Anthropic Needs to Disclose Its Tokenomics Before Its IPO
Ilan Strauss, Tim O'Reilly / Nov 3, 2026Ilan Strauss and Tim O’Reilly are co-directors of the AI Disclosures Project.
In this photo illustration, the icon for the AI app "Claude" by Anthropic is seen through running water as it is displayed on a digital screen on September 16, 2026 in London, England. (Photo by Leon Neal/Getty Images)
When Anthropic filed its S-1 statement on October 26, it gave us our first largely audited look into the financials of one of the world’s leading AI companies. An amended filing on Monday added third-quarter results. This S-1 is only a preliminary submission, however, and the Securities and Exchange Commission (SEC) can request additional information.
Anthropic’s filing will establish a precedent for other AI labs preparing to go public. What it leaves out may therefore matter just as much as what it discloses. Two disclosure gaps stand out. Together, they make the AI market more opaque, increase financial risk and fragility, and undermine the efficient allocation of capital. The SEC should exercise its authority in the public interest to request additional information to address these gaps.
The first concerns Anthropic’s relationships with the firms that finance it, distribute its models, and sell it computing power. More than 40% of its revenue comes through the cloud marketplaces of Amazon, Google, and Microsoft, which offer Anthropic’s Claude models to their own cloud customers. Yet Anthropic does not disclose how much it pays each of them or how much it earns through each. It also has not yet filed with its S-1 the contracts to buy computing power from these companies, despite SEC rules requiring disclosure of contracts on which a business is substantially dependent.
The second disclosure gap concerns the “token”: a small piece of text or other data processed by a model, and the unit in which most of Anthropic’s usage-based sales are billed. The filing does not report token volumes (how many tokens Anthropic’s models processed for customers) or revenue and costs per token. It also does not say how much of the computing capacity Anthropic has contracted for has been delivered or is in use. The SEC already requires oil and gas producers to report how much oil and gas they produce, the average price they receive for each, and the average cost of producing a barrel of oil or its equivalent in gas, all using common definitions. Frontier AI developers need the same, including a common definition of a token, so investors can compare revenue and costs per token across companies.
Existing SEC rules can largely close the first gap. The second calls for standardized disclosures specific to the AI industry. Anthropic’s revenue is growing extraordinarily quickly, yet investors still cannot assess whether the computing capacity behind that growth will earn a return.
Disclosure matters all the more because ordinary shareholders will have limited influence over the company’s governance. After the IPO, Anthropic’s seven founders will hold 50.1% of votes on general shareholder matters, while the Long-Term Benefit Trust (an independent body of five trustees with no financial stake in the company) will select four of the seven directors. This unusual division of control strengthens the case for transparent, comparable disclosure.
Anthropic’s changing cloud economics
The S-1 filing gives outsiders their first audited annual accounts for Anthropic, alongside unaudited results for the first nine months of 2026. The filing reports revenue of $4.6 billion in 2025 and $33.2 billion in the first nine months of 2026, including $17.0 billion in the third quarter alone.
But is Anthropic profitable? Its headline net loss of $104.0 billion for the first nine months of 2026 is largely a noncash accounting consequence of its rising valuation. Most of this loss ($94.3 billion) comes from convertible notes (loans from investors that will turn into shares) and similar contracts. As the company’s valuation rose, the shares it owes these investors became worth more. Accounting rules count that rise as a loss, though no cash is paid out.
The underlying business tells a different story. Anthropic lost $1.7 billion from operations in the first quarter of 2026, but made its first quarterly operating profit in the second: $974 million, or about 8 cents on each dollar of revenue. Over the nine months, $2.4 billion in interest on its cash and investments more than covered its $1.3 billion operating loss, not counting those much larger accounting charges.
The airline problem: paying for tomorrow’s capacity
Anthropic typically contracts for computing capacity 12 to 18 months ahead, so its near-term costs are largely fixed. This resembles airlines leasing aircraft: payments begin when the planes arrive, whether or not the seats are full. Profits then depend on filling that capacity at sufficiently high prices. In the second quarter, Anthropic’s compute costs rose 64% from the first quarter, while revenue jumped 144%. The amended filing shows how quickly results can swing. In the third quarter, compute costs rose 77% while revenue rose 47%. The $974 million operating profit became a $567 million loss, mainly because of $5 billion of short-term compute expenses, priced well above its long-term contracts, and a $2 billion noncash charge for matching employees’ charitable donations with company shares.
Anthropic’s business model is changing though, and with it the meaning of its reported profits. Until recently, it predominantly purchased computing services from cloud providers such as Amazon and Google, recording expenses as capacity became available. It is now increasingly leasing computing equipment directly, including $161 billion in five-year equipment leases with Broadcom. These are treated more like investments in equipment, much as an oil company treats a new well: their costs are spread out over several years, with the interest on the leases recorded separately from operating profit.
Anthropic puts its leases and purchase commitments at a staggering $895 billion. Yet its balance sheet shows just $4.5 billion in lease liabilities, because leases are recorded only once the equipment is ready for use, and purchase commitments not at all.
This will make Anthropic’s future operating profits challenging to interpret. Rising operating margins will not necessarily mean less cash is being committed to computing infrastructure.
Where Anthropic’s expenses go
In the first nine months of 2026, Anthropic recorded 57 cents in compute and infrastructure costs per dollar of revenue. Personnel accounted for another 15 cents. Sales, marketing, and partnerships cost another 17 cents, much of it in fees to cloud marketplaces distributing Claude.The trouble in tracking these flows is that Amazon and Google are simultaneously distributors of Claude through their own marketplaces, investors in Anthropic, and suppliers of its compute.
As distributors: In the first nine months of 2026, 44% of Anthropic’s revenue came through Amazon, Google, and Microsoft cloud marketplaces. Anthropic books these sales at the full customer price and pays the cloud providers distribution fees, although it does not disclose what share of the price they keep.
As investors: In 2026, Google bought nearly $20 billion of Anthropic’s preferred stock and Amazon bought $10 billion, and both have arrangements for further billions in future financing.
As compute suppliers: Anthropic has committed to spending at least $121.3 billion with Google, $110 billion with Amazon, and $44.9 billion with Microsoft on compute over multiyear contracts extending into the 2030s. These are mostly future spending commitments. The Google and Amazon agreements require minimum payments through 2033 and 2036, respectively. Subject to exceptions, Anthropic must pay any shortfall if spending falls below those minimums. By contrast, its own revenue, the filing notes, is “primarily consumption-based, and much of it is not subject to long-term contractual commitments.”
Applying existing SEC rules
What the filing leaves out about Anthropic’s cloud contracts and future payment obligations can largely be addressed through existing SEC disclosure rules. We made this case a year ago in Tech Policy Press.
Shortly after we published our recommendations, in December 2025, the SEC’s Investor Advisory Committee, a body Congress created to advise the agency, published its own recommended AI disclosure guidance. It built on existing disclosure rules and argued that a separate set of AI disclosure rules may not be necessary. Even so, it describes its recommendations as only the “initial scaffolding” of a framework that “may be further built upon as technology develops.”
In light of Anthropic’s S-1, two pressing disclosure obligations stand out:
First, Anthropic should file its main compute contracts. Under SEC rules (Item 601(b)(10) of Regulation S-K), a company must file with the SEC any contract its business is “substantially dependent” on, including contracts to buy “the major part” of what it needs. Yet Anthropic has not filed its compute deals with Google, Amazon, Microsoft, or SpaceX’s AI unit (formerly xAI). Each of the first three commits it to at least $44.9 billion, more than its revenue through September 2026. The SpaceX deals commit it to spend up to $120.5 billion.
Beyond filing the contracts, Anthropic should ideally disclose the money it pays to and receives from Amazon, Google, and Microsoft, partner by partner: what each invests in Anthropic, what Anthropic pays them, and what Anthropic earns through their marketplaces. SEC rules already require a company to discuss commitments that could materially affect its cash position or revenue and to add any material information needed to keep the filing from being misleading. And the contract-filing rule lets a company withhold contract terms only when those terms are not material, meaning a reasonable investor would not consider them important. Given how central these relationships are to Anthropic’s revenue and computing supply, their materiality is difficult to dismiss.
Second, and arguably more consequential, Anthropic should show how much of its future compute spending it is legally bound to pay. The amended filing puts its compute commitments at roughly $895 billion. About $214 billion of that can be canceled or depends on conditions, such as partners securing financing. A further $150 billion to $175 billion in estimated payments, for deals with no fixed dollar amount, is left out of that total. The filing provides a rough timeline for when these commitments fall due, but investors need to know how much Anthropic must actually pay, and what happens if revenue growth disappoints. The SEC could easily press for greater transparency before the IPO. Item 303(b)(1) of Regulation S-K is a natural lever: it requires companies to explain how much cash they will need to meet their contracts and other known obligations.
Beyond these immediate demands, the SEC should apply the successful cybersecurity disclosure model to material AI risks and safety incidents. The SEC began issuing cybersecurity guidance in 2011 and introduced dedicated requirements in 2023; AI disclosure could follow the same approach, we argued in previous research. This would mean adding a dedicated AI section to the annual 10-K report and reporting material AI incidents as they happen through 8-K filings – the reports companies must file after major events.
Missing breakdowns: compute use, product revenue, and data costs
Anthropic’s S-1 provides some helpful operating measures, including customer retention and the number of large customers. But three other gaps make the economics of its business difficult to evaluate:
- Anthropic does not say how much of its computing goes to training new models and how much to serving customers. It reports a single compute bill and shifts computing between these uses as needed. So investors cannot tell whether the fall in compute costs relative to revenue since early 2025 comes from serving customers more cheaply or from shifting computing away from training temporarily.
- Anthropic reports no revenue by product line, so investors cannot tell whether Claude Code (its coding assistant), the API (which lets developers build Claude into their own products), or the consumer plans drive growth and margins. With Prof. Mariana Mazzucato of University College London, we have called for enforcement of proper segment reporting disclosures, so that product-by-product financials are broken out rather than buried in the aggregate.
- The filing also says little about the cost of training data and outside labor. We are only told that the costs of data quality and model performance evaluation services rose by $757 million in the first nine months of 2026, compared with a year earlier. Anthropic does not report the total cost of these services or how much is paid to outside workers preparing data and evaluating models.
Industry-specific disclosures: lessons from oil and gas
Existing SEC rules can largely close the gaps above. But some figures, such as token volumes, the cost of serving each token, and how fully computing capacity is used, need standard definitions written for the AI industry. The SEC already sets tailored requirements for oil and gas producers, mining companies, and banks, because each makes money in a different way.
The oil and gas rules are probably the closest parallel. Like an AI developer, an oil producer commits to heavy spending long before output arrives, and its returns turn on volume, price, and cost per unit. Item 1204 of Regulation S-K requires oil producers to report how much they produced, the average price they received, and the average cost of producing each unit, using common definitions set by the SEC. Investors can then see whether each producer’s average margin per barrel is widening or shrinking, and compare producers.
For AI developers, the equivalent unit is the token. Anthropic records revenue according to the number of tokens customers use, and this usage-based revenue made up 79% of its total in the first nine months of 2026. Yet the filing does not report how many tokens Anthropic processed for customers or what it earned or spent per token. Anthropic says it cut customers’ cost per task by about 40% with its Opus 5.5 model “while preserving our margins.” Price cuts and discounts took about $600 million off third-quarter revenue. Disclosing revenue and cost per token would show whether those margins are holding. As we argued with Mazzucato in the Oxford Review of Economic Policy and a related policy brief, regulators should require companies to disclose the key operating figures they use internally to run their business lines.
The obvious objection is that tokens are not all alike. Anthropic’s filing itself notes that the same text can be split into a different number of tokens depending on the model, and that prices vary with the model chosen. But oil and gas faced the same problem: oil, gas, and natural gas liquids are different products with different prices. So the SEC required volumes and prices to be reported for each product separately. Yet oil and gas often come from the same wells, so their costs cannot be cleanly split. Producers therefore convert all output into a common unit, usually barrels of oil equivalent (converting gas by its energy content), and report one cost per unit, comparable across companies.
The goal over time should be common measures of AI developers’ output and of the computing power used to produce it, with tokens counted the same way at every firm. That would make efficiency and costs comparable across AI models and firms. Without a common count, a firm whose models split text more finely would appear to sell more tokens, at a lower price and cost per token, for the same work. In 2022, SEC staff told a mining company that required production figures could not be withheld as confidential.
AI developers could begin by reporting input and output token volumes (the tokens customers send in and those the model sends back) for each model class, such as Opus or the low-cost Haiku, along with the average revenue and cost per token. Because Anthropic pools its computing across research, training, and serving customers, it should also explain how it splits those shared costs to arrive at a cost per token.
The token operating figures would show how many tokens customers buy and at what price. Investors also need to know how much computing capacity Anthropic has to produce those tokens. Airlines voluntarily report revenue per seat alongside how many seats they filled: Delta’s “load factor” was 84% in 2025, for example. AI developers could, likewise, disclose contracted computing capacity, capacity actually delivered, and how intensively it is used, explaining the definitions behind those figures. Together, these figures would show investors whether compute efficiency is real, and whether today’s commitments can earn a return.
Conclusion: What the market needs to know
The SEC already requires industry-specific disclosures when conventional financial statements fail to capture important features of a business. Oil and gas companies must report standardized information about their reserves. Even the SEC’s Investor Advisory Committee, which favored working through existing rules, hoped the SEC would offer examples so that “issuers in similar industries” could disclose “in a standardized format.”
Anthropic has shown how quickly its revenues can grow. But it has not given investors nearly enough information to assess its returns on the computing capacity behind that growth, or how the revenues and costs are divided between Anthropic and its cloud and distribution partners.
Frontier AI model developers should be required to report their physical computing capacity, how much of it they use, and what they earn and spend per token, alongside meaningful measures of AI risks and public benefit. Anthropic’s IPO is an opportunity to establish these expectations before its disclosures become the template for the industry.
Authors
