Anthropic’s IPO: Who Bears the Risk When the Compute Arrives?

Anthropic’s first public IPO filing describes a company that generated $16.286 billion of revenue, a $693 million operating loss and $3.929 billion of operating cash flow in the first half of 2026. Its $80.748 billion net loss is real accounting, but it is a poor standalone description of the business. Most of that loss comes from revaluing financing instruments, rather than paying to run Claude.

The question I followed through the October 26 S-1 was what happens after the next wave of computing capacity arrives. Anthropic is selling services customers can often scale down, while signing infrastructure contracts that can last for years. It proposes to sell public investors an economic interest while keeping decisive shareholder voting power with its founders and majority board-election rights with its Long-Term Benefit Trust.

My reading is that the central IPO risk is the gap between flexible customer spending and committed computing costs, under a governance structure public investors will have limited power to change. Exceptional demand can make that arrangement enormously productive. Slower demand, cheaper competing models or an expensive safety intervention can expose its rigidity. The filing supplies evidence for both outcomes.

This analysis uses the initial public S-1 filed October 26, 2026, accession 0001628280-26-067168. Annual financial statements are audited; the first-half and quarterly 2026 figures discussed here are unaudited. September operating metrics and subsequent-event disclosures have their own dates. I keep those periods separate.

What Anthropic has actually filed

Anthropic has filed a registration statement for a proposed initial public offering of Class A common stock. It has applied to list on the Nasdaq Global Select Market as ANT. Its announcement says the number of shares and the offering price range have not been determined, and the registration statement is not yet effective.

The distinction between company proceeds and stockholder proceeds matters. A large total offering would not necessarily deliver the same amount to Anthropic’s balance sheet. Until the quantities are supplied, a headline about the size of the IPO cannot answer how much additional funding the company will receive. See the prospectus cover and use of proceeds.

Start with three numbers, not the $80.7 billion loss

The income statement and cash-flow statement answer different questions. Treating their measures as interchangeable produces a misleading story in either direction.

Revenue grew about fourteenfold between the two first-half periods. Operating losses narrowed despite substantial increases in compute, personnel and distribution spending. Those are meaningful commercial improvements. They should survive a careful reading of the loss headline.

Of the $80.748 billion first-half net loss, $72.388 billion came from changes in the fair value of convertible notes, derivative liabilities and SAFEs. The filing attributes those changes primarily to increases in Anthropic’s enterprise valuation. It also reports a separate $8.720 billion noncash adjustment for recognition and fair-value changes in other financial instruments in the cash-flow reconciliation. These are distinct entries; combining them into one convertible-note charge would be inaccurate. See pages 23 and F-13, and the non-GAAP reconciliation on page 157.

That does not make dilution or financing terms irrelevant. A noncash charge can reflect economically valuable claims held by somebody else. It does mean the net loss should not be described as $80.7 billion of cash burned developing and serving Claude.

The operating result is a cleaner starting point for examining the service business. The cash-flow statement then tests how earnings translate into money. Neither resolves the future cost of infrastructure that has not yet entered service.

The margin table is in Note 8

My most useful accounting finding came from reading the expense presentation alongside the supplemental table in Note 8, page F-40. Anthropic changed its presentation in 2025: the face of the income statement groups expenses by their nature, including compute and infrastructure and personnel. It no longer separately displays cost of revenue and gross profit. The notes retain a functional breakdown.

That breakdown puts first-half 2026 cost of revenue at $2.380 billion, research and development at $9.741 billion, sales and marketing at $3.808 billion, and general and administrative expenses at $1.050 billion. They reconcile to the same $16.979 billion total operating expenses. The presentation change does not change the operating loss. Note 1 explains the reclassification; Note 8 supplies the table.

Subtracting $2.380 billion from $16.286 billion gives an 85.4% gross margin, calculated from the functional table. That is a useful measure of customer-serving economics under Anthropic’s allocation. It is not the margin left after funding the research needed to keep customers buying the next model.

Two qualifications make this reconciliation more informative than the gross-margin figure alone.

First, the company’s $8.945 billion compute-and-infrastructure expense is not its cost of revenue. It includes compute used for research, model development, customer service and internal workloads. Deducting all compute to calculate a conventional gross margin would use the wrong definition. Deducting cost of revenue and then all compute would double-count costs.

Second, distribution is expensive. The nature-based sales, marketing and partnerships line reached $3.168 billion. Management says its year-on-year increase included $2.7 billion of additional channel-partner fees; that is an increase, not a disclosure of total partner fees. The functional sales-and-marketing line is larger because it includes other allocated costs. Both presentations include these expenses in the operating result. See the MD&A, pages 146 and 150.

My interpretation is that the business has demonstrated attractive incremental customer-serving economics, while the cost of renewing its technical lead and distributing the product still absorbs almost all that gross profit. An investor needs to underwrite the entire cycle, including the next generation of research.

The quarter that made the bull case credible

The quarterly table on pages 158-159 shows revenue rising from $4.731 billion in Q1 2026 to $11.555 billion in Q2. The operating result moved from a $1.667 billion loss to a $974 million profit. That second-quarter profit is a GAAP operating result, not merely an adjusted earnings claim.

The same table shows compute and infrastructure expense rising from $3.387 billion to $5.558 billion. Revenue more than doubled while that expense rose by roughly 64%. Anthropic extracted much more revenue from its cost base. That is the strongest evidence in the filing for the argument that scale and engineering efficiency can improve the economics.

There is a timing qualification. Management says it typically contracts for long-term compute 12 to 18 months before availability. Costs enter as capacity becomes available, sometimes in large steps. The company also expects greater near-term use of more expensive short-term compute. It explicitly warns that margins may compress when new capacity arrives. See MD&A, pages 134-138.

The profit proves that a favorable combination of demand, capacity and spending can produce an operating surplus. It does not establish that each subsequent quarter will retain that combination.

A simple test of how much room there is

I calculated a static sensitivity using first-half revenue of $16.286 billion and expenses of $16.979 billion. It changes revenue while holding all operating expenses fixed. This is an illustration of exposure to revenue variation, not a forecast or an estimate of actual marginal costs. Some costs would change with usage, and future capacity, hiring and pricing could change the expense base substantially.

Under that assumption, about 4.3% more first-half revenue would have covered the reported expenses. Looking at Q2 alone, about an 8.4% revenue shortfall would have erased its reported operating profit if expenses had stayed unchanged. Those are calculations from historical statements, not predictions of a downturn.

This explains why utilization matters so much. Strong growth spread across an already-committed cost base can create operating leverage. A revenue miss against that same cost base can work in the opposite direction. The company's account of the timing of compute delivery is therefore central to interpreting its profits.

The real contract mismatch

Consumption-based revenue represented 80% of first-half 2026 revenue, or approximately $13.1 billion. Subscription revenue supplied about $3.2 billion. Some enterprise customers make annual or multiyear minimum-spend commitments, but the risk factors say much revenue lacks long-term contractual commitments while compute obligations extend over multiple years. See MD&A, page 139, and risk factors, pages 32-34.

That mismatch is the economic center of this IPO. A customer can optimize an application, route a task to a cheaper model or slow an experiment. Anthropic cannot necessarily unwind the computing contract that was signed to serve the anticipated demand.

The filing includes several measures of commitments. They have different dates, scopes and conditions, and should not be added together into a single liability.

The subsequent-events notes, pages F-64-F-66, add further dated agreements and estimates. They are not a license to combine every disclosed number. Some arrangements overlap other categories; some remain conditional or depend on commercial terms that are not finalized. The initial public filing leaves blanks in its updated aggregate summary and next-12-month payment total. I cannot turn those disclosures into a reconciled current maturity schedule.

There is a credible industrial logic behind Anthropic Direct, its move to lease chips and use dedicated facilities: the company says separating facility commitments from accelerator commitments lets it secure slower-to-build capacity while deferring larger hardware decisions. Lower-cost dedicated infrastructure could improve economics. The financial consequence is that more of the business depends on decisions made well before the revenue materializes.

A financing backstop that gets less generous under stress

The Broadcom-related financing facility deserves to sit beside the computing commitments, rather than in a separate mental box. The Infrastructure Partner 2026 note-purchase agreement permits up to $42 billion of potential issuances to finance quarterly rent under the 2026 lease arrangement, subject to conditions. The filing does not describe that amount as cash already borrowed or unrestricted cash available for any purpose.

The terms on pages 166-168 include post-IPO conversion at 90% of a trailing 30-day volume-weighted average share price. Notes issued when specified revenue-performance conditions are not met carry 13% annual interest and convert at 85% of the applicable price. The ability to issue notes is itself conditional and can terminate after sustained failures of specified conditions. The company does not expect notes under this facility to be issued before the offering. See the MD&A financing tables and footnotes.

My inference is that the backstop does not remove utilization risk; it can change its form. If cash generation disappoints, funding rent with convertible notes could create interest and dilution costs. The stated stressed terms make that funding more expensive precisely when revenue performance is weaker. That deserves more attention than a simple tally of available capital.

Cash is substantial; the date and source matter

At June 30, cash, cash equivalents and short-term investments totaled $131.651 billion. The first-half cash-flow statement records $110.432 billion of net proceeds from redeemable convertible preferred stock. The size of the cash reserve therefore reflects a very large financing operation alongside improving commercial performance. It is not evidence that the service business generated that reserve through earnings.

Operating cash flow was positive $3.929 billion. Subtracting $482 million of property-and-equipment purchases leaves $3.447 billion, a calculated historical cash-flow subtotal before other investing activities and finance-lease principal payments. I would not extrapolate it into future free cash flow. Management expects the infrastructure buildout to produce significant cash outflows in operating and investing activities, potentially even as financial performance improves. See pages 162-163 and F-13.

The $17.181 billion investing outflow also needs care: the statement includes $18.972 billion of purchases of short-term investments and $2.610 billion of sales and maturities. Moving cash into securities is different from spending it on data centers. Counting every investing outflow as infrastructure burn would overstate the latter.

Finally, the $131.651 billion reserve is a June observation. The September liquidity summary is left blank in this initial filing. Dividing June cash by a later, incomplete commitment total would give an apparently precise coverage ratio built from incompatible observations.

The customer evidence is impressive, but 900% is not a guarantee

As of September 30, Anthropic reports more than 11,900 customers generating at least $100,000 of trailing-12-month revenue, more than 1,900 at $1 million, and approximately 250 at $10 million. Its reported net revenue retention is approximately 900%. Those figures provide a strong demand-side counterweight to the obligations. See pages 135-136.

Read the definition on pages 3-4. The retention metric compares the same qualified customers’ revenue in the quarter ending at the measurement date with their revenue in the corresponding quarter a year earlier. A qualified customer needs at least $1 of base-period revenue; Claude Pro and Max subscribers are excluded. It measures expansion, contraction and churn across that cohort. It is not the percentage of customers retained, and it does not mean contracted future revenue is nine times larger.

My interpretation is that it demonstrates unusually powerful expansion from an early adoption base. It cannot establish the mature growth rate. The company itself expects revenue growth to decline over time.

Concentration remains relevant: the top five customers supplied about 25% of revenue in the nine months through September, although none exceeded 10%. Separately, cloud marketplaces accounted for 48% of first-half revenue. Those are different measurements: one concerns end customers; the other concerns distribution. The company recognizes marketplace sales gross because it says it is the principal, maintaining customer contracts, providing the service and setting prices. The cloud platforms collect and remit payments. See risk factors, pages 57-58, and Note 1, F-17 and F-21-F-22.

Cloud partners can consequently be investors, financing providers, infrastructure suppliers and distribution channels, sometimes while offering competing AI products. This creates interdependence and possible conflicts. It does not, by itself, establish fictitious revenue or improper accounting. The diligence question is how durable independent customer demand and access to distribution remain if partner economics change.

Safety can affect the income statement immediately

One disclosure connects the mission directly to revenue quality. Anthropic estimates that customers it subsequently determines violated its usage policy or commercial terms represented 5% to 7% of revenue in the nine months ended September 30, 2026. It recognizes revenue when service is delivered before detecting, investigating and restricting the activity. Management expects the share to fall after stronger enforcement, but the timing and revenue impact are uncertain. See MD&A, pages 145-146.

That estimate covers violations of terms as well as usage policy. It does not mean every affected customer committed a crime. Nor can I apply the percentage to first-half revenue to manufacture a dollar amount: the periods differ, and the filing does not provide the necessary nine-month revenue total in that disclosure.

The commercial question is whether compliant demand replaces usage removed by enforcement, and on what timetable. A company with committed capacity can incur the cost of responsible enforcement before it finds replacement revenue. That is a concrete route by which safety decisions can affect utilization.

The risk factors also disclose an October 2026 FTC civil investigative demand concerning consumer protection, data protection, privacy, and product safety and security. An investigative demand is not a finding of misconduct. It is, however, an actual event rather than a purely hypothetical regulatory risk. See risk factors, page 52.

Public shareholders buy economics with limited control

The governance section, pages 242-251, separates economic ownership from several forms of authority. These are the arrangements described for completion of the offering, not a claim that every contemplated issuance has already occurred.

The founders’ general voting majority and the trust’s board-election power operate on different tracks. Saying the founders simply elect the entire board would be wrong. Saying Class A shares have no vote would also be wrong. The consequential fact is that public shareholders cannot outvote the founders on general matters while the Class F arrangement operates, and do not elect the trust’s majority of directors.

There are important qualifications. The trust’s removal of a Class T director without cause generally requires concurrence from the nominating and governance committee or a specified employee vote; the trust agreement can be amended under defined thresholds. The trust does not manage the business day to day. The founder structure also has transition provisions. It is neither ordinary public-shareholder control nor an unalterable safety guarantee.

Anthropic’s stated rationale is coherent: keep a consequential technology from being managed solely to quarterly financial pressure. The investment implication is equally clear. Public benefit decisions may reduce revenue, increase costs or dilute shareholders. Buyers must value a security whose governing purpose deliberately includes objectives beyond their financial return.

The public-benefit allocation needs its own dilution analysis

Anthropic says it contributed 846,840,960 Class N shares to its wholly owned Public Benefit Fund, representing 5% of fully diluted shares as of June 30 under its stated Amazon-note conversion assumption. The offering assumptions distinguish legally outstanding shares held by that subsidiary from shares treated as outstanding for accounting purposes. This is not equivalent to an immediate 5% cash distribution to an unrelated charity.

The board has also approved a discretionary framework allowing additional Class N contributions of up to 2% at successive 50% share-price appreciation milestones, subject to trading-day tests. Future contributions are not automatic. The contributed shares may subsequently be transferred, donated or sold to advance public-benefit initiatives. See pages 144-145 and the offering assumptions on page 20.

This mechanism deserves analysis separately from employee compensation and the founders’ personal charitable pledges. A founder donating existing personal shares reallocates that founder’s holdings. Company share contributions and their later disposition have different implications for other holders. The prospectus itself warns of potential dilution from public-benefit initiatives. It would be premature to attach a definitive post-IPO per-share cost while the fully diluted offering capitalization remains incomplete.

The IPO will also change how employee pay reaches earnings

One reason to be careful with the profitable-quarter narrative is that the offering itself will trigger an accounting change for employee awards. The risk factors on pages 65-66 say that, had the offering completed on June 30, Anthropic would have recognized $2.9 billion of cumulative RSU compensation expense, with another $23.7 billion to recognize over a weighted-average period of about 2.1 years. The performance condition had not been satisfied, so that RSU expense was not yet recognized at June 30. These figures do not cover all other awards or all later grants. See the risk factors and Note 11.

That is a June hypothetical, not a forecast of the exact IPO-quarter charge. Later grants, modifications and final offering terms can change the amount. It does show why the first public-quarter earnings will require another reconciliation. A newly recognized noncash compensation expense does not establish that customer-serving economics suddenly deteriorated. Excluding it from an adjusted result does not make employee equity free to the other shareholders.

The amendments that will decide the investment case

The initial filing establishes extraordinary adoption and a business capable of quarterly operating profitability. It does not establish an investable price. My next reading will focus on information that could strengthen or weaken the utilization thesis:

1. The updated commitment schedule. A reconciled total, amounts due within 12 months, lease-versus-service treatment, commencement dates and conditional obligations would make the funding burden assessable.
2. Current liquidity and financing availability. September cash, restricted amounts and draw conditions matter more than treating the June reserve as permanent protection.
3. Quarterly margins after capacity enters service. Revenue, compute cost and operating profit should be assessed together, rather than celebrating one profitable quarter or assuming one weak quarter disproves the model.
4. Compliant customer expansion. The next disclosures should show how enforcement, top-customer concentration and cohort growth evolve. A spectacular retention number needs a continuing definition and comparable periods.
5. The final security and price. Primary proceeds, secondary sales, conversion mechanics, RSU settlement, fully diluted shares and potential public-benefit dilution determine what each public share represents.

A sustained improvement in operating margins after major capacity deliveries, alongside growing compliant usage and a clear funding schedule, would support the bull case. Capacity costs rising faster than durable customer spending, or reliance on increasingly dilutive conditional funding, would weaken it. Those are observable tests, not a target share price.

The distinction I would keep in view is simple: selling more intelligence can be highly profitable at the margin; committing to the infrastructure that makes it possible is a different financial decision. Anthropic’s IPO asks public shareholders to participate in both, within governance arrangements that limit their ability to redirect the second. The offering’s eventual price must compensate for that whole package.

For background, the Anthropic ownership investigation and valuation history cover separate questions. Their dated observations should not be substituted for the still-unpriced IPO.

Methodology and calculation notes

I reviewed the October 26 public S-1’s cover, summary, financial statements and notes, operating discussion, selected risk factors, financing terms and governance provisions, and checked the registrant and accession against the SEC filing index and EDGAR submission record. The company’s same-day announcement independently confirms the proposed listing terms and unpriced status. Page references are the prospectus’s printed page numbers, including F-numbered financial pages, rather than a browser’s page count.

All financial figures are USD. Billions shown in tables are converted from the filing’s millions. Gross profit is revenue less Note 8 cost of revenue; expense ratios divide the stated expense by same-period revenue. The static sensitivity changes revenue by 10% and deducts unchanged expenses. Its break-even calculations divide the operating loss or profit by corresponding reported revenue. Calculated values are rounded and are not company forecasts or company non-GAAP measures.

I have not independently audited Anthropic’s accounts, reconciled all individual infrastructure contracts, reviewed every exhibit or compared every confidential draft. The initial filing contains blanks that prevent a complete current commitment and per-share valuation model. This article offers document-based analysis, not a recommendation to purchase an unpriced offering. Later amendments may change the facts or the assessment.