Can a Frontier AI Lab Make Money?

On July 31, 2026, Amazon filed its report for the second quarter. Below the cloud business sat a line called other income: a net gain of $53.4 billion for the quarter and $69.1 billion for the first half of the year. The gain, it said, was "primarily from upward adjustments for observable changes in price relating to our nonvoting preferred stock in Anthropic" and from gains on Anthropic notes it had converted.

Put that next to Anthropic's own numbers and the picture turns over. In the same six months, Anthropic's revenue was $16.3 billion. One of its suppliers booked a net gain more than four times that, on paper, which it said came primarily from its stake.

On October 26, 2026, Anthropic filed its S-1 to go public. The headline number in it was a net loss of $80.7 billion for the first half. Same company, same boom, two ledgers. When Anthropic became more valuable, the notes it owed to its investors became more valuable too. Anthropic booked the difference as a loss. Its investors, who hold those notes and its preferred stock, booked gains.

That is the question this Deep Dive puts to the record. Ask whether a frontier AI lab can make money, and the popular answer is no: these labs are money pits, kept alive by the cloud companies that sell them computing power. We tested that against Anthropic's S-1, Amazon's filings and the Census Bureau's survey of American businesses. The first half of the claim is going out of date. The second half is only half right: the money runs both ways. The lab is starting to make money. So far, at least one of its suppliers has booked more from it, on paper.

The case for the money pit is not a straw man. It was laid out, in part, by Anthropic's own chief executive. In a podcast published in August 2025, Dario Amodei walked through how the economics look in a conventional reading. A lab trains a model for $100 million, then the next one for $1 billion, then the next for $10 billion. Looked at "in a conventional way at the profit and loss of the company," he said, "you've lost $100 million the first year, you've lost $800 million the second year, and you've lost $8 billion in the third year, so it looks like it's getting worse and worse." His own answer was that each model, taken alone, can be profitable; the losses come from paying for the next, bigger one.

Add the scale of the spending, and the story writes itself. The cloud companies are building data centers at a pace the economy has not seen since the dot-com years. Deep Dive 1 traced who is paying for that and found the money coming mostly from the cloud companies' own cash, with the risk sitting in promises from a handful of AI developers. Anthropic is one of those developers. If its revenue cannot carry its bills, the received view says, the suppliers will have to keep carrying it.

It is a fair worry. It just rests on two assumptions that the S-1 lets us test: that the lab loses money on its operations, and that the money flows one way, from suppliers to the lab.

Start with the business itself, before any of the financing. Revenue was $376 million in the first quarter of 2025. Five quarters later, in the second quarter of 2026, it was $11.6 billion. The filing gives the reason in plain terms: "successive model and product releases that drove higher usage, expansion, and broader adoption across our offerings."

The costs grew too, but not as fast. Anthropic's operating margin, its operating income or loss as a share of revenue, was minus 249% in the first quarter of 2025. In the second quarter of 2026 it was plus 8%: operating income of $974 million on revenue of $11,555 million. In the S-1's words, the company "generated operating income of $1.0 billion for the three months ended June 30, 2026, as revenue growth outpaced the growth in total operating expenses."

That turn is the first thing the received view gets wrong, or at least gets late. The losses in Amodei's example grow because each new model costs ten times the last. Anthropic's real numbers show what happens when revenue grows faster than the bill. The biggest cost, computing power, is bought "12 to 18 months in advance of availability," so it is "generally fixed in the near term." When revenue outruns a bill that was set more than a year earlier, the margin swings fast, and in the second quarter it swung past zero.

So what about the $80.7 billion loss? Most of it never left the building. $72.4 billion of it is the change in the fair value of convertible notes, derivatives and SAFEs, which the filing says rose "primarily" because of "an increase in our enterprise valuation during each of the respective periods." Another $8.8 billion of other expense came mostly from recording the note purchase agreements at fair value. Amazon's gains trace to the same cause, a higher valuation for Anthropic, recorded on different instruments in different ways.

The fair caveat is that one quarter is not a business, and the S-1 itself warns as much. "We expect our revenue growth rate to decline over time." And as new computing power comes online, "the associated expense will be recognized in advance of the full revenue that the compute will ultimately support, and we expect operating margin and Adjusted Operating Margin may compress in those periods." The verdict on the money pit: out of date, for one quarter, with the company's own warning attached.

The second half of the received view says the money flows from the suppliers to the lab. Follow it through the filing and you find it flowing in at least four directions at once.

First, the shop window. Nearly half of what Anthropic sells goes through the cloud companies' own marketplaces. That share was about 11% in 2023, 32% in 2024, 47% in 2025 and 48% in the first half of 2026. The marketplaces "bill end customers, collect payments on our behalf, and remit such amounts to us, net of applicable platform and other fees." At June 30, 2026, 58% of the money customers owed Anthropic was owed through them. The fees are not small. Sales, marketing and partnership costs rose by $3.0 billion in the first half of 2026, mostly from "a $2.7 billion increase in fees paid to channel partners."

Second, the factory. Anthropic buys most of its computing power from the same companies. The S-1 lists the commitments supplier by supplier. To Amazon, "at least $110.0 billion between May 2026 and April 2036, regardless of actual usage." To Google, "at least approximately $114.5 billion between April 2026 and July 2033, regardless of actual usage." To Microsoft, "at least $44.9 billion." Then come the newer names: SpaceXAI, Nscale, Lambda, AMD, and Broadcom, whose equipment leases carry about $161.2 billion of rent. The terms differ, so a total is only a sense of scale, but the first seven commitments in the table come to about $637 billion. The filing leaves the overall figure blank.

Third, the bank. Here the circle closes. Several of the suppliers have also agreed to provide financing to a customer that buys from them. Anthropic can issue up to $42.0 billion of notes to Broadcom, "to finance obligations under the 2026 Lease Arrangement": the leases of Broadcom's own equipment. Google can take up to $20.0 billion more of notes or stock, Amazon up to $15.0 billion, and AMD has agreed to buy up to $5.0 billion of stock as its chips are delivered and deployed. Across the three note agreements, up to $77.0 billion of notes or stock could be available, subject to conditions.

Fourth, the owners. Amazon, Google and Microsoft "have made significant equity investments in our company," the S-1 says. "These CSPs are also our customers." Amazon's side of that trade is in its own filings: $8.0 billion of convertible notes from 2023 to 2025, $10.0 billion of preferred stock in the second quarter of 2026, a financing facility of up to $20.0 billion, and an expansion of Anthropic's AWS commitment "by more than $100.0 billion over 10.0 years." The S-1 adds that notes issued under Amazon's agreements can be settled "by set-off against amounts Amazon owes the Company under certain commercial agreements with Amazon."

The S-1 has one sentence that says all of this at once. The cloud companies, it says, "simultaneously serve as our investors, customers, distribution partners, and competitors," while Broadcom "simultaneously serves as a supplier and lessor." Strategists have a word for a partner who is also a rival: a frenemy. This is something more tangled. It is a market where the same three or four names sit on every side of the table.

Does that make the received view right after all? Half right. The suppliers are keeping the lab supplied, and some will finance it. But look at who has booked the gains so far. Amazon's filings show $18.0 billion invested: $8.0 billion of notes and $10.0 billion of preferred stock. Its net gain for the first half of 2026 alone was $69.1 billion, "primarily" from Anthropic, and came with $15.9 billion of net discrete tax expense "primarily attributable to the upward adjustments to our investments in Anthropic." The lab is not just being carried. It has made at least one supplier richer, on paper, than its own revenue.

One more finding, from what is missing. We read the latest annual and quarterly reports of Alphabet, Microsoft and NVIDIA for the word "Anthropic". It does not appear in any of them. Google's commitments and investments are described in Anthropic's S-1. Alphabet's own 10-Q reports "$20.0 billion of future capital funding commitments with a private company" through 2030, the same amount and end date as the remaining Google agreement in the S-1, and second-quarter gains on equity securities of $99.0 billion from "SpaceX and a private company." It does not name the company. Neither do Microsoft or NVIDIA.

That is not a scandal. Companies disclose what is material to their own investors, and a stake that is large for Anthropic can be small for Alphabet. But it is a lesson in where to look. If you want to understand a web like this one, read every party's filings. The richest description of the relationships came from the company with the least choice about disclosure: the one going public.

A lab that makes money for one quarter is interesting. A lab whose customers stay is a business. Here the record is more mixed, and the S-1 is unusually frank about it.

The customer count is the strong part. Customers spending $100,000 or more a year went from about 20 at the end of 2023 to more than 11,900 by September 30, 2026. Net revenue retention, which compares what a group of customers spends now with what they spent a year earlier, was about 900%, leaving out the fixed-price Pro and Max plans. And the business is spreading out. The five largest customers made up about 36% of revenue in the first quarter of 2026, by our arithmetic, and about 21% in the third.

Now the frank part. Model versions released in the prior six months "accounted for over 80% of our revenue" in the third quarter: the product turns over fast. Spending fell by about 20% over the year-end holidays, when many companies freeze their code. "Many of our top customers are not subject to long-term contracts with us" and can "reduce or in some cases cease spending at any time and for any reason." And the filing estimates that "approximately 5% to 7%" of revenue in the first nine months of 2026 came from customers later found to be breaking its usage policy or commercial terms, a share it expects to fall after stepped-up enforcement in August and September.

Then there is the runway. Over half of the Fortune 500 now spend at least $100,000 a year with Anthropic, yet that spending is "on average, less than 0.1% of their total expenses." The Census Bureau's survey points the same way across the economy. In the latest wave, 24.6% of US businesses said they had used AI in the previous two weeks, with information companies far ahead of restaurants and hotels.

Read together, that is a business with deep, fast-growing relationships that are not yet locked in. The revenue is real. Its durability is still being earned, one model release at a time.

The last test is not about money at all, which is exactly why strategists should read it. The S-1 pulls apart two things most companies keep together: who gets the vote, and who gets the money.

Public buyers will get Class A shares, one vote each. The strategic partners hold Class N shares, with economic rights but no vote. The seven co-founders, through a single Class F share, will hold "50.1% of the total voting power" on matters put to shareholders, with no economic rights. A Long-Term Benefit Trust of five financially disinterested trustees holds Class T stock and elects the directors "who together will constitute a majority" of the board, also with no economic rights. Employees get Class E shares that vote on limited matters.

On top of that, Anthropic has put 5% of its fully diluted shares into a Public Benefit Fund it owns, and each co-founder has pledged 80% of their personal equity to philanthropy. The S-1 is careful to call those pledges "voluntary, revocable expressions of each co-founder's philanthropic intent" that "do not constitute legally binding obligations."

Put that next to Test two and the design becomes clear. The suppliers and investors get the economics. The founders and the trust get control. Whatever you think of the mission, in effect it works as an answer to the frenemy problem: the companies on every side of the table own a lot of the upside, but they do not get to steer.

The comparison worth making is with other founder-controlled listings. In the familiar dual-class design, such as Alphabet's or Meta's Class B shares, the founders' high-vote shares also carry economic rights. Here the founders' share carries the vote and no money at all, and a trust with no money elects the board's majority. In effect, it bets that control without economics produces better long-run judgment than control with them. The filing cannot prove that. Only the next decade of decisions can.

Suppose, as a thought experiment and not a forecast, that Anthropic's revenue had stopped growing after the first quarter of 2026 and stayed at $4,731 million. Most of its computing power would have arrived on schedule anyway, because it was ordered 12 to 18 months earlier. Second-quarter compute and infrastructure cost $5,558 million. Against first-quarter revenue, that bill would have been 117% of revenue.

That is the arithmetic behind the company's own warning about margin compression, and behind the clause that recurs in its commitments: "regardless of actual usage." It is also why the gains on the other side of the trade matter. In a stall, the commitments stay put. The paper gains do not.

A habit worth keeping: read a deal from every chair at the table. The S-1 rewards it, because each party can point to real receipts, and each has something that cuts against it.

From the lab's chair, the model business works. Revenue was $16.3 billion in the first half of 2026, against $1.2 billion a year earlier. Adjusted operating margin, which leaves out stock-based pay, non-cash gifts to charity and a legal settlement, was 10% for the half, up from minus 159% a year before. Customers paying $100,000 or more a year passed 11,900. What cuts against it: growth has come in jumps tied to model releases, and the company itself expects its growth rate to decline.

From the skeptic's chair, computing power eats the business. Compute and infrastructure cost $8,945 million in the first half of 2026, 55% of revenue. Most of it is bought a year or more ahead, and much of the long-term bill is pledged "regardless of actual usage." The total of the commitments is blank in this draft of the filing; the S-1 says only that substantially all of what falls after 2030 runs from 2031 to 2045. What cuts against it: the ratio is falling fast. Compute cost 251% of revenue in the first quarter of 2025 and 48% in the second quarter of 2026.

From the supplier's chair, this is the best customer in the business. Anthropic trains and serves its models on four chip families: Amazon's Trainium, Google's TPUs, NVIDIA's GPUs and AMD's GPUs. On our reading, that spreads its spending across rivals, so no single supplier can take the account for granted. Each of the three cloud companies gets three things at once: a large, committed buyer, a product to resell in its own marketplace, and a stake whose paper value has risen. What cuts against it: Amazon, Google and Microsoft also sell competing models, and the S-1 says the arrangement gives them "visibility into our pricing and commercial terms," a conflict Anthropic lists as a risk.

From the customer's chair, the bargain is capable models that keep improving, sold through a cloud account you already have. The S-1 describes a "joint go-to-market approach" in which Anthropic contracts directly with customers and pays each cloud company a fee, which it says lets it "avoid channel conflict, set pricing." What cuts against it: the product turns over within months, so whatever you build on this year's model, you may need to rebuild on next year's.

From the public investor's chair, the pitch is growth inside a governance wrapper. Anthropic will be a controlled company under Nasdaq's rules, because the trust can elect a majority of the board, so it may opt out of some of the usual board requirements. Concentration is easing but not gone: one customer, the S-1's Customer C, was 11% of revenue in the first half of 2026. And the board may give the Public Benefit Fund up to 2% more of the shares each time the share price rises 50% above the prior milestone.

None of these chairs is wrong. That is the point. A deal that looks good from every seat is either very good or not yet tested, and one profitable quarter cannot tell you which.

Here is the received view, line by line, against the record.

If you buy AI at scale, find out where your money goes. Nearly half of Anthropic's sales pass through a cloud marketplace that takes a fee and also sells competing models. Buying through the marketplace may suit your procurement; just know that you are paying two companies, and that one of them is also a rival of the other.

If you sell to a fast-growing customer, decide whether you are also its lender. Broadcom, Google, Amazon and AMD have each agreed to provide financing to a customer that buys from them. That can win the deal. It also means your revenue and your credit risk are the same bet.

If you commit capacity ahead of demand, match the length of the commitment to the life of what it builds. Anthropic's commitments run to 2033 and beyond, "regardless of actual usage," for products that turn over in months. The second quarter of 2026 shows that the bet can pay. The thought experiment shows what happens if it doesn't.

If you advise companies on AI, the S-1 hands you a checklist most vendor pitches leave out: how much of the vendor's revenue comes through a reseller, how far ahead it buys its capacity, how much of its revenue comes from model versions less than six months old, and who controls the board. Each is disclosed here. Ask any vendor you recommend for the same four.

If you sit on a board, notice that control and economics can be separated on purpose. Anthropic's structure gives partners the upside and keeps the steering wheel elsewhere. Whether you copy it or avoid it, decide deliberately.

Then take the question to your own business: how much of your revenue comes from somebody who is also trying to replace you?

Every figure comes from a filing, a federal survey or, for Amodei's illustration, a published transcript, and every one is in the claims ledger behind this page. The core sources are Anthropic's preliminary S-1, filed October 26, 2026, and Amazon's 10-Q for the quarter ended June 30, 2026. We also read Alphabet's 10-Q for the same quarter, Microsoft's annual report for the fiscal year ended June 30, 2026, and NVIDIA's 10-Q for the quarter ended July 26, 2026. The Census Bureau's Business Trends and Outlook Survey supplies the AI-use figures, as in Deep Dive 1.

Definitions travel with the numbers. Operating margin is operating income or loss over revenue, as printed in the S-1. The top-five share is our arithmetic: 100% minus the share the S-1 gives for customers outside the top five. The commitments in the table are stated in different ways ("at least", "up to", rent excluding interest), so their total is a scale, not a precise sum.

The filings are as of October 27, 2026; Amazon's, Alphabet's and Microsoft's reports for the third quarter will update them. There is also a lot we could not see. The S-1 is preliminary: third-quarter revenue, the total of the compute commitments, and the offering price are blank, and an amendment may change other figures. Amazon's gains are Amazon's measurement of its own holdings, and part of the $69.1 billion may come from other investments, so we do not net Amazon's gain against Anthropic's loss. They measure different instruments in different ways. Alphabet, Microsoft and NVIDIA do not name Anthropic in their filings, so their side of the relationships is described here only through Anthropic's S-1.

This is analysis, not investment advice. Stratrix's research tools run on Claude, an Anthropic model; this Deep Dive was edited, checked against its sources and approved by people.

Can a frontier AI lab make money? On the evidence of one quarter, yes. The better question is who else is making it. In a boom, follow the paper gains: the company that books the gain owns the upside, and the company that signs the commitment owns the risk. When the same names sit on both sides of the deal, read it from both sides.