Anthropic’s IPO: $968 billion basic equity value and long-term compute obligations

This review uses the November 10, 2026 S-1/A Amendment No. 2, not the assumptions in the late-October filing. It remains a preliminary prospectus: $70–$78 is an estimated range and ANT a proposed Nasdaq symbol, not evidence of completed pricing, listing or regulatory endorsement.^[1]

1. Separate proceeds, secondary sales and valuation

The public offering comprises 825 million Class A shares: 801,351,354 newly issued by the company and 23,648,646 sold by existing holders. At the indicated range, total public gross proceeds would be $57.75 billion–$64.35 billion, including the secondary sale. Not all of that funds the company. At the $74 midpoint, primary gross proceeds are approximately $59.3 billion and secondary proceeds $1.75 billion; the issuer receives none of the latter.^[1]

A separate $26 billion concurrent private placement issues non-voting Class N at the IPO price. Google and NVIDIA each commit $10 billion, AMD $2.5 billion, Broadcom, Micron and Microsoft $1 billion each, and TSMC $500 million. Google's $10 billion is part of the $26 billion, not additive again. Closing remains conditional on the offering and specified requirements.^[1][2]

The issuer estimates combined net proceeds of approximately $84.6 billion at $74, or $93.7 billion if the underwriters fully exercise their additional-share option. This combines primary public issuance and the private placement, subtracts estimated issuer expenses and underwriting discounts, and excludes secondary proceeds.

Supplier investment demonstrates access to financing, not independently proven end-customer demand. Investors, chip suppliers, cloud providers and distributors can overlap. That is not by itself evidence of false revenue, but funding, procurement obligations and ultimate customer receipts must be analyzed separately to avoid counting the same capital repeatedly as demand.

2. Revenue, operating earnings and cash tell different stories

Nine-month revenue was $33.242 billion, versus $2.545 billion a year earlier. Q3 revenue was $16.956 billion, following $11.555 billion in Q2. Scale is no longer hypothetical.^[1]

Nine-month GAAP operating loss was $1.260 billion, while adjusted operating income was approximately $4.1 billion. GAAP net loss was much larger, $104.039 billion, including $94.324 billion of fair-value losses on convertible notes, derivatives and SAFEs. Non-cash remeasurement associated with a higher enterprise valuation is not current cash burn. Adding it back, however, does not eliminate the dilution economics of financing.

Q3 recorded a $567 million GAAP operating loss after Q2's $974 million operating profit. Adjusted operating income increased from approximately $2.2 billion to $2.5 billion, but adjusted margin fell from 19% to 15%. The company used approximately $5 billion of more expensive short-term compute in Q3 and recorded approximately $2 billion in non-cash donation-matching charges.^[1]

Consequently, a “90% incremental gross margin” should not become an assumed permanent company-wide GAAP gross margin. Compute and infrastructure are presented as operating expenses and include research, training, internal usage and customer service. The disclosure does not provide a clean traditional-software cost-of-sales allocation. Any marginal calculation needs explicit periods and exclusions rather than silently omitting distribution, personnel, additional capacity and replacement costs.

3. Why efficiency can improve while margins retreat

Compute and infrastructure expense was $9.863 billion in Q3, approximately 58.2% of revenue, versus $5.558 billion and 48.1% in Q2. These are disclosed expense-to-revenue ratios, not GAAP cost-of-sales percentages. They show that procurement mix can offset part of the efficiency improvement.^[1]

Compute is typically committed 12–18 months ahead, and short-term capacity currently costs substantially more than multi-year commitments. Additional capacity enables sales but introduces a step-up in cost. Revenue growing against a previously fixed cost base can produce an exceptional incremental margin; expensive new supply, discounts or insufficient utilization can reverse that effect.

The company estimates Q3 price reductions and discounts lowered sequential revenue by approximately $600 million. Efficiency gains may therefore accrue partly to customers rather than shareholders. The relevant unit is cash retained per paid, quality-adjusted task, not only the cost of an individual token.

Anthropic Direct separates data-center capacity from accelerator commitments, potentially securing long-lead facilities earlier while deferring chip choices. That can improve flexibility without eliminating leases or minimum purchases. Support for four chip families is also not proof of instantaneous, lossless workload substitution.

4. $900.9 billion is neither tomorrow's bill nor an irrelevant headline

The filing aggregates commitments outstanding at September 30, subsequent agreements and uncommenced leases to approximately $900.9 billion, substantially all compute-related. Approximately $451.6 billion is expected before the end of 2030, with substantially all the rest during 2031–2045. Approximately $70.7 billion falls in the twelve months after September 30, including $4.2 billion subject to specified commencement conditions or termination rights.^[1]

Other arrangements without fixed-dollar commitments have estimated payments of $150 billion–$175 billion outside that aggregate. Capacity, timing, pricing and conditions remain uncertain; they are not interchangeable with fixed debt.

Approximately $186.2 billion of the aggregate involves exercisable termination rights or pre-commencement conditions, such as counterparties securing financing, facilities or approvals. This is meaningful protection, not an unrestricted right to cancel every obligation at no cost. Remedies can include credits or rescheduling rather than termination.

A valid stress test matches timing. Subtracting lifetime undiscounted commitments from today's cash and declaring immediate insolvency is wrong. Ignoring future minimum payments because a quarter produced positive operating cash is equally wrong.

5. Substantial liquidity, supported substantially by financing

September liquidity was approximately $146.6 billion: $95.6 billion cash and equivalents, $36.3 billion short-term investments and $14.8 billion available credit. Rounded components differ slightly from the rounded total, and available credit is not cash held. Strategic partner note and equity capacity is conditional; part decreases when the private placement closes and must not be added in full to the $84.6 billion expected net proceeds.^[1]

Nine-month operating inflow was $6.569 billion, investing outflow $36.546 billion and financing inflow $110.735 billion. Much investing cash purchased short-term securities: purchases of $40.318 billion against sales and maturities of $6.900 billion. Exchanging cash for investments is not equivalent to consuming it all in training or facilities.

Cash property/equipment purchases were $925 million and leased-asset/equipment prepayments $1.955 billion. Subtracting these from operating cash yields a narrow Atlas bridge of $3.689 billion. This is not company-defined free cash flow and excludes all future lease principal and minimum compute payments. Operating improvement is real; the stock of liquidity nevertheless owes much to financing.

6. The share denominator changes the valuation

At the $74 midpoint and filing assumptions, post-offering Class A is 7,427,991,791 shares and Class N 5,658,680,800. Their combined economic basic shares are 13,086,672,591. Multiplication by $74 gives approximately $968.4 billion of basic equity value, calculated by Atlas—not enterprise value or a fully diluted valuation.^[1]

The denominator excludes outstanding options, remaining RSUs, certain convertible notes, matching obligations and other potential issuance. The filing's expanded capital bridge shows potentially material additions, some dependent on conditions, vesting and share price. Loss-statement weighted-average shares, the 825 million offered shares alone, or non-economic Class T/F shares are inappropriate shortcuts.

Primary issuance dilutes owners while adding cash; secondary sales do not fund the issuer. Because the $26 billion placement is fixed in dollars, its share count varies inversely with price. Applying the same fixed midpoint share count across the entire proposed range would not be strictly comparable.

As a sensitivity, not a forecast, $968.4 billion at 30, 40 or 50 times sustainable annual earnings would require approximately $32.3 billion, $24.2 billion or $19.4 billion of annual profit. This does not subtract cash, incorporate potential dilution or predict these profits. It illustrates the need for substantial enduring earnings attributable to common holders, beyond a temporarily exceptional incremental margin.

7. Economic ownership without corresponding control

The Long-Term Benefit Trust's non-economic Class T generally elects and removes a board majority. The founders' LLC holds non-economic Class F carrying 50.1% of general voting power. Class A has one vote per share; Class N is non-voting with specified conversion on unaffiliated transfers. The company expects controlled-company treatment.^[1]

This goes beyond a generic dual-class label: board selection and general shareholder votes are affected through different mechanisms. Public holders bear compute, dilution and operating risks with limited influence over strategy and capital allocation. Public-benefit duties can support long-term safety and research while conflicting with short- or medium-term shareholder returns.

Donation matching paid in shares is non-cash but economically dilutive. Concurrent private-placement shares also cannot be assigned the same lock-up assumptions as all existing holders: they are not covered by the same market-standoff or underwriter lock-up agreements, yet remain restricted securities subject to registration or exemption requirements. Neither immediate unrestricted trading nor a universal 180-day prohibition is established.

8. Three paths and the next falsifiable tests

The favorable case combines persistent enterprise spending, timely cheaper long-term supply, less short-term procurement and costs per accepted task falling faster than prices. Distribution and personnel efficiency improve. Revenue covers new minimums and financing becomes optional optimization rather than a gap filler.

The middle case retains strong growth but introduces margin steps as capacity arrives. Expenses and dilution mean per-share progress trails total revenue. IPO cash strengthens the buffer without immediately proving the valuation. Several quarters of operating cash, prepayments and dilution are needed.

The adverse case combines price competition and delayed capacity while long commitments precede revenue. Conditional partner financing leaves common holders exposed to additional funding needs. Lower utilization or customer expansion would consume liquidity. These are scenarios, not failures already established by the filing.

Next checks are the final offering terms and actual shares; cash received net of reductions in existing capacity; short-term compute mix; customer receipts and distribution costs; delivery and payment against the $70.7 billion twelve-month obligations; and continuing share issuance.

Conclusion

The documents do not show capital abandoning Anthropic: large financing commitments, operating cash generation and rapid revenue growth show both funding and demand. Nor do they show effortless software margins funding unlimited infrastructure. Procurement is becoming larger and longer-dated, public control is limited and dilution matters. The investment framework should center on duration matching, cost steps, per-share cash returns and governance—not one incremental-margin headline.

Source scope and limitations

Materials checked November 12, 2026. These sections and agreements belong to one issuer's disclosure system, not multiple independent commercial verifications. Complete customer/supplier contracts, a live bookbuilding record and the final effective offering document are unavailable. Valuation arithmetic and scenarios are Atlas analysis, not a price target, trading recommendation or guarantee of completion.