1. What is being offered—and why does it matter?
Has the IPO happened?
Anthropic filed its S-1 on 26 October 2026; proposed Nasdaq symbol ANT. Price and offer quantity remain blank. Filing is neither SEC approval nor a completed listing. [2] The S-1 cover lists Goldman Sachs and Morgan Stanley first among the underwriters. [1]
Where does this fit in investment banking?
The primary market is where new securities raise capital for the issuer. An IPO may also include existing investors selling shares. Those secondary-sale proceeds go to the sellers. Underwriters help with diligence, valuation, investor marketing, book-building, pricing and allocation; the prospectus makes the business, use of funds and risks reviewable.
Will the customer value created by AI translate into durable cash flow per share after computing costs, research, distribution and dilution?
2. How does Anthropic earn revenue?
Claude is sold through applications and developer access. Claude Code extends this into coding workflows; enterprise deployment adds administrative controls and cloud integration. Product capability is the vendor’s proposition, not proof that every customer achieves the same return. [8]
Revenue sources: latest disclosed breakdown
Source: S-1 Note 2; six months ended 30 June 2026, unaudited. The pie shows billing models, not an invented split by Claude product. [1]
| Billing model | How money is earned | Economic implication |
|---|---|---|
| Consumption | Metered use; spending follows workload | High expansion potential, but usage and realised pricing can change |
| Subscription | Access and usage capacity for a recurring fee | Predictable entry price; resource-intensive usage affects economics |
Do not add cloud distribution to this pie: it is a sales channel, not a third mutually exclusive billing model. A subscription or consumption customer can purchase through a cloud marketplace.
Why might this business scale differently?
Software can distribute the output of a relatively small team to many users. Agentic workloads can also execute many more tasks per customer. The tension is that each additional workload consumes resources: more use is valuable only if realised revenue covers incremental service cost and contributes towards research and fixed obligations.
3. How fast is revenue growing?
Actual reporting periods; FY bars cover 12 months, H1 bars six months. Compare H1 with H1 for growth. Annual results audited; interim results unaudited. [1]
H1 year-on-year growth: (16.286 ÷ 1.163 − 1) × 100 = 1,300.3%. This is about 14 times the previous level, not a 14% increase.
Does annualised revenue equal a forecast?
No. Recognised revenue records completed-period activity. A run rate extrapolates a short observation window; media run-rate figures are of this kind. [4] A forecast models future demand, pricing and capacity. Revenue can also be confused with bookings, customer commitments or cash collected; each measures something different.
Test growth by decomposing these drivers. Falling price can be offset by rising use, but only if capacity and unit economics support it.
4. Revenue per employee: extraordinary, but incomplete
Revenue per employee · approximately 5,300 employees [1]
What is the correct calculation?
The filing’s figure uses trailing revenue and a stated measurement date. Do not divide H1 revenue by September headcount and call the result an annual metric. For a fresh analyst calculation, average headcount is preferable when available.
FY2024 analyst calculation: $386m / approximately 940 employees = $0.41m. TTM September 2026: company-reported >$6.6m; bar uses the lower bound 6.6. Both use period-end rather than average headcount. Different reporting dates limit direct comparison. [1]
Does this prove superior profitability?
No. Labour productivity can look exceptional when computing, cloud services and outsourced infrastructure perform much of the work. Compare periods, headcount conventions and outsourcing before comparing companies. A high ratio measures sales scale relative to people; it does not measure return on all capital employed.
Research and engineering create the model and product.
Infrastructure trains models and serves workloads.
Channels and enterprise integration bring customer demand.
5. Which profitability measures tell the useful story?
| Metric | H1 2026 | Meaning |
|---|---|---|
| Operating margin | −4.3% | Operating result ÷ revenue |
| Net margin | −495.8% | GAAP net result ÷ revenue; heavily affected by financial remeasurement |
| Compute / revenue | 54.9% | Broad compute burden, not the conventional cost-of-revenue ratio |
| Operating cash / revenue | 24.1% | Cash conversion including working capital |
How can the company have a huge loss and positive operating cash flow?
The two statements answer different questions. The income statement includes accruals and revaluations; the cash-flow statement reverses non-cash items and incorporates collection and payment timing.
Actual reconciliation: H1 2026, unaudited · US$ billion
| Step | Item | Signed adjustment / balance ($bn) | Plain-language explanation |
|---|---|---|---|
| Start | GAAP net loss | -80.748 | Accrual-accounting result; includes substantial non-cash charges. |
| Non-cash adjustment | Share-based pay | +2.185 | Expense without equivalent current-period operating cash payment; dilution still matters. |
| Non-cash adjustment | Depreciation / amortisation | +0.097 | Allocates past investment cost; add back for operating cash reconciliation. |
| Non-cash adjustment | Financing fair-value changes | +72.388 | Remeasurement of convertible notes, derivative liabilities and SAFEs. |
| Non-cash adjustment | Donation-matching liability | +0.126 | Liability-related expense adjusted in the cash-flow statement. |
| Non-cash adjustment | Other financial-instrument recognition / revaluation | +8.720 | Separate from the financing adjustment above; not an operating cash outflow in this period. |
| Other reconciliation | Other | -0.007 | Residual reconciliation adjustment. |
| SUBTOTAL | Before changes in operating assets and liabilities | +2.761 | Subtotal, not another amount to add. |
| Operating balance change | Customer receivables | -4.982 | Revenue recognised faster than cash collected ties up cash. |
| Operating balance change | Prepayments / other assets | +2.094 | Net movement releases cash in the operating reconciliation. |
| Operating balance change | Supplier payables / accruals / other liabilities | +2.679 | Net increase supports current-period cash; obligations remain. |
| Operating balance change | Lease liabilities | -0.003 | Operating lease balance adjustment. |
| Operating balance change | Deferred revenue | +1.380 | Cash collected before the related revenue is recognised. |
| TOTAL | Net operating cash inflow | +3.929 | Reported operating cash generated; not the increase in total cash. |
Source: consolidated statement of cash flows, six months ended 30 June 2026 (F-12/F-13). Figures converted from US$ million to billion. Subtotal and total are balances, not additional adjustments. [1 · SEC S-1]
+2.761 + 1.168 net operating-balance benefit = +$3.929bn operating cash flow
The +$72.388bn adjustment does not mean customers paid that amount. It reverses a non-cash charge already included in net loss. Cash collection and payment timing then determine the operating cash result.
What should investors conclude?
Positive operating cash flow does not erase the accounting loss or prove that every expense is irrelevant. Share awards can dilute ownership; unpaid supplier amounts must ultimately be settled; customer prepayments create service obligations. Track whether these cash-flow supports persist.
New borrowing and equity issuance belong to financing cash flow, so they are not the explanation for this positive operating cash figure. Cash purchases of equipment belong to investing cash flow. Future compute commitments also require a separate funding assessment. Do not confuse operating cash flow with free cash flow or the total increase in cash.
Operating results divided by same-quarter revenue. Q1 −$1.667bn / $4.731bn; Q2 +$0.974bn / $11.555bn. Bar length is absolute magnitude; signs and colours identify loss versus profit. [1]
The move into a quarterly operating profit is meaningful, but one quarter is not a stable earnings record. Future model development, new capacity, price reductions and stock compensation can change the picture.
Why not call revenue less all compute expense “gross profit”?
Training, internal research and customer-serving compute have different functions. A broad expense classification cannot simply replace cost of revenue. Comparing gross margins across AI and software companies requires consistent definitions.
6. Projected revenue: what must happen for the story to work?
Reuters reported in August that Anthropic’s 2028 revenue projection was $190–200bn, according to people familiar with its finances. This is a media-reported projection, not an achieved result or a filled-in S-1 forecast table. [3]
Blue = historical result; gold = reported projection range, dated August 2026. These are annual periods. Omitted 2026–2027 bars indicate no assumed interpolation. [1,3]
Use this chart to examine the scale of the proposition, not as a continuous growth path. Test whether workloads, pricing, capacity delivery and customer economics can jointly support it.
Why is AI forecasting unusually sensitive?
- Demand can expand with capability: a better model may unlock tasks customers could not previously automate.
- Price and quality move together: cheaper models may increase adoption while reducing revenue per token.
- Agents can multiply usage: more automated steps can raise both billing and compute consumption.
- Research is recurring: maintaining competitiveness may require another expensive development cycle.
Bankers should value the downside as carefully as the upside. A large terminal market is not evidence of one issuer’s eventual share or margin.
7. Capex versus compute commitments: keep the categories separate
Cash purchases of property and equipment.
Multi-year purchase commitments.
Undiscounted leases not yet commenced; conditions differ.
These are different measures and horizons. They are not three components of current-year capex and should not be stacked into a capex total. Source: S-1 cash flow, commitments and lease notes. [1]
What does the capex projection show?
This article does not invent an annual company capex forecast. The lab below lets you test illustrative investment levels against illustrative cash flows. Actual spending will depend on deployment schedules, contractual conditions and financing choices.
Why does “asset light” need qualification?
Renting capacity instead of owning equipment can reduce reported cash capex while leaving substantial contractual cash obligations. An analyst must understand who owns the hardware, who guarantees capacity, when payments commence and whether commitments can be reduced.
Some demand can adjust quickly; capacity is ordered ahead of use. Underutilised infrastructure can therefore hurt cash flow even when a product remains popular.
8. Profitability Index: test investment value, not just margin
Is PI the same as a profit margin?
No. In capital budgeting, Profitability Index (PI) = present value of future net cash inflows ÷ initial investment. PI above 1 indicates positive NPV under the specified assumptions. It is not revenue growth, an operating ratio, or a reported Anthropic metric.
Compute investment stress laboratory
Five year-end inflows; no terminal value. Inflows are assumed after operating costs, tax, working capital and subsequent reinvestment. The initial investment occurs at time zero.
A simplified project appraisal. It does not value Anthropic shares or replace a full issuer cash-flow model. Changing any input recalculates PI and NPV.
What would a company-specific PI require?
A defensible schedule of incremental cash flows and investment, consistent treatment of leases and commitments, a supported discount rate and a clearly defined appraisal boundary. Without these, an exact “Anthropic PI” would be false precision.
9. How should an investment banker approach valuation?
Equity value → enterprise value adjusted for financing and other claims
Value per share → equity value ÷ consistently diluted shares
A revenue multiple can be a starting comparison, especially when earnings are negative. It cannot reveal the capital needed to support those revenues. A low multiple of an aggressive distant forecast may still represent a demanding valuation.
A simple expectation test
Suppose—not as a company forecast—equity value is $1trn and annual sales $50bn. That is 20× sales. If sales double and the multiple halves, equity value is unchanged before other effects. Shareholders need outcomes that outperform the expectations embedded in their purchase price.
Use scenarios for usage, realised price, service cost, research, distribution, reinvestment, working capital and dilution. Avoid a P/E multiple based on adjusted earnings that removes the costs central to the business model.
10. What can undermine the investment thesis?
| Risk | Financial transmission | What to watch |
|---|---|---|
| Capacity mismatch | Committed capacity underused as demand slows | Payment schedules, utilisation and cancellation rights |
| Competition | Prices fall or customers change model provider | Price per successful task and renewal cohorts |
| Partner dependence | Suppliers and distributors influence access and margins. Cloud marketplaces reportedly account for about 48% of H1 2026 revenue. [9] | Cloud fees, concentration and independent demand |
| Customer concentration | A few spending decisions affect sales. The top five customers reportedly account for about 25% of revenue in the nine months to September 2026. [6] [9] | Top-customer trends and enforceable commitments |
| Safety, reliability and IP | Claims, restrictions and remediation consume resources | Incidents, litigation and deployment rules |
| Talent and dilution | Retention costs rise; more shares share future returns | Stock awards, vesting and diluted capitalisation |
| Governance | Minority holders may have limited strategic influence | Voting rights and board-appointment powers |
Reuters describes cloud partners occupying several roles: supplier, investor, distributor and competitor. This can deliver reach while concentrating bargaining power. Such relationships require analysis; they do not by themselves establish improper accounting. [5]
How control is allocated. The S-1 sets governance by share class: Class A shares carry one vote each; Class N shares, held mainly by strategic partners, carry no general voting rights; the single Class T share lets the Long-Term Benefit Trust elect the Class T directors, who form a majority of the board; and the single Class F share gives the founders’ LLC 50.1% of general voting power. Together, the Trust and the founders’ LLC can elect the entire board. [1] [7]
Public Class A investors therefore buy an economic interest with limited control, which makes the distinction between economic ownership and governance rights especially important.
Is the customer base evidence of a moat?
The S-1 reports approximately 900% net revenue retention (NRR) and more than 11,900 customers generating at least $100,000 each in trailing-12-month revenue, as of 30 September 2026. These measure different aspects of adoption. [1 · SEC S-1]
What does 900% NRR mean?
It measures the revenue generated by the same qualifying customer group in comparable periods, including the effects of expansion, reduced spending and departures. Revenue from newly acquired customers is excluded from that group’s comparison.
Anthropic compares corresponding three-month periods one year apart. Its definition excludes Claude Pro and Max subscriptions. A qualifying customer had at least $1 of revenue in the base quarter.
Illustration: the same customer group, one year apart
| Illustrative measure | Amount / calculation |
|---|---|
| July–September 2025 revenue from the base customer group | $10m |
| July–September 2026 revenue from that same group, after expansion, contraction and churn | $90m |
| Net revenue retention | $90m ÷ $10m × 100 = 900% |
| Revenue increase | ($90m − $10m) ÷ $10m × 100 = 800% |
| Revenue from customers first acquired after the base quarter | Excluded from this same-group calculation |
What does the $100,000 customer threshold mean?
Each of the more than 11,900 qualifying customers generated at least $100,000 for Anthropic during the preceding 12 months. This is a historical spending threshold, not a statement that each customer signed a guaranteed $100,000 annual contract. It is separate from the NRR calculation.
Why can a small pilot make the ratio unusually high?
Hypothetically, a customer moving from a $10,000 trial quarter to a $90,000 deployment quarter generates 900% NRR. The same dollar increase from a much larger starting base would produce a lower ratio. This illustrates a denominator effect; it is not evidence that Anthropic’s reported growth is unsustainable.
| What to examine next | Meaning | Why it helps |
|---|---|---|
| Mature customer cohorts | Customers grouped by when they began, followed over several years | Shows expansion after the initial pilot-to-deployment jump |
| Customer churn | Customers that stop spending, alongside revenue lost | Large expansions can offset departures in an aggregate revenue metric |
| Customer profitability | Revenue after the relevant serving and acquisition costs | Spending growth is valuable only if its economics are attractive |
Interpretation: Existing customers are expanding spending rapidly. Assess the durability of that expansion alongside customer departures, spending by longer-established accounts and the cost of delivering the service.
11. What should the primary-market investor ask?
- How much of the offer is new capital and how much is selling-shareholder liquidity?
- What are the final price, fully diluted shares and post-offer claims?
- How much revenue is contracted, and how much depends on discretionary use?
- What happens if workload grows but realised price falls?
- Which infrastructure obligations survive a slowdown?
- Do cash inflows reflect recurring economics or temporary working capital?
- Can the business sustain research while earning a return on new capital?
- What control and information rights do public investors receive?
Evaluate customer value, the cost of sustaining capability, the conversion of growth into cash flow, governance and the offer price together. Strong technology and a good investment price are different judgments.
12. Sources, periods and calculation notes
Research cut-off: 29 October 2026. No live share price is implied. Recheck amendments before relying on final offering terms.
- SEC · S-1 · 26 Oct 2026
- Anthropic · filing announcement
- Reuters (via Investing.com) · reported 2028 forecast · Aug 2026
- Reuters (via Investing.com) · run-rate reporting · 17 Aug 2026
- Reuters (via Investing.com) · distribution and business model
- Moneycontrol · adoption and customer concentration
- AP · proposed IPO and governance
- Anthropic · Claude Code enterprise
- Stock Titan · S-1 guide: customer and channel concentration
H1 figures end 30 June; TTM employee productivity is measured at September. Ratios are author calculations from disclosed data. Forecasts and the PI laboratory are labelled separately. Charts include underlying figures or accompanying tables and print without external chart libraries.
Educational analysis. This article is not an offer document or an investment recommendation. Forecasts can change materially; the preliminary prospectus may be amended.