Anthropic IPO: An MCDA Assessment

Analysis outlook: conditionally allocatable. On a consistent TOPSIS method, Anthropic scores a closeness coefficient of about 0.50 in the base case and about 0.35 if API pricing compresses 40% with usage flat. In Venture Science's framework, 0.40 is the house threshold below which an ECM allocation is not made. The offering clears it today, with no margin for a margin miss.

The case rests on three facts the filing establishes and one it does not:

• Growth is real and the operating business is near breakeven. First-half 2026 revenue was $16.3 billion against an operating loss of $693 million. Third-quarter revenue reported to investors was $16.8 billion, up 45% quarter over quarter.
• The $80.7 billion net loss is an accounting artifact. Roughly $72 billion is a non-cash fair-value charge on convertible instruments that grows as the private valuation rises. It will keep growing until the notes convert at the IPO, then largely disappear.
• Compute obligations are the balance sheet. Disclosed commitments run from at least $518 billion to roughly $637 billion over eight to ten years, about 80% non-cancelable. That is a $50–65 billion annual payment floor against roughly $67 billion of annualized revenue.
• Gross margin is still not pinned down. Pre-filing estimates clustered at 44–50%. The compute floor is survivable at 50% and not at 35%. This single number decides the investment case.

What this means in practice for an allocator: participate only at a discount to the reported $1.5–2 trillion range that compensates for a governance structure in which, as filed and reported, public holders control neither the board majority nor stockholder votes, size the position to the first lock-up window of 27 days rather than a conventional 180, and hedge. If the gross margin disclosed in the amended filing is below 40%, the position is not allocatable at any price in the reported range.

Much of the public discussion mixes filed facts with press reporting. The distinction matters for scoring, so the table separates them.

If the reported range holds, the IPO would price at roughly 1.5–2.1x the May private round, six months later.

Anthropic's revenue growth is without precedent among issuers of comparable scale, and the mix is healthier than a pure API story. The question the filing leaves open is what that revenue costs to produce.

What is strong

Revenue growing four to five times faster than compute cost is the operating-leverage signal the market has been waiting for from a frontier lab. A fifth of revenue from subscriptions and more than a quarter from outside the US reduce dependence on any single enterprise channel.

What is unresolved

1. Gross margin. The filing says substantially all cost of revenue is compute and infrastructure. Pre-filing analyst estimates put gross margin near 44% (Morningstar) to roughly 50% (company projections shared with investors), with a stated path toward 77% by 2028. The audited figure in the amended filing will reset consensus in one direction or the other.
2. Gross versus net. Channel partner fees paid to Amazon, Google and Microsoft for marketplace distribution are booked in sales and marketing ($3.8 billion in H1) rather than netted from revenue. Headline revenue is therefore not directly comparable with peers that report net of partner take.

Neither point undermines the growth. Both determine whether the growth can carry the compute obligations discussed below.

The headline loss is a valuation artifact, and the operating line is within rounding of breakeven. Readers should separate the two.

Sources: IBTimes, Motley Fool, WION, Crypto Briefing.

Why the loss is so large

Anthropic raised capital through convertible notes and similar instruments carried at fair value. As the enterprise valuation rose from the 2025 rounds to $965 billion in May 2026, the liability was marked up and the increase ran through the income statement as a loss. The higher the private mark, the larger the GAAP loss: $72.4 billion of the $80.75 billion is this charge.

What happens next

Expect a further large charge in the second half of 2026 as marks move from $965 billion toward the reported IPO range. At the IPO the instruments convert to equity and the charge stops. The relevant run-rate for valuation is the operating line, which the company projected to investors would turn positive in the second quarter, and operating cash flow, reported as positive in coverage of the filing. Neither is yet confirmed by audited quarterly figures.

The caution

A business that is breakeven on $17 billion of half-year cost has no cushion. Every dollar of the compute floor in the next section comes out of a margin that does not yet exist in audited form.

Anthropic has committed to between $518 billion and roughly $637 billion of compute, cloud and infrastructure spend over the next eight to ten years, most of it payable regardless of usage. This is the balance sheet, and it is the line on which the investment case turns.

Form S-1 text and reporting on it (Motley Fool, Fortune, CNBC) · as of October 26, 2026.

The filing's own contractual commitments table was left blank in the public version, so the counterparty figures above are drawn from the S-1 text and from reporting on it (Motley Fool, Fortune, SEC). Fortune tallies the total at $542–567 billion; earlier reporting on the draft put it at $518 billion. Beyond these, Anthropic has committed to procure an estimated $125–150 billion of additional TPU compute for 2028–2031 that sits outside the table.

Terms that matter

The Google and Amazon agreements are non-cancelable except for the provider's uncured material breach, and Anthropic must pay the shortfall if actual spend falls below the contracted amount. The SpaceX arrangements, by contrast, are mostly cancelable on 90 days' notice and carry a run-rate of about $3.1 billion a month once all capacity is delivered. That is the only large flexible tranche, and the first one that reprices if demand slows.

The arithmetic

Spread evenly, $518–637 billion over eight to ten years is a payment floor of roughly $50–65 billion a year. Annualized third-quarter revenue is about $67 billion. At a 50% gross margin, today's revenue covers the floor with little to spare. At a 35% gross margin, or with a 40% price cut that is not offset by volume, it does not. Rating agencies and credit analysts will treat the non-cancelable portion as a debt equivalent, and so does this framework.

Mitigants the bear case understates

Usage, not price, has driven revenue so far, and the orchestration layer the filing describes lets Anthropic shift workloads across Trainium, TPU, Nvidia and AMD silicon. Vendor financing is substantial: Amazon has agreed to provide up to $15 billion in $5 billion tranches, Broadcom up to $42 billion and Google $20 billion in potential financing. These reduce near-term liquidity risk; they do not reduce the obligation.

Anthropic's three largest cloud counterparties are simultaneously its investors, its suppliers, its distribution channels and its competitors. The filing says so directly, naming Amazon, Google and Microsoft as competitors that develop and deploy rival models through their own cloud platforms, and describing the compute relationships as creating dependencies and potential vulnerabilities (CNBC).

Three consequences follow for an allocator:

1. The leases are not arm's-length. A supplier that is also a shareholder has reasons to extend terms in a downturn, which softens the compute floor, and reasons to prioritise its own models, which hardens it. Neither is priced in a conventional lease analysis.
2. Channel economics are opaque. Marketplace fees to the same three companies sit in sales and marketing. Whether those fees are fixed or scale with revenue determines how much operating leverage the business really has.
3. The dependence runs both ways. SpaceX now relies on Anthropic for a third of its revenue. Concentration of this kind tends to stabilise relationships in good times and to transmit stress quickly in bad ones.

If the reported founder share class is adopted as described, public Class A holders will have a decisive vote on neither the board majority nor most stockholder matters. That is a more complete exclusion than SpaceX's dual-class structure, though with a different purpose and different safeguards.

Sources: SEC Form S-1, TechCrunch via AI Weekly, iTechGuides.

The two layers

The Trust controls who sits on the board. The founders, through Class F, control the outcome of stockholder votes. Neither group is answerable to Class A. Because the Trust can elect a board majority, Anthropic qualifies as a controlled company under Nasdaq rules (Unite.ai).

What mitigates it

The filing states the company expects to rely only on the nominating-committee exemption, and that the board has a majority of independent directors with fully independent audit and compensation committees. LTBT trustees hold no equity and do not share in profits (WSJ via iTiger). Control is therefore split across a non-economic trust and seven founders who each own roughly 2%, rather than concentrated in one person. The public benefit purpose, responsibly developing advanced AI for humanity's long-term benefit, is a statutory obligation to balance, not a licence to ignore stockholders.

What does not

The trust agreement and its arrangements with the company are not public, including when and how stockholders could override the Trust. Class F is new and its sunset terms are undisclosed. Proxy advisers will flag both. Index eligibility should be checked against each provider's current multi-class rules rather than assumed.

The framework scores governance at 1.5 of 5: a structural discount that no amount of revenue growth offsets, partly relieved by the independence of the committees that oversee the numbers.

The one live comparable for a mega-cap listing with a ~5% float is SpaceX, and it argues that any scarcity premium at the Anthropic open will be short-lived.

SpaceX, June 2026

SpaceX priced its IPO at $135, raised roughly $75–86 billion, opened at $150 and rose about 19% on day one (Nemo, The Edge). The stock peaked above $225 on June 17, then fell to the $135 IPO price by July 15 with only about 4% of shares trading (AI Weekly). By mid-2026 it had traded roughly 36% below its post-listing high. The premium lasted about five weeks.

Why Anthropic's supply picture is tighter, not looser

• Some employee and early-investor selling is permitted on the first day of trading.
• Lock-up releases follow at 27, 48, 62 and 76 days, rather than a single 180-day release (CNBC).
• Selling stockholders are part of the offering, so the primary raise is not the only source of float.

On a mid-November pricing, the four releases land between mid-December and early February, the same window in which the fourth-quarter fair-value charge and the first public earnings report arrive.

The implication

The usual trade (accept full allocation, let the squeeze run, collar the position) still applies. But the collar should be sized to the first lock-up window of 27 days, not to the conventional 180, and any thesis that counts the lock-up structure as a stabiliser has it backwards.

Anchors

Press accounts of a $10 billion Nvidia commitment and a $5 billion Microsoft commitment refer to the 2025 financing at a ~$350 billion valuation, not to the IPO book (MarketScreener). No IPO cornerstone investors have been disclosed in the filing.

At the reported $2 trillion, Anthropic would price at roughly 30 times annualized third-quarter revenue, about where SpaceX priced on its own run-rate in June. The multiple is defensible only on the growth rate, and the growth rate is the input most exposed to pricing.

Projections per Futurum; run-rate per WION. Projections are management figures reported in the press, not filed guidance.

A reverse-DCF framing

For $2 trillion to be fair value at a 25x terminal free-cash-flow multiple, the business needs about $80 billion of annual free cash flow. At a 25% free-cash-flow margin, that is roughly $320 billion of revenue, more than four times the 2028 projection. Investors paying the reported range are therefore underwriting either a decade of sustained hyper-growth beyond 2028, or a margin structure well above anything a compute-intensive business has yet shown.

The price-compression test

The sensitivity in Appendix C models a 40% cut to API pricing with usage flat. That is the open-source-commoditisation scenario. Under it, revenue quality falls from a score of 4.2 to 2.8 and balance-sheet durability from 1.6 to 1.3, because the compute floor is fixed while revenue is not. The closeness coefficient drops below the 0.40 house threshold.

What a discount should compensate for

Two structural deficits are permanent at any price: the governance exclusion of Class A, and the debt-equivalent compute floor. Venture Science's framework asks for an entry discount of 15–20% to the reported range to offset them, which at the top of the range implies roughly $1.6–1.7 trillion. That is a framework output, not a prediction of where the deal prices.

On the corrected facts and a consistent method, Anthropic scores 2.74 of 5.00 on weighted utility and a TOPSIS closeness coefficient of 0.498 in the base case. The 40% price-compression case falls to 0.347.

Reading the result

A closeness coefficient near 0.50 means the offering sits roughly equidistant from the ideal and the worst-case profile in this framework. It is allocatable under Venture Science's 0.40 house threshold, but the margin is small and rests on one undisclosed number. Below 0.40 the framework does not allocate; the price-compression case lands there.

Conditions for participation

1. Audited gross margin in the amended filing at or above 40%. Below it, the position is not allocatable in the reported range.
2. An entry price that discounts the reported range by enough to compensate for the governance exclusion and the compute floor.
3. Position sizing and hedging calibrated to the 27-day first lock-up release, not a 180-day assumption.
4. A re-score when the contractual commitments table, the trust agreement terms and the Class F sunset provisions are disclosed.

What would move the outlook upward

A disclosed gross margin above 50%, confirmation that marketplace fees are largely fixed, or a cornerstone book that materially extends the effective lock-up.

What would change it downward

Evidence of API price compression in the third-quarter numbers, a contractual commitments total above the $637 billion upper bound used here, or any provision allowing the Trust or founders to block a change of control without a stockholder vote.

The framework combines three standard multi-criteria methods: the Analytic Hierarchy Process (AHP) to set criterion weights, Multi-Attribute Value Theory (MAVT) to convert disclosures into utility scores, and TOPSIS to express the result as a distance from ideal and worst-case profiles.

Step 1: AHP weighting

Criteria are compared pairwise on Saaty's 1–9 scale and the principal eigenvector of the comparison matrix gives the weights. For mega-cap technology offerings the hierarchy and weights are:

Sub-criteria weights sum to their parent's weight. A consistency ratio below 0.10 is required for the pairwise matrix to be accepted. The underlying pairwise matrix and its consistency ratio are maintained in Venture Science's scoring workbook and are not reproduced here.

Step 2: MAVT scoring

Each criterion receives a utility score from 1.0 (critical risk) to 5.0 (ideal quality), anchored to the filing. Sub-criteria are scored first and aggregated by their local weights. Appendix B gives the rationale.

Step 3: TOPSIS

Each score is multiplied by its weight to give a weighted value vector V. The positive ideal solution (PIS) is the vector of weights × 5.0; the negative ideal solution (NIS) is weights × 1.0. Euclidean distances to each are computed and the closeness coefficient is:

A coefficient of 1.0 is the ideal profile; 0.0 is the worst case. Because the weights differ across criteria, the coefficient is not a linear function of the weighted utility sum: a weak score on a heavily weighted criterion pulls the coefficient down faster than the simple average would suggest.

House threshold

Venture Science treats a closeness coefficient below 0.40 as non-allocatable for an unhedged long position in an equity capital markets transaction. This is a firm policy, not an industry standard.

Scores are anchored to the public S-1 and to reporting on it as of October 26, 2026. Where the filing leaves a figure blank, the score reflects the best available reporting and is flagged for re-scoring.

Changes from the preliminary scoring

An earlier internal draft scored C2 at 3.2 and C4 at 2.0, on the assumption of a $245 billion compute commitment, a single 180-day lock-up and a Nvidia anchor in the IPO book. The public filing established commitments at least twice that size, a staggered lock-up with day-one selling, and no disclosed cornerstones. C3 moved from 1.4 to 1.5 to credit the independent committees, offset by the founder class.

All three cases below use the same distance method, so the coefficients are comparable with one another.

Ideal and worst-case vectors

With weights (0.35, 0.25, 0.25, 0.15), the positive ideal solution is PIS = (1.750, 1.250, 1.250, 0.750) and the negative ideal solution is NIS = (0.350, 0.250, 0.250, 0.150).

Base case

Sensitivity: 40% API price compression, usage flat

Revenue quality falls to 2.8 as top-line growth decelerates mechanically; balance-sheet durability falls to 1.3 because the compute floor is fixed while revenue is not. Capital structure and governance are unchanged.

Comparison with the preliminary scoring

The earlier internal draft reported a base-case coefficient of 0.488 and a stress-case coefficient of 0.378. The 0.488 was the normalised weighted utility, (2.955 − 1) / 4, not a TOPSIS distance; computed on the same distance method as the stress case, the preliminary base case is 0.534. The table below restates all three on one method.

Break-even on the house threshold

Holding C2, C3 and C4 at revised base-case levels, the coefficient crosses 0.40 when C1 falls to roughly 3.3. In the framework's terms, that corresponds to an audited gross margin in the high-30s with no offsetting volume response, the number flagged in the summary as the condition for participation.

The framework is a disciplined way of making judgments explicit. It does not remove them. Readers should weigh the following.

1. Single-alternative TOPSIS is a benchmark score, not a ranking. TOPSIS is designed to rank several alternatives against each other. With one issuer and synthetic ideal and worst-case vectors, it reduces to a weighted distance score. A full application would score Anthropic alongside a peer set: SpaceX as the listed comparable, OpenAI once it files, and capital-intensive infrastructure names such as CoreWeave as the bear-case reference.
2. The weights overlap on compute. Pricing power (C1.3), adjusted free cash flow (C4.1) and off-balance-sheet intensity (C4.2) respond to the same underlying risk. The price-compression scenario therefore registers in three places. The result is directionally right but the magnitude of the swing is partly an artefact of the hierarchy.
3. Scores are anchored to incomplete disclosure. The public filing leaves the contractual commitments table, share count and price range blank. Counterparty figures are drawn from the filing's narrative text and press reporting on it. The trust agreement and Class F terms are undisclosed. Every score touching those items is provisional.
4. Projections are management figures reported second-hand. The 2028 revenue and cash-flow projections, and the gross-margin path, were shared with investors and reported in the press. They are not filed guidance and should not be read as such.
5. The house threshold is a policy choice. The 0.40 cut-off reflects Venture Science's risk tolerance for unhedged ECM positions. Other allocators will set it elsewhere.
6. AI assistance. The analysis was prepared with assistance from an AI model developed by the issuer. The facts were checked against the cited sources, but the content has not been independently verified, no responsibility is accepted for its accuracy or completeness, and the conflict is disclosed at the top of this note.

Primary

• Anthropic, PBC — Form S-1, SEC EDGAR
• Anthropic — S-1 filing announcement, October 26, 2026

Reporting on the filing

• CNBC — Anthropic IPO live updates
• Investing.com — Anthropic files for IPO
• Investing.com — mid-November IPO timing
• Fortune — compute strategy in the S-1
• Motley Fool — five key numbers from the S-1
• Unite.ai — S-1 structure and controlled-company status
• IBTimes — loss breakdown
• WION — revenue, operating loss and run-rate
• Traders Agency — counterparty dependencies and SpaceX concentration
• Dealroom — draft S-1 summary
• Crypto Briefing — 2025 figures
• Yahoo Finance — the $518 billion commitment
• BigGo — SpaceX compute terms

Governance

• TechCrunch via AI Weekly — founders' 50.1% voting class
• Reuters via The Star — supervoting shares
• WSJ via iTiger — the Long-Term Benefit Trust
• iTechGuides — ownership and Trust-appointed board majority

Margins and projections

• Futurum — projections shared with investors
• Morningstar — gross margin estimate

Comparable: SpaceX

• Nemo — SpaceX IPO performance
• AI Weekly / TechCrunch — SpaceX back to the $135 IPO price
• The Edge — SpaceX second-day trading
• MarketScreener — 2025 Nvidia and Microsoft commitments