth has to come from companies already using the tools spending more. Ramp's head economist said the same thing: first-time AI buyers still pick the American model companies, but future growth has to come from existing users, especially heavy ones.

There are places where the punishment has already been recorded. The rows below all point the same way.

Even with the best model on the shelf, businesses put barely a tenth of their budget on it. The rest went to cheaper tiers. And the average price falling more than forty percent in a matter of months is not only a matter of vendors cutting rates. It is also the result of companies shifting the same work onto cheaper models. Pebblous covered one side of this before, in how a falling unit price still produces a bigger bill; this data shows the other side. When unit prices fall and companies downgrade tiers at the same time, the two effects stack.

How the downgrade happens is in the data too. Ramp reported that some companies are cutting frontier-model usage by setting a company-wide default, on the grounds that the standard tier does the job well enough and costs less. This is a procurement and finance decision, not an individual developer's preference. A default, once set, presses on spending every month without anyone having to be persuaded, and undoing it requires another approval. The conclusion Ramp drew is correspondingly sharp: they had found a ceiling on what businesses are willing to spend on AI, and above that line better performance does not pay for itself. Raising the line again would take proving performance clearly beyond today's frontier models while keeping competitors nowhere near — which, with cheap open-weight models arriving a few months behind, the index argued, looks increasingly hard.

The same economist revised that diagnosis a month later, and that belongs here too. The squeeze in spending does not appear to be the result of switching to open-weight or Chinese models. Companies taking that route account for 6.4% of businesses spending on AI, and even that measurement counts routes that also broker closed models, so the true figure may be lower. What is pressing on spending, in other words, looks less like vendor switching than like falling prices and downgraded defaults.

The distribution of spending shows where the forecast's weight rests. Median monthly spend per employee is $7,205 at companies in the top 1%. At the top 10% it is $650, and at the median company it is $11.95. The first figure is from August 2026 and the other two from July, so they are not the same month. The orders of magnitude are unambiguous regardless. Today's revenue curve is being pulled by a very small number of deeply committed organizations, while most companies still spend about the price of three coffees a month. For the 2028 forecast to hold, that middle group has to move up. Moving up is what descending the three layers in Section 4 amounts to.

In August, though, the movement went the other way. Per-employee spend at top-1% companies fell 9.7% in a single month, from $7,976 to $7,205. Since that group pulls most of the model companies' enterprise revenue, this one figure tests the paragraph above it head-on. Ramp attached several caveats. The top 1% is a small set of companies, so the value swings hard. The July figure is itself a revision upward, from about $7,400 to about $8,000, as more transactions came in, and August brings a seasonal decline as developers take vacation, similar to what shows up every November and December. Reading one month as an inflection would be hasty, then. What is worth watching here is not whether the value rose or fell. The metric capable of testing the 2028 forecast is published monthly, and it bends under things like seasonality and a company-wide default setting. The denominator two years out gets filled by twenty-four months like these.

The competitive side is not quiet either. Reuters reported that Anthropic is weighing additional model releases as OpenAI's latest model gains traction with enterprise customers. Figures purporting to set the two companies' shares of enterprise AI spending side by side circulate in re-reported coverage, but this article does not use them. That story is taken up separately in Section 7.

The scale of the cost side can be gauged from the company's own announcements. Anthropic has said it will have access to up to 5 gigawatts of compute with Amazon Web Services through the end of 2026, spending more than $100 billion over ten years. Another 5 gigawatts is scheduled with Google and Broadcom starting in 2027. The Series H announcement added one more: access to GPU capacity at SpaceX's Colossus 1 and 2. One of the three companies used in Section 2 as a yardstick for pricing this one reappears in the same paragraph as a compute supplier. When the comparable and the counterparty overlap, how independent the comparison is becomes its own question. On top of that, compute is projected to go from roughly 5 gigawatts at the end of 2026 to nearly double that by the end of 2027 — an investor projection relayed by the Journal rather than a company statement, so it carries a different grade. What investors are betting on comes down to one thing: that revenue grows faster than these costs and the margin widens. A projected rise in gross margin from about 50% in 2025 to about 77% in 2028 is the same bet in another form, and it too comes from the investor side.

Setting two numbers side by side shows where the bet is asymmetric. More than $100 billion committed over ten years averages out to a little over $10 billion a year, which looks small next to a $190 billion revenue forecast for 2028. But one is signed spending and the other is unconfirmed income. And if compute doubles within two years, that money goes out ahead of the revenue. The logic holding the price up hangs on a single sentence — revenue grows faster than costs — and in that sentence the cost side is already written into contracts while the revenue side is not written anywhere yet. Pebblous traced the path capital took into the inference layer once before. If that piece asked which layer the money went to, this one asks which year's revenue it is scheduled to come back through.

Forces slowing growth from inside the company have been reported as well. Chief executive Dario Amodei is said to have pushed to slow the pace of model releases even while preparing the listing. The request is grounded in safety, but slower releases slow commercial growth with them. Around August, several things were cited at once: that the flagship model costs more than competitors', cheaper Chinese alternatives, an argument that the Commerce Department's June export controls held back that month's revenue growth, the dispute with the Defense Department, and customers' sensitivity to AI costs. Pebblous covered the Defense Department case separately, and the question of evaluator independence sits in another piece. Here they stay a list.

Finally, there are precedents: companies that listed on a price built from forward projections, and what happened next. Reuters cited two in the same story. Cerebras investors pointed to 2028 revenue expectations ahead of its listing, and SpaceX's projections ran out to 2029. Cerebras went public on May 14 at $185 a share, rose 68% on the first day, and by mid-September had come back to 5–8% above the offer price — close to half off its high. SpaceX also swung widely after listing at a record valuation. The business models differ too much for a direct comparison, and the only thing being carried over here is the size of the swings. A price set on a forecast in the denominator gets re-set after the listing.

A remark from one investor Reuters quoted is the right place to close this section. David Merkel of Aleph Investments put it this way.

Merkel added one more question in the same interview: whether this really generates that much additional productivity. Anyone setting a price, Merkel said, ought to be asking it. The answer to that question is not in the vendor's results. It is in the books of the organizations using the tool.

Reduced to a sentence, everything so far comes to this. The $190 billion of 2028 is money nobody has earned, and a large share of it has to come out of next year's budgets at organizations currently running pilots. So the document to check, for an organization reading this listing story, is not the vendor's results. It is your own contract.

Reverse the direction and the checklist is already written. Recall the diligence list from Section 1. Investors putting money into this company look at how fast revenue grows, what compute costs, how concentrated the customers are, and what it takes to train and run the models. If the people funding a vendor ask how concentrated that vendor is in one or two customers, the people paying that vendor should be asking how concentrated they are in one supplier. It is the same question, pointed the other way.

The six items below are not a new industry standard. They are problems Pebblous has covered separately over the past few months, rewritten as lines in a contract. Each one links to the piece that dealt with it.

Read the six again and not one of them asks about model performance. They are all questions about contracts, records, and portability, and answering them rarely requires contacting the vendor at all. Nor is the list an instruction to distrust the vendor. Investors on the other side are putting the same items to the vendor's books, so this is closer to applying equal diligence to your own.

Pebblous wrote earlier about startups being judged in diligence on the state of their data management. That piece and this one run in opposite directions. There, the company raising money was the one being assessed; here, the company using the tool is the one becoming a line in a vendor's forecast. What is striking is how little the two checklists differ — what to prepare when you are being assessed, and what to verify when you are the thing a valuation rests on. Both end up asking where your data sits and in what condition.

No single department can fill in all six. Items 1 and 5 are held by the engineers who did the integration, items 2 and 3 by legal and data management, items 4 and 6 by procurement and finance. In the room where adoption was decided, those three usually sat apart. That the others' cells were empty only comes out after the contract is signed. The usefulness of the list, then, is less in filling the answers than in getting the three sides in front of the same table once. Just identifying which cells are blank settles which clauses to negotiate at the next renewal.

Which leaves the opening question intact. Most organizations can work out which line they are on. The hard part is saying where their own terms are written on that line. If you cannot answer that, the document to open right now is not the vendor's registration statement. It is in your own desk drawer.

The price tag on this listing hangs on one denominator: revenue arriving in 2028. For that denominator to hold, agents have to keep moving into companies' context, procedures, and judgment — and that interior is where Pebblous works. Across six sections, the work was not forecasting but checking the denominators one at a time. Four reasons why that is our work.

7.1Define the yardstick before the grade

In DataClinic we attach a grade and a defect type to every record. The first step in that work is not assigning grades but defining what the grade measures. Figures accumulated without a defined yardstick look more precise the more of them there are, which is exactly what makes them dangerous. This case is the same work applied to financial metrics. The "N times" hanging off $2 trillion means nothing without the name of the metric and the year underneath, and those two are the first things to fall off as the number gets re-reported.

7.2Four values circulating under one word

In this case the word revenue points to four: annualized run rate, quarterly results, fiscal-year revenue, and the 2028 forecast. Add the gross-versus-net boundary from Section 3 and the same business shows up at several different sizes. Each pass through a re-reporting outlet drops one more caveat, and that is exactly what happened here.

One re-reported article set two numbers side by side under the label of enterprise AI spending share. What the original data provider had published were two different yardsticks: the share of businesses paying a given vendor and the share of spend one model accounts for within a single vendor. Neither can stand two companies next to each other. That is why Section 5 does not use them. Figures that have lost their metadata look more precise the further they are re-reported, which is the lesson this case leaves, and it is why AI-Ready Data insists that a value travel bundled with its definition, its reference date, and its aggregation scope. What happens to training data happened here to financial metrics — nothing more exotic than that.

7.3The records to check are usually already inside the organization

When a customer says "we're adopting agents too," the thing to verify is not a model performance table. The six cells in Section 6 are the list: what the agent reaches, whether our data is used for training and how long it is kept, whether agent actions are logged and can be deleted, how price changes are notified, whether the work is portable, and concentration in both directions. All six are data and record questions rather than model questions, and all six are work Pebblous already does with customers. The records needed to answer them mostly sit in your systems rather than the vendor's. If they do not exist, that absence is the first finding.

7.4Where Pebblous stands

Pebblous does not take the position in this article that $2 trillion is a bubble, or that it is cheap. Where we stand is on opening up the yardstick attached to a price. Which year's revenue it was divided by, which metric sat in the denominator, when the value was observed, and whose what would have to be true for that denominator to hold. That the last question's answer sits on the reader's side is this article's turn, and the turn happens when the work done on somebody else's financial metric gets done once more on your own organization's.

The facts here come from four streams. The pricing method and the verbatim quotations come from the Reuters exclusive of August 14 and from company announcements; the enterprise spending and token prices from figures the data provider published directly; the adoption-depth numbers in Section 4 from research and surveys the vendor produced itself; and the September 19 timing and the year-end projection from an article re-reporting The Wall Street Journal. The last stream could not be checked against the original and the one before it consists of vendor-produced figures, and the text says so at both points.

Primary reporting and company materials

• 1.Echo Wang. "Exclusive-Anthropic IPO valuation hinges on $190-200 billion 2028 revenue forecast, sources say." Reuters, 2026-08-14. The primary source for this article. The two sentences stating the pricing method, the three comparables' multiples (LSEG data), the Q2 projection and first quarterly operating profit, the Cerebras and SpaceX precedents, and David Merkel's remarks all come from this story.
• 2.Anthropic. "Anthropic raises $65B in Series H funding at $965B post-money valuation." Company announcement, 2026-05-28. The source of the Alfred Lin quotation in Section 4 (in full, including the opening clause), the $47 billion May run rate, the Amazon, Google/Broadcom, and SpaceX compute commitments, and the fact that Claude sells across all three major clouds.
• 3.Anthropic. Series G announcement, 2026-02. The source of Claude Code's $2.5 billion run rate, the enterprise share of that revenue, and the count of customers spending over $1 million a year.
• 4.Anthropic. "Agentic coding and persistent returns to expertise," 2026-06-16 (often referred to as "How Claude Code is used in practice"). The source of the roughly 400,000 sessions and 235,000 users, the 70%/20% split in planning and execution decisions, the shifts in task composition and task value over seven months, and the action and output volumes by expertise level. The limit that non-interactive use is excluded is also stated in this document.
• 5.Anthropic. "2026 Agentic Coding Trends Report." Carries the gap between using AI for roughly 60% of the work and fully delegating 0–20% of it, though the report states that the figure came from the company's Societal Impacts research. The report describes itself as a forward-looking document for 2026.
• 6.Anthropic and Material. "The 2026 State of AI Agents Report." A late-2025 survey of more than 500 US technical leaders. The source of the 86% running agents on production code (91% enterprise, 83% SMB), the 42% who would trust agents to lead development work, the adoption barriers at 46%, 43%, 42%, and 39%, and the conclusion that the bottleneck is data accessibility and context assembly rather than model capability. The economic-index analysis in the same document (3.5 million conversations) is the source of the 77% automation share in business use and the relationship between context length and output.

Original sources for the spending data

• 7.Ara Kharazian. "Cracks in the AI Thesis." Ramp AI Index, 2026-08-12 (July data). The source of the July payment rates and the month-over-month increases, the top model's share of tokens and spend within Anthropic, the roughly $10 per million tokens, the distribution of per-employee spending, and the conclusion about a ceiling on what businesses will spend on AI. The caveat that the sample skews more technical than a general index sample, so real adoption may be lower, is also the author's.
• 8.Ara Kharazian. "Cracks in the AI Thesis Part 2." Ramp AI Index, 2026-09-09 (August data). The source of the August payment rates and the deceleration, the 9.7% drop in top-1% per-employee spending and its caveats, the 41% fall in token prices, the 53%-to-45% shift in frontier-model token share, the company-wide default settings, and the open-weight adoption rate.
• 9.Axios, 2026-05-13. Reported the figures for April, the first month the payment rates flipped.

Re-reported coverage and analyses

• 10.Lawrence Mondal. "Anthropic targets November IPO at potential $2 trillion valuation." crypto.news, 2026-09-19. An article re-reporting The Wall Street Journal. The November delay and its reason, the $110 billion year-end run-rate projection, the compute expansion projection, and the diligence items were confirmed through this route. The definition of run rate and the statement that no public registration statement has been confirmed are also in this story.
• 11."Anthropic IPO 2026 Explained: From $965 Billion to a Possible $2 Trillion Listing." GraniteShares, 2026-08-31. The ~10x in the first row and the ~31x in the third row of the Section 2 table are both GraniteShares' own calculations, and the caveat that they are illustrative rather than predictive came from the analysis itself. Also the source of the confidential filing and the 15-day public-filing rule, the statement that the company tightened its run-rate definition, the roughly $5.6 billion loss in 2024, the description of the 2028 forecast as a contracted compute construction schedule, and the caveat attached to the run-rate comparison with OpenAI.
• 12.Cris Tolomia. "Anthropic investors target $2 trillion IPO valuation in October." Quartz (via Yahoo Finance), 2026-08-13. A snapshot from the day before the Reuters exclusive; the timing changed afterward. The source of the investor remark in Section 2 about thirty times getting to $3 trillion, the year-end projection of $100–120 billion against $47 billion in May, the three underwriters, and the list of headwinds cited in August. The article itself rests on Financial Times reporting.
• 13.Bloomberg. "Anthropic Finalizing $15 Billion Pre-IPO Credit Facility," 2026-09-03. The source of the revolving credit facility expansion in Section 1. The reporting on Q1 and Q2 2026 results also rests on documents Bloomberg reviewed.
• 14.Josipa Majic. "OpenAI And Anthropic Count Revenue Differently." Forbes, 2026-03-25. The source of the gross-versus-net treatment in Section 3. This article treats that material at summary grade and quotes none of it verbatim.
• 15.FASB, ASC 606 "Revenue from Contracts with Customers," the provisions on principal-versus-agent determination. The accounting description in Section 3 refers to this standard.