A Critical Reading of Anthropic's pre-IPO Filing

On October 26, 2026, Anthropic filed its public S-1 registration statement in preparation of a planned IPO. While the document primarily reflects Anthropic's own view of its business and framing of relevant economic and societal questions, it still represents a rare window into the workings of one of the major drivers of large-scale AI in the global economy.

Anthropic's flagship products, Claude models and Claude Code, are undoubtedly impressive, consistently topping benchmarks and enjoying a strong reputation among their core base of software developers. However, commanding the kind of exorbitant market-entry stock prices the company is aiming for requires a business story beyond simple product popularity; the goal of the registration statement is to put forward such a story, one of emerging profitability (albeit under a particular kind of accounting) layered with near-unlimited market potential.

But a deeper look at the same numbers mustered to support this account also shows evidence that goes against prevailing narratives on AI innovation and economic growth, adding further support to the warning flags organizations like AI Now (among many others) have been raising with respect to circular financing patterns, distorted information ecosystems, and concentration of economic power.

This article offers a critical reading of the S-1 registration statement with a view toward surfacing evidence related to these concerns. Across three sections, we show how the numbers disclosed in the filing paint a picture of success supported by financial instruments as much or even more than technological success, how this economic power translates into demand creation strategies that pay dividends for the company but obfuscate the limitations of more “organic” adoption of their services, and finally review the cost of this economic power in terms of obligations to hyperscalers; and what they mean for other categories of prospective shareholders.

Dual Currencies, Economic Power, and Profit Accounting: How Billions Pay for Themselves

Anthropic's economic situation needs to be read in the context of its exceptional fundraising history ($141 billion raised as of June 2026 by the company's own account, with nearly $90 billion more in committed facilities), which has given it singular levers of economic power through its stratospheric cash reserves and high-value stock. Neither of these two mechanisms is entirely novel: the rise of venture capitalism as the main under-writer of technology development has seen other companies leverage early cash reserves (Amazon, Uber capturing markets through sustained losses) and stock as a private currency (e.g. Facebook's pre-IPO acquisition of Instagram) to great market power effect. But the scale of the phenomenon for recent AI champions is such that we often struggle to properly appreciate its impact on their economic trajectory. Thankfully, the statement does offer a partial view of how the company puts these affordances to use:

• Deriving income from treasury yield. This might be the purest expression of market power afforded by cash reserves, and illustrates how investment patterns turn into self-fulfilling prophecies. As of June 30, 2026, Anthropic holds $131.6B of investor money in cash equivalents and short-term investments, which earned over $1.1B of interest in the first half of 2026: about two-thirds as much as the entire Claude business by the company's own preferred measure, with nothing to do with business performance and everything to do with invested amounts.
• Absorbing labor and innovation costs. Using stock to pay for labor and M&A is standard practice; but exceptionally valued stock lets the company purchase commensurate resources without booking the equivalent expense on the income statement; so long as the market trusts the company's ability to grow, which it in turn supports. In the case of Anthropic, this labor cost adds up to $2.2B of stock-based compensation in the first half of 2026, and reported acquisitions total $566M over the same period, with the disclosed case of Stainless Software structured as 96% shares. Both stay out of the company's "adjusted operating income" figure: the $2.2B by an add-back, without which the advertised +$1.6B for the same half-year would be a loss; the $566M by accounting rule, since share consideration becomes goodwill rather than expense.
• Funding strategically aligned philanthropy. A third, more novel use of the stock's high value is philanthropic funding. The statement mentions two main mechanisms: employee donation matching programs totaling $667M to date, and the newly endowed Public Benefit Fund, holding 5% of the company's shares and established inside the company's own governance structure. Both extend the company into an ecosystem that shapes its public image and strategic narratives: non-profits drive public discourse, policy debate, and adoption of new technology, as also reflected by the 75% discount Anthropic provides them outside of these mechanisms. In addition to leaving the cash ledgers intact, gifts paid in stock see their value follow the fortunes of the company: recipients become shareholders whose budgets depend on the valuation.

None of these points invalidate Anthropic's revenue growth story in its most basic form: surging revenue from token-based consumption has radically changed the company's balance sheet over the last year. But they show how far the company's influence is a direct product of the funding and valuation it has received rather than of commercial success. Anthropic's lifetime revenue of $21B remains dwarfed by the $141B its investors have contributed, nearly seven to one. The "free" levers outlined above add up to several times the value of a full "unicorn" company, still a standard marker of start-up success: $1.1B for a half-year of interest, $2.2B for a half-year of equity pay, $667M in donation matching, 5% Public Benefit Fund endowment; combining to fund operations, labor, generosity, and narrative support without weighing on the cash ledger. Where earlier generations of venture-funded companies used cash reserves and private stock as bridges to business performance, this unprecedented scale of funding allows them to function as a substitute for it, as long as the valuation holds or increases.

The flagship profitability figure also looks different when measured against these affordances. Companies have wide latitude in deciding what counts as a cost, and Anthropic reports several profit figures, but one in particular is showcased in the Business section to support its profitability story: the "adjusted operating income" line, reported at $1.6B for the first half of 2026 and presented for the second quarter as an "Adjusted Operating Margin" of 19%. But that figure remains smaller than the $2.2B of stock-based compensation paid to employees over the same period: counting it as a labor expense (as the official accounting already does, and as an innovation model that values "talent" as much as compute makes hard to deny) brings the books back to an operating loss of -$0.7B. Other comparisons add perspective: of the same order as the $1.1B of purely financial interest earned on the cash pile over the same period, and much smaller than the $2.8B of channel fees Anthropic pays its distribution partners, an administrative expense in all but name, itself dwarfed by the $8.9B compute bill.

With $131.6B in the bank and nearly $90B more committed, the public offering serves less to raise cash than to ratify what the stock has already paid for and to expand the economic power expressed in its value (with the risk transferred to buyers). Under these conditions, the prevailing narrative, and the IPO pricing it produces, translate directly into economic power for the company, likely outweighing commercial success alone. The next two sections follow this dynamic: how economic power is bargained into adoption and narrative control, and where the revenue money flows from and to.

The Best Demand Money Can Buy: Strategies for Sustaining the Fervor

After examining the funding side of the company's finances, we turn to the demand for Anthropic's products: how it's sustained, at what cost, and how it measures up against the company's characterization of its market. In Anthropic's telling, the service it provides is on the cusp of becoming such a ubiquitously indispensable tool that its near-term opportunity reaches past the $4.3 trillion of global IT spend to an estimated $40 trillion of economic value in "the work that knowledge workers perform." This account implies organic demand and a natural market fit; yet the filing itself documents the battery of methods deployed to create and sustain it, on a scale only made possible by the exceptional economic power outlined above.

That battery combines traditional strategies at a new scale with more novel ones. On the more conventional side, Anthropic's marketing and advertising costs have risen every period on record to a $660M annualized rate in the first half of 2026. For context, the company discloses its significant customers (organizations spending $100k+ a year): 6,000 at mid-2026, 11,900 at the latest reported date. Pair the two series either way and the ad spend per $100k+ customer lands between $55k and $110k a year, which suggests an investment in public discourse, including the perceptions of investors and policymakers, rather than a typical revenue funnel.

This push is also supported by extensive in-house work to broadcast the company's framing and predictions of its services' economic roles. The filing highlights, without quantifying their cost, the company's policy proposals, an Anthropic Economic Index that takes the company's job replacement ambitions as a foregone conclusion to quantify under its own metrics, and an Anthropic Institute created to "deepen and broaden the reach of this work". Corporate funding of framing research is an established practice, infamous in particular for the role the tobacco and oil industries played in introducing counternarratives against growing awareness of their harms. The role here is inverted, manufacturing certainty about positive GDP and productivity outcomes rather than doubt. In addition, the "independent" economic analysts of the AI boom who have validated these framings have a very specific stake in the value of the company's IPO, as the twenty-two underwriters listed on the cover, paid a fee indexed on the sale value. Their research desks, drawing in part on the company's own publications, have spent years publishing the enterprise statistics and prognostics on which the filing's growth projections rest: yet another example of the dual-currency dynamic outlined in the previous section. Research produced outside this apparatus, and without anything like its amplification, stands in sharp enough contrast to at least put these prognostics in question: workslop studies documenting work displaced on colleagues rather than "augmented" (HBR, BetterUp/Stanford, The Guardian), findings outlining the rarity of profitable sustained adoption (roughly 5% of organizations in an MIT study), and METR's finding that perceived speed gains from AI code assistants outrun measured ones.

Another demand expansion strategy takes proven technology company playbooks to a new scale: free and discounted access to the company's products, combining broad public access with a strategic focus on groups who are likely to support the company's reputation and to drive second-order adoption. While the filing does not provide a dedicated budget line for free access as an advertising expense, it does acknowledge that it exists "in significant part to build awareness of, familiarity with, and engagement with our products." We estimate rough bounds on what this costs the company. On the lower side, a conservative estimate of 50M free users using 300k/100k input/output monthly tokens with Anthropic's own ratio of customer-serving compute to revenue of 15% would add up to over $200 million a year. On the higher side, the expense is included in the $3.8B sales and marketing line, of which channel partner fees and media spend account for $3.2B: at most the remaining $640M for the half-year could cover free access, a ceiling it shares with S&M personnel and overhead. Overall, this points to a few hundred million dollars a year just for compute to support free usage, developing a user base and capturing further training data to help align the model to user expectations. This strategy is extended in different ways toward educators and nonprofit organizations. US K-12 teachers, who are in a privileged position to normalize technology and shape new adults' digital skills, are eligible for a full year of free premium access, with a promise not to use their interactions as training data. Nonprofits, who play a dual role in introducing new technology to underserved communities and shaping public narratives on societally relevant topics, get 75% off Team and Enterprise plans. Both are harder to price.

The most novel strategies are the numerous channels through which Anthropic purchases or otherwise financially incentivizes both individual and institutional champions to push adoption. The filing highlights a few of these, including the Claude Partner Network launched with $100M to fund "partner training, technical certification, and joint market development with the firms that bring Claude into enterprise environments", and the Claude Frontier Academy, backed by another $100M, and aiming to train 10,000 engineers beginning with consulting-firm cohorts, "[to lead] a Claude deployment at their own organization". Beyond the filing, the company has announced an additional 1,000 Claude Corps fellowships backed by $150M to "teach 1,000 fellows how to use Claude well" and pay to place them inside nonprofits.

Other significant channels within this broad strategy are harder to price out individually. The filing mentions partnerships with major consulting firms Accenture, Deloitte, KPMG, and PwC to extend Claude's reach to their customer bases, but folds the corresponding channel fees, unsplit, into an estimated $2.8B total shared with the compute platforms. Accenture's relationship with Anthropic runs even deeper, as the filing commits "at least $1 billion over the next five years to build external reviewer capacity with Accenture" for "independent verification of compliance with future regulatory standards," the terms not yet final; while the latter is not a fee paid to push adoption, it does make the firm a direct beneficiary of Anthropic's continued growth. Ode with Anthropic, co-founded "in partnership with leading private equity firms," sells Claude implementations to companies these firms fund: a model of adoption through ownership rather than the market.

These items together offer a sharp counterpoint to Anthropic's own characterization of its product adoption. The filing foregrounds a narrative of organic adoption, supported by figures like a Net Revenue Retention of approximately 900%; but that figure depends on a scoping that qualifies customers with as little as $1 of prior revenue, excludes consumer subscribers, and starts from a base period predating the current surge, so the ratio soars by construction. Similarly, while the filing touts adoption among Fortune 500 companies, the average share of their total expenses committed to Claude remains under 0.1%. In contrast, the demand-expansion channels above carry individual costs in the hundreds of millions to $1B: the advertising line, the estimated cost of free access, and the championship programs add up, again, to the value of several "unicorns." User appetite for the service is real, as attested by surging revenue; but the filing cannot distinguish demand that is bought from demand that would exist anyway, and the intensity of the effort strongly hints that Anthropic's $40 trillion characterization describes a market that needs to be created rather than just tapped.

Round and Round the Money Goes: All Roads Lead to the Hyperscalers

In the first half of 2026, Anthropic reports a record revenue of over $16B. While remarkable, it is exceeded by another set of figures disclosed in the filing: multi-year commitments to its compute suppliers. Anthropic is on the hook for at least $114.5B in cloud spending with Google (through July 2033), at least $110B with Amazon (through April 2036), and at least $44.9B with Microsoft (through May 2033), each binding "regardless of actual usage," in addition to approximately $161.2B under five-year equipment leases with Broadcom, excluding interest. This creates an asymmetry: the commitments run for up to a decade, while the revenue that must service them is, the filing notes, "primarily consumption-based, and much of it is not subject to long-term contractual commitments." Since year-end 2025, new and amended compute and infrastructure agreements, the four above included, total roughly $660B in future minimum payments, with another $125-150B of 2028-2031 capacity whose terms are not yet final. That is more than thirty times the company's cumulative lifetime revenue of roughly $21B, and more than sixty times the $10.8B of revenue under contract (remaining performance obligations) as of June 30, 2026. The same structure extends to the energy supply: dedicated power arrangements, some with on-site natural-gas generation, run twenty-year terms with take-or-pay obligations and no cap on fuel costs, all applying regardless of utilization. The company has also publicly committed to covering the electricity price increases its data centers impose on consumers. These commitments rest on structural entanglements documented across the filing: how the suppliers underwrote the company's trajectory, and how they benefit in turn.

These relationships are difficult to describe because of their multifariousness: compute suppliers, particularly hyperscalers, are at the same time owners, suppliers, distributors, customers, and even competitors of Anthropic, all five roles the filing itself acknowledges without fully connecting the dots. Amazon's funding instruments (namely, convertible notes: cash now, shares later, at prices fixed when written) convert at the IPO into roughly a tenth of the company's shares, approaching a 12% ownership ceiling, while Google's ownership will grow toward its ceiling of 14.99% as its 2026 commitments are drawn over the coming years. Anthropic describes compute as its largest operating expense, reaching $8.9B in the first half of 2026, most of it directed by those commitments to the hyperscalers above. Claude models distributed through hyperscaler platforms rather than bought directly account for 48% of first-half revenue, a share growing year over year. These roles were cemented in the same negotiations as the commitments above: Google's $114.5B in March 2026 alongside $10B of equity and a $30B note facility, Amazon's $110B in April alongside $5B and a $20B facility. Together, they amount to deep dependence that materially constrains Anthropic's strategic leeway: the company's projections only work if it keeps growing its user base and if its services remain inherently compute-intensive enough for their price to cover the commitments. A trajectory favoring compute-efficiency over scale, however (e.g. smaller or more purpose-specific models), threatens to put the company in an untenable position, caught between its competitors on one side and its suppliers on the other. Narratively, this dependence also explains Anthropic's focus to date on "frontier" AI and on scale as a competitive advantage.

The commercial structure of the distributor role outlined above is particularly telling. For the 48% of revenue billed through the cloud providers' platforms, Anthropic pays service fees to these same firms; the filing acknowledges the fees, and folds them into an aggregated "channel partner" expense within sales and marketing without disclosing the amount directly. Reconstructing their size from that line and the disclosed revenue shares: roughly $360M of channel fees against $2.1B of platform-billed revenue in 2025 corresponds to an effective toll of 17%. Channel fees in the first half of 2026 rose to $2.8B, against $7.8B of platform-billed revenue; even allowing for up to $1B of that to represent fees to other channel partners, the distributor share lands between 23% and 36%, both ends far above the single-digit listing fees standard on cloud marketplaces, with Microsoft and Google each at 3%. This toll can be read through two lenses. The first is bargaining power: the fees indicate terms skewed in the compute providers' favor, unsurprising given the dependence above. The second is alignment: at this size, distributing Claude is a lucrative business in itself, which matters because these distributors also sell access to competing models, whether their own (Google's Gemini) or a partner's (Microsoft, as OpenAI's largest shareholder).

So far we have focused on what Anthropic owes and pays the hyperscalers, but the filing also documents the return flow: hyperscalers supporting a meaningful share of the $21B in cumulative lifetime revenue, either directly or through other organizations they fund. First, Anthropic acknowledges the hyperscalers themselves as customers, purchasing its services "for internal use or for their own products and services" at a scale it never discloses. Still, Google's commercial agreement is named in the notes, and Amazon's purchases run large enough that its multi-billion-dollar notes carry set-off rights against what Amazon owes Anthropic under their commercial agreements.

Second, hyperscaler-originated funds also support Anthropic's revenue less directly, through the broader customer base. The filing characterizes that base in ways that differ across sections. The Business section foregrounds the Fortune 500: every company in the Fortune 10 and over half of the list purchase at least $100,000 of Anthropic services a year, some relying on Claude for "mission-critical workflows." However, the statement later notes that the average Fortune 500 customer spends less than 0.1% of its expenses on the product, putting the cohort's aggregate at a low-single-digit share of Anthropic's revenue. The Risk Factors section, meanwhile, acknowledges that "a significant portion of our customer base consists of startups and technology-oriented enterprises," which "may reduce or delay spending during a downturn," and that early-stage customers "depend on venture capital or other external financing to fund their operations." The revenue figures favor the second reading: the top five customers alone accounted for roughly 21% of revenue in the third quarter, a scale no Fortune 500 customer approaches. The Business section itself names this population: "leading AI-native companies" building "next-generation coding tools, design platforms, and productivity applications" on Claude, for whom model usage is a cost of goods sold rather than a productivity tool. While the filing does not disclose more detail on this customer category, the ecosystem it belongs to feeds on hyperscaler funding through two channels. The first is direct funding, a channel at a record high: more than half of all capital raised by startups in 2025 came from rounds with at least one corporate backer, with big tech companies featuring prominently in the largest AI rounds. The second is in-kind incentives, with AWS and Google steering startups onto their platforms through credits of up to $200,000 and $350,000 respectively, the former redeemable on third-party models such as Claude, the latter bundled with an on-ramp to Google's marketplace.

For the hyperscalers, the arrangement is self-hedging. Between their stakes in Anthropic and Anthropic's obligations to their clouds, they win either way: a share of any AI company the startups build, or the invested funds returned as Anthropic revenue, to which their commitments, channel fees, and equity give them first claim.

Anthropic's interdependence with the hyperscalers gives it unprecedented access to funds and compute, which it has leveraged as substantial economic power, as outlined in the first section. But it also carries two less visible consequences that matter for the upcoming IPO. First, the dependency constrains as much as it supports: its committed expenses and structural ties narrow Anthropic's strategic options and force a race against its own compute bills. Second, it produces a three-tier ownership structure in which the shares offered to the public in the IPO are the least attractive tier. The first tier consists of the founders, who retain voting control through the proposed governance structure while holding a class of shares with no economic rights. The second consists of investors with a structural hold over the business, above all the hyperscalers, whose combination of commitments, ownership, and market control amounts to a first claim on the company's revenue streams. Public investors come last: they stand to benefit if the company's value keeps rising after the IPO, but only after their investment has helped meet the obligations already outstanding.

Conclusion

Filing for an IPO is an exercise in storytelling, a skill Anthropic has shown great proficiency in as a company. The tale it tells this time measures a fourteen-fold surge in revenue against compute expenses said, at last, to be coming under control: the former will keep expanding in perpetuity, capped only by a $40 trillion market, while the latter is wrangled into something manageable, all on the strength of unprecedented innovation meeting the digital economy's latent demand. Pull back the curtain, however, and the story loses its footing. The registration statement requires the company to finally give its own account of figures it had heretofore kept close to the chest, supporting an analysis considerably more critical of its preferred narrative. We use those to put the story back in its context of unprecedented economic privilege, extensive demand creation strategies that shape a market to fit the product instead of shaping the product to meet the market, and past successes bought at the cost of deep obligations to hyperscalers now positioned to claim the lion's share of the revenue the company reports. The latter is particularly relevant to prospective buyers weighing the attractiveness of Anthropic shares in this offering. First, because the filing spells out a three-tier ownership structure that severely dilutes the economic benefit of ownership for shareholders who are neither company founders nor holders of a structural claim on its profits. Second, because it reveals the company's real existential risk: not so much regulation, reputation, or catastrophic events of limited likelihood, but rather scenarios in which the cost of AI compute falls far enough that even the exponential growth the company is banking on fails to cover its commitments. From a broader societal perspective, that same dependence also puts the company's trajectory at odds with some of the major challenges of our time: energy scarcity, climate change, and growing resource disparities.

@misc{jernite2026critical,
  author = {Jernite, Yacine},
  title  = {A Critical Reading of {Anthropic's} pre-{IPO} Filing},
  year   = {2026},
  month  = nov,
  day    = {4},
  note   = {Annotated edition},
}