Politics & institutions
Feeding the Elephant
Anthropic has bought its food years in advance. Can each dollar of it keep earning enough?
Read Here
From Heavy Lifting · Public feed content · Heavy Lifting
This piece is co-authored with Shivaram Rajgopal and available on The Googly.

Napoleon’s elephant for the Place de la Bastille, as its architect, Jean-Antoine Alavoine, drew it: a bronze fountain about 24 meters high, to be cast from captured cannon. Conceived in 1808, it never got past a full-size plaster model, which stood from 1814 until 1846, crumbling, harboring rats its neighbors petitioned against, and sheltering Gavroche in Les Misérables. The commitment outlived the empire that made it.
Jean-Antoine Alavoine, final design for the Elephant fountain, Place de la Bastille, 1809–1810, watercolor, countersigned by Vivant Denon. Musée du Louvre, Paris. Public domain, via Wikimedia Commons.
In our first piece we asked what kind of animal Anthropic is and found evidence for three. A spider builds a product once and sells it cheaply many times. An elephant buys enormous capacity in advance and must fill it. A shark must keep producing, because whatever it made last soon loses its customers. The evidence pointed to the shark, but the only species whose required growth looked even close to achievable was the elephant.
This piece examines the elephant. Elephants are judged by one thing—whether they can eat what they have stored. Anthropic has committed about $895 billion to its suppliers, against $13.8 billion its customers have committed to it, and by 2028 its annual compute bill will be roughly three times its current rate. The ratio at the center of this piece is one Anthropic now uses itself: revenue per dollar of compute. It was $2.08 in the second quarter of 2026 and $1.72 in the third. To keep the 45 percent margin after compute on which the elephant’s valuation in our first piece rests, each dollar must bring in at least $1.82, which is exactly what it brought in, on average, over the first half of the year. So, the question this piece asks is simple. Can revenue per dollar of compute get back above $1.82 and stay there while the compute behind it roughly triples?
The third quarter also changes the picture of the elephant. Until now the worry about Anthropic has been whether it could use all the capacity it had bought. In the third quarter it could not get enough. Demand outran the compute it had contracted, so it bought $5.0 billion of compute on short-term contracts that, in the words of its amended filing, “currently cost significantly more” than its long-term ones, and that premium is what pulled the ratio below the bar. Anthropic expects to buy more short-term compute in the near term. For now, the elephant is not struggling to eat what it stored. It has run short and is buying at the corner store until the larder fills. Whether the stored food proves a bargain or a burden depends on what compute is worth when it arrives.
Airlines, which also commit to capacity years before they know the demand, have a way of answering that question. They split revenue into three parts: capacity (seats flown), load factor (the share of seats filled by paying passengers) and yield (what each paying passenger brings in). Anthropic’s revenue splits the same way. Capacity is the compute it has bought. Load factor is the share of that compute serving paying customers. Yield is what each unit of that serving compute earns. The sections that follow take each term in turn, then ask who else takes a share and whether the product clears the bar. Every figure comes from the S-1 and its amendment of November 2 unless we say otherwise.
Capacity: Food Bought In Advance
This is an industry in which capacity has to be locked in well before demand arrives. Anthropic buys compute 12 to 18 months before it is delivered, and the data centers behind it need power, equipment, and permits that take longer still. Table 1 lists the largest commitments. In all, Anthropic has committed about $895.4 billion, roughly half of it, $447.5 billion, before the end of 2030 and most of the rest between 2031 and 2045. That total leaves out a further $150–175 billion of compute it has agreed to buy without a fixed price. Against them stands $13.8 billion of remaining performance obligations at September 30, the contracted revenue Anthropic has yet to recognize, about three-quarters of it due within a year.

Almost all of it is take-or-pay. The contracts with Google, for at least $121.3 billion, and Amazon, for $110 billion, require Anthropic to spend those sums “regardless of actual usage” and to pay any shortfall. Microsoft’s contract uses the same words. The Broadcom leases, about $161.2 billion of rent over five years for tensor processing units (TPUs—the AI chips Google designs and Broadcom builds), cannot be canceled except on default, are backed by cash in escrow, carry cross-default clauses, and can begin charging rent before a data center is ready. Nscale and Lambda charge only as capacity is delivered, but once it is delivered it cannot be canceled either. The largest exception is SpaceXAI, the AI unit of SpaceX and formerly Elon Musk’s xAI.
Anthropic has agreed to spend up to $120.5 billion with it, about $3.1 billion a month once all the capacity arrives in December, and after initial periods ending in January and March 2027 either side can cancel on 90 days’ notice. In all, about $213.8 billion of the commitments can be terminated, or depend on conditions such as a supplier securing financing or a site before they begin.
Spread evenly over their terms, the contracts come to roughly $105–120 billion a year once all the capacity has arrived, around 2028. That is our estimate, and actual payments will be lumpier, but it is consistent with the company’s own figure of $447.5 billion before the end of 2030. In the third quarter Anthropic was spending on compute at a rate of about $39 billion a year, half of it on short-term contracts. About $70–80 billion of the 2028 bill is a fixed core that cannot be canceled; SpaceXAI adds about $37 billion that can. And the larder is still being stocked. Since September 30 alone, Anthropic has signed about $100 billion of new commitments.
What does that buy, in physical terms? The S-1 does not say, so we speculate. Recent contracts suggest that renting a gigawatt of AI computing for a year costs something like $12–20 billion. On that range, Anthropic was paying for roughly 2–3 gigawatts in the third quarter and will have roughly 5–10 in 2028. The conversion is uncertain, but nothing below depends on it, because the test is set in dollars: revenue per dollar of compute.
Nor is the bill a one-time purchase. Because Anthropic mostly rents, its suppliers bear the chips’ depreciation inside the price. But the contracts run about five years, and chips stop being worth running long before they stop working. When the contracts end, the capacity must be bought again, at whatever it costs then. Like an airline’s fleet leases, the bill is a permanent cost of flying, a form of maintenance capital spending that never appears as capital spending in Anthropic’s accounts.
Load Factor: Who Eats the Food?
Most of the food does not go to paying customers. Cost of revenue was $4.8 billion in the first nine months of 2026, against $18.8 billion of compute and infrastructure, so at most about a quarter of compute served paying customers. In airline terms, that is Anthropic’s load factor: only about one seat in four carries a paying passenger. But the other seats are not empty. They are taken by research, the training of new models, free users and Anthropic’s own staff, in proportions the S-1 does not give. Figure 1 shows why: the seats that earn nothing today are where the next model is built.


An ouroboros, the serpent that feeds on its own tail, in a Greek alchemical manuscript copied by Theodoros Pelecanos in 1478. Anthropic’s loop differs in one respect: the serpent only has to keep eating itself, while the loop has to take in more at every turn.
Drawing by an anonymous illuminator in a copy, made by Theodoros Pelecanos in 1478, of an alchemical tract attributed to Synesius. Bibliothèque nationale de France, Codex Parisinus graecus 2327, fol. 279. Public domain, via Wikimedia Commons.
An airline’s empty seat earns nothing. Anthropic’s spare compute is never idle. The same chips that serve customers in the afternoon can train the next model at night, and Anthropic moves compute wherever “the marginal unit is expected to be most productive.” But compute spent on training earns nothing until the models sell. For an elephant, the question is not whether the chips are busy but whether revenue pays for what they do.
That makes the load factor the place where the second kind of maintenance spending hides. Warren Buffett defined maintenance capital spending as what a business must spend just to keep its competitive position and volume; anything beyond that is growth. Training that merely keeps Anthropic level with its rivals is maintenance. Training that opens new kinds of work is growth. The three margins in our first piece are three answers to that split. The 85 percent gross margin treats all training as growth. The 45 percent margin after compute and the 26 percent margin after R&D treat it all as maintenance. The S-1 gives no split, but the next section suggests where it lies.
Yield: The Red Queen’s Race

“Now, here, you see, it takes all the running you can do, to keep in the same place.” The Red Queen to Alice, in John Tenniel’s illustration for Through the Looking-Glass, 1871.
John Tenniel, engraved by the Dalziel Brothers, in Lewis Carroll, Through the Looking-Glass, and What Alice Found There, Macmillan, 1871. Public domain; scan from the Victorian Web.
Revenue is quantity times price, and for Anthropic the obvious quantity is tokens, the fragments of words a model reads and writes. But a token is not a standard unit. An Opus token costs several times a Haiku token; a reasoning model can spend thousands of tokens thinking before it answers; and a newer model may charge more per token yet need fewer of them. That is why the S-1 reports how much each new model lowers the cost of a typical task, rather than the price per token, and why yield is better measured per dollar of compute than per token. Counting tokens is like counting airline passengers without asking how far they flew.
Yield is under constant pressure. Models released in the six months to September 30 produced more than 80 percent of third-quarter revenue. In 1973 the biologist Leigh Van Valen proposed that species must keep evolving just to hold their place, because their rivals are evolving too, and named the idea after the Red Queen. A Claude model does not wear out; it loses customers when something better or cheaper appears. That is strong evidence that most training is maintenance, spent to stay in place, which puts Anthropic nearer the shark.
Each lap of the race cuts yield directly. Four-fifths of Anthropic’s revenue is billed by usage. Fable 5.1 and Opus 5.5 cut the cost of a typical task by about 25 and 40 percent, and for usage-based revenue a cut in the cost per task is a cut in revenue per task. After a 40 percent cut, customers must do two-thirds more work just to keep revenue level. The S-1 says the cuts preserved margins, which is a statement about margins, not revenue. The amended filing puts a number on the revenue. Price cuts in new models and contractual discounts took about $600 million off it between the second and third quarters of 2026. New chips push yield back up by doing more work for each dollar of compute, although they also cost more to rent. And usage can grow in two ways. Cheaper tasks bring more of the same tasks, much as more efficient steam engines raised Britain’s use of coal in Jevons’s day, and better models bring new kinds of work. The second is stronger, and it depends on staying at the frontier.
The race may be narrower than the 80 percent suggests. Shiva, who often uses older models to save tokens, notes that customers sort themselves: Haiku for routine volume, Sonnet for everyday work, Opus and above for the hardest tasks. Only the top tier is a race. The lower tiers drift toward commodity prices, contested by open-weight models that anyone can run, and the 80 percent partly reflects Anthropic’s own choice to make each new model the default. Geography splits the market too. Customers billed in the United States provided 72 percent of revenue in the first half of 2026, although most consumer use of Claude comes from abroad. Yield depends on whether the top tier, in the richest markets, will pay.
So far usage is outrunning the cuts, but by less and less. Quarter-on-quarter revenue growth reached 144 percent in the second quarter of 2026 and slowed to 47 percent in the third, when revenue was $17.0 billion. Ramp’s data on business AI spending show the same pattern. In the four weeks to October 11, volume across all providers rose 52 percent while spending rose 23 percent, even as Anthropic’s share of spending rose from 54 to 60 percent. Usage still outruns price, but spending is growing at less than half the pace of volume: yield is falling.
Who Else Feeds on the Yield?
In the first piece we described CoreWeave, which rents out AI capacity, as an elephant investors prize because it holds roughly $100 billion of signed contracts from its customers. Anthropic is its mirror image. Its suppliers hold its signed contracts, and it holds very few from its customers. That overstates the mismatch somewhat. Usage-based revenue never enters the backlog however reliable it prove; third-quarter revenue alone, at an annual rate, is about five times the backlog, and while compute is scarce, guaranteed supply is an asset. But it shows who bears the risk if yield falls short. The suppliers are protected by contracts. Anthropic is protected only by its customers using more, by the cash it has raised, and by the one contract it can cancel.
The third quarter shows what that asset can be worth. A company that cannot get enough compute pays a premium for it, and a rival without capacity cannot serve customers at any price; the S-1 itself warns that competition may come to depend on “access to large-scale compute rather than on model development capability.” Contracts that lock in capacity at today’s prices are then a moat. But they are also a bet. A take-or-pay contract fixes the price for five to twenty years. It is an asset as long as compute stays scarce and its market price stays above the contract price, and a burden if the industry builds more than it needs and the price falls. Every leading lab and cloud company is making the same bet, so if it goes wrong, it goes wrong for all of them at once. Anthropic is short of compute today. The question is whether it will be long when $447.5 billion of capacity arrives by 2030.
The same suppliers also take a share of the yield. The S-1 describes Amazon, Google, and Microsoft as companies “who simultaneously serve as our investors, customers, distribution partners, and competitors.” Their cloud marketplaces handled 44 percent of revenue in the first nine months of 2026, up from 11 percent in 2023. From the fees Anthropic paid all channel partners, we estimate that the marketplaces keep up to about a third of what passes through them, a cut taken before Anthropic’s yield reaches its own accounts. Through those channels the providers also see Anthropic’s prices, customers and growth, and the filing names Google as one of its two primary competitors. Table 2 sets out the roles each plays.


Oxpeckers on a Cape buffalo in South Africa’s Kruger National Park. They were long celebrated as the tick-eating friends of the large animals they ride. In 2000 the zoologist Paul Weeks reported from Zimbabwe that red-billed oxpeckers did not reduce the ticks on the cattle they fed on, and that wounds healed faster when the birds were kept away, because they keep wounds open and drink from them. The cattle still tolerate them. A partner that helps and feeds at once is how the S-1 describes Anthropic’s.
Photograph by Derek Keats, 2021, licensed under CC BY 2.0, via Wikimedia Commons.
And they finance the shortfall if there is one. Beyond the $131.8 billion of cash and short-term investments Anthropic held at September 30, Google offers $20 billion of note facilities, Amazon $15 billion and Broadcom $42 billion, and AMD has agreed to buy up to $5 billion of stock. Some of this is circular by design. Amazon’s facility opens in $5 billion steps as Amazon delivers capacity, with each draw settled in Anthropic shares, and Broadcom’s notes can be used only to pay rent to Broadcom. The suppliers are lending Anthropic part of the money to buy their own capacity and taking equity in return. The closest precedent is Lucent, which committed about $8 billion of financing to customers buying its equipment in 1999 and 2000 and set aside some $3.5 billion for bad debts in 2001 and 2002 when many of them failed.
Does the Product Clear the Bar?
Put the three terms back together and the test from the start of this piece becomes concrete. On the 2028 bill of $105–120 billion, keeping revenue at $1.82 or more per dollar of compute requires revenue of about $190–220 billion. The spider’s 85 percent gross margin gives nearly the same answer as long as the load factor stays near 27 percent, about where it is now, because serving costs must then stay within 15 percent of revenue. Anthropic has reportedly told investors to expect $190–200 billion in 2028, about $1.60–1.90 per dollar of compute. That clears the bar only at the low end of the bill or with SpaceXAI canceled. So, on the company’s own forecast, whether capacity matches revenue turns on a single contract. That right to cancel runs both ways. SpaceXAI, which builds a rival model, could take its capacity back on three months’ notice.
Table 3 lists the five assumptions the loop depends on, from the register of assumptions Shiva built for boards, with his judgment of the chance that each holds. Each corresponds to a term in the identity. The first is yield times volume and the second is the absorption of capacity. The third and fourth protect the load factor from a halt, the third from a pause Anthropic chooses and the fourth from a shutdown the government imposes. The fifth protects yield from rivals.

Of course, these probabilities are completely subjective. Each assumption is more likely than not, but all five together are not. If they were independent, all five would hold about one time in nine, and allowing for the links between them, perhaps one in five. The first two also pull against each other. Growth of 45 percent a year from the July run-rate puts 2028 revenue at about $137 billion, or $1.14–1.30 per dollar of compute. That is well short of the $1.82 bar, and it would leave compute at roughly 75 to 90 percent of revenue. The base case for usage does not eat the food already bought.
And clearing the bar buys less than it seems. As our first piece calculated, a $2 trillion price implies revenue within five years of about $305 billion for an elephant, $510 billion for a spider and $1.1 trillion for a shark; at $1.5 trillion, about $225 billion, $375 billion and $800 billion. Shiva’s base case, about $300 billion by 2030, meets only the elephant’s requirement. Even if everything goes right, the price works only if the market values Anthropic as an elephant, the species whose fate turns on the number this piece began with.
The Cost of Standing Still
One accounting choice makes the bar look higher than it might be. Anthropic expenses all of its training as it goes and capitalizes none of it. Suppose instead it counted the training that only keeps it level with rivals as a running cost and spread the cost of training that opens new kinds of work over a model’s commercial life of about a year. While spending is growing fast, margins would look better, and the $1.82 bar would fall to about $1.60–1.70. But most of the gain is timing; as growth slows, the bar returns to about $1.76–1.79. Changing the accounting changes when training is counted, not how much of it there is.
The larger question is how fast the cost of keeping up grows. Margins stay constant only if the cost of staying at the frontier grows no faster than revenue. Epoch AI estimates that the cost of frontier training runs has grown about 2.4 times a year since 2016. If the cost of keeping up grew at that pace, revenue would have to grow about 140 percent a year just to hold margins where they are. So far revenue is winning. Anthropic’s R&D grew 5.1 times in the past year while its revenue grew 14 times, and R&D fell from about 160 percent of revenue to 60 percent. The valuation requires that race to keep going Anthropic’s way for five more years.
What It All Adds Up To
Strip away the detail and this piece comes down to one ratio: revenue per dollar of compute, which Anthropic now uses too. It was $2.08 in the second quarter of 2026 and $1.72 in the third. The margin behind the elephant’s valuation needs at least $1.82, and the compute bill will roughly triple by 2028. Everything above is a reason that ratio might hold or might not.
First, capacity is fixed. Anthropic has bought its food years ahead on take-or-pay terms, and it must buy it again as the contracts run out. Of the three terms that make up revenue, capacity is the one it can least change. Only about a quarter of the commitments, led by SpaceXAI, can be canceled or depend on conditions before they begin.
Second, the other two terms are under pressure. Most of the compute does not serve paying customers, because it is building the next model. Each new model cuts what a task earns by 25 to 40 percent, and price cuts took about $600 million off third-quarter revenue. So far usage has made up the difference, but across the industry spending is now growing at less than half the pace of volume.
Third, the numbers barely reach the bar. In the third quarter the ratio fell below it, though for a reason that may pass. On the company’s own forecast, revenue per dollar of compute in 2028 is about $1.60–1.90. That clears $1.82 only if the bill comes in low or SpaceXAI is canceled. On Shiva’s base case it is $1.14–1.30. Changing the accounting does not rescue it, because once growth slows, counting training differently moves the bar by pennies.
Fourth, the cost of standing still decides it. If the cost of keeping up at the frontier grows as fast as frontier training costs have historically, about 2.4 times a year, no plausible path for revenue holds margins. If it slows to the pace of revenue, the elephant’s price is within reach. The S-1 does not say which, because it does not separate the training that keeps Anthropic in place from the training that moves it forward.
Hence the takeaway. At $1.5 trillion to $2 trillion, Anthropic is priced as an elephant, a business whose value depends on filling capacity it has already bought. It behaves like a shark, spending most of that capacity just to keep its place. The price holds only if the shark comes to need less food for each dollar it earns. Investors should watch two numbers every quarter: revenue per dollar of compute, and how much compute Anthropic buys on short-term contracts. In the third quarter the ratio fell below $1.82 because demand outran the capacity Anthropic had contracted and it paid a premium to keep up. That is the better kind of shortfall. If short-term buying falls as the contracted capacity arrives and the ratio climbs back above $1.82, the stored food was a bargain. If short-term buying falls and the ratio keeps falling, the elephant is not eating what it stored, and the larder becomes a debt.
That leaves one decision Anthropic still controls: whether to keep running. In February 2026 it dropped its binding commitment to pause development that outran its safeguards. It kept only narrower promises to delay if it ever finds itself far ahead of its rivals, along with the discretion to pause, and the S-1 reports a recent delay of more than a month to pre-training a new model. A pause would not switch revenue off, but in a Red Queen’s race standing still means falling behind, and the contracts would keep billing.
Our next piece, “Can a Shark Stop Swimming?” asks what remains of Anthropic’s power to stop, what using it would cost, and why a governance structure that looks like a weakness from outside may be what makes stopping possible. The fourth and last, “The Paper Tiger,” turns to the accounting. Most of the $80.7 billion loss in the headlines comes from how Anthropic’s accounts value what it owes Amazon, a measure that makes the reported loss larger the better the company does.
A Note on Method
We wrote this as a hybrid team of two authors and AI models. Each of us worked with Claude, Anthropic’s own model, a fact readers should weigh: both of us used Claude Opus 5.5. Before publication, Google’s Gemini and OpenAI’s ChatGPT each reviewed the draft as an adversary, independently and with the same instructions; the appendix reports what they found and what we changed. Every figure taken from the S-1 and its November 2 amendment was checked against the filings. The annual compute bill, the revenue needed to absorb it, the rate of marketplace fees, and the revenue each valuation implies are our estimates; the probabilities in Table 3 are Shiva’s judgments. Several figures come from outside the filing and are labeled as such: Ramp’s token data, from a sample of its own customers that Ramp says is not representative; the valuation range of $1.5–2 trillion, the July run-rate and the 2028 forecast, which are press reports of what investors were told; and the Lucent figures, from a compilation by Tomasz Tunguz. The revenue each valuation implies is derived from the required growth rates in our first piece, which assume constant margins and a 10 percent annual return.
Appendix: How This Piece Was Made
This is the second of four weekly pieces on Anthropic’s S-1. The first, “What Kind of Animal Is Anthropic?,” classified the specimen. This one describes how it feeds. The third, “Can a Shark Stop Swimming?,” asks whether it can stop, and the fourth, “The Paper Tiger,” explains why the loss in the headlines is not the business.
Who did what. Shiva supplied the analysis on which the piece rests: the commitments set against the backlog, the five assumptions and their probabilities, and the arithmetic of the annual compute bill. After reading an earlier draft, he rebuilt it around a single ratio and the airline’s way of splitting revenue. Bob supplied the general reader’s test, in dozens of questions, along with the animal titles and the images. Claude read and searched the filing, ran and reran the calculations in code, and drafted and revised the text. Gemini and ChatGPT reviewed the draft as adversaries. The company’s amendment of November 2, which reports the third quarter, arrived as the piece was being finished, and Claude checked the new figures against it. Every figure from the filings was checked, and the two of us take responsibility for every claim.
What we learned: reviewers with a stake. Every model in this exercise has a maker with an interest in its subject. Claude is Anthropic’s. Google, which makes Gemini, is Anthropic’s shareholder, lender, supplier, distributor, and competitor. OpenAI, which makes ChatGPT, is its main rival. Both reviewers found real errors, and ChatGPT found the most important one. Anthropic had dropped its binding commitment to pause in February 2026, a change our drafts had missed although Shiva’s own workings recorded it. But neither reviewer discussed her maker’s interests until we asked. When we did, each found something. ChatGPT concluded that she had given the case for heavy compute commitments, which her own maker makes, the benefit of the doubt while demanding hard proof of the case against them. The lesson is practical. When the analysts have makers with a stake in the subject, use more than one, ask each about her maker, and check what each says against the documents. (The app identified the reviewers as Gemini 3.1 Pro, which described herself as 1.5 Pro, and ChatGPT, which described herself as GPT-6, a version we could not confirm.)
Sources
Anthropic, PBC, Form S-1, October 26, 2026, and Form S-1/A, November 2, 2026. Ramp, Ramp AI Index, data through the week ending October 11, 2026. Reuters, “Anthropic IPO valuation hinges on $190–200 billion 2028 revenue forecast,” August 15, 2026. The Hill, “Anthropic narrows AI safety policy pledge,” February 25, 2026. Epoch AI, “How much does it cost to train frontier AI models?,” June 2024. Tomasz Tunguz, “Circular Financing: Does Nvidia’s $110B Bet Echo the Telecom Bubble?,” October 3, 2025. Microsoft, Form 10-K for fiscal 2026. Paul Weeks, “Red-billed oxpeckers: vampires or tickbirds?,” Behavioral Ecology 11 (2000): 154–160. Leigh Van Valen, “A New Evolutionary Law,” Evolutionary Theory 1 (1973): 1–30.
This is the public content supplied by Substack. Any subscriber-only content remains on Substack.






















































































































