Anthropic IPO roadshow: about $1T of equity value against about $900B of compute commitments

TL;DR

The roadshow put two numbers side by side. At the midpoint of the price range, Anthropic's implied equity value is about $968.4B; its disclosed compute commitments are about $900B. The first is 35% to 52% below the valuations that circulated in October; the second is $150B above the widest figure compiled then.

• Valuation range $916.1B to $1,020.8B. Price range $70–78, about 13.087B shares after the offering. The midpoint is 35.4% below the $1.5T the October piece used as a hypothesis and 51.6% below $2T.
• The multiple more than doubles depending on which revenue you divide by. The midpoint is 12.0x the October run-rate of $81.0B, 14.2x Q3 annualized revenue of $68.0B, and 27.4x trailing-twelve-month revenue of $35.28B.
• Net loss $104.04B, operating loss only $1.26B. About $94.3B of the gap is fair-value remeasurement of convertible notes, derivative liabilities and SAFEs, not operating cash burn; operating cash flow for the first nine months was +$6.6B.
• Compute efficiency slipped in Q3. Revenue over compute expense fell from 2.08 in Q2 to 1.72. Using the roughly $11.6B Q2 revenue cited in the October piece, compute expense rose 77.2% quarter over quarter against revenue growth of 46.6%.
• The next twelve months of commitments come to $70.7B, about 1.79 times Q3 annualized compute expense. Holding the Q3 revenue-to-compute ratio on that spend would take about $121.6B of revenue, 50% above the October run-rate.
• Three labels in the draft do not reconcile: growth multiples mix two conventions; compute commitments appear as both $900B and $900.9B; and the roughly $5B of short-term compute cannot all have landed in Q3.

This piece is built on an external draft. The draft says it relies on the second amended S-1 of November 10, the IPO fact sheet, the founders letter and five customer case studies. The snapshot is 2026-11-11; the final price is expected on November 18 with trading on the 19th, and the registration statement was not yet effective as of November 10. Figures are kept as the draft states them and the SEC filing was not reopened for this piece. Everything computable was rechecked; the comparisons and derivations in sections 2, 5, 9 and 10 are mine.

1. Offering terms: four different numbers answer "how much"

Listing on Nasdaq under ANT. Timeline: first S-1 on October 26, first amendment on November 2, second amendment and roadshow launch on November 10.

Transaction value, proceeds and market value are three different things. $57.75B to $64.35B is what the shares sold in this deal are worth; what reaches the company is the new shares plus the placement, net of fees, about $84.6B; market value multiplies the full share count, covered in the next section.

2. What changed since the October piece

The October 26 piece was written the day the first S-1 was filed. There was no price range then, so valuation could only be hypothesized from press figures. The amendment two weeks later pinned several of those numbers down:

The interesting part is two changes moving in opposite directions: the price the market will pay is well below October's talk, while the compute spend already signed is above October's widest figure. Midpoint value $968.4B, commitments $900B; the latter is 92.9% of the former.

That ratio has obvious basis problems — commitments span many years, carry conditions and termination rights, and are nominal rather than present value, while valuation is a one-time price on equity — so it does not mean the company has pledged its whole value to compute. But it shows the scale: a large share of this company's cash flow for the next decade is already allocated to compute suppliers.

Section 8's P/S table in the October piece is therefore superseded. That table had 22.3x Q3 annualized revenue at $1.5T; on the same basis (Q3 annualized $68.0B) the midpoint is now 14.2x. The qualitative analysis in its sections 3 and 7 still holds.

3. Revenue: recognized, run-rate and trailing twelve months are three numbers

From about $0.4B in 2025 Q1 to $17.0B in 2026 Q3, quarterly revenue grew to 42.5 times its starting level.

The three measures have to be kept apart:

• Trailing twelve months, realized: 2025 Q4 $2.04B plus 2026 first nine months $33.242B, about $35.28B. This already happened.
• Q3 annualized: $17.0B × 4 = $68.0B. Assumes the Q3 level holds for a year.
• October run-rate: about $81.0B, 19.1% above Q3 annualized. Assumes the latest month holds for a year.

The last two are extrapolations. In a period of fast growth they sit closer to the current scale of the business than the trailing figure, but they lean harder on growth not slowing.

4. Profit and cash flow: the two loss figures describe different things

First nine months of 2026:

About $94.3B of the net loss comes from fair-value changes on convertible notes, derivative liabilities and SAFEs, 90.6% of the total. The higher the company's valuation goes, the larger the book loss from remeasuring these instruments — this part is the accounting shadow of a rising valuation, not cash consumed by operations.

The adjusted number cannot be read alone either. Q3 GAAP operating margin was about -3%, adjusted +15%; the 18 percentage points of adjustments, stock compensation among them, are real costs that simply do not use cash. The prospectus also warns that as infrastructure spending such as Anthropic Direct expands, future operating and investing cash flow may come under pressure — positive operating cash flow today does not show that free cash flow has settled.

5. Compute: the efficiency ratio, short-term capacity and $900B of commitments

Revenue over compute expense was 0.40 in 2025 Q1, reached 2.08 in 2026 Q2 and fell back to 1.72 in Q3. The company credits the rise to algorithms, scheduling, a multi-chip fleet (AWS Trainium, Google TPU, NVIDIA GPU, AMD GPU) and infrastructure work, and attributes part of the Q3 decline to buying higher-priced short-term capacity.

This ratio is neither GPU utilization nor gross margin — training and inference compute share one denominator.

Using the roughly $11.6B Q2 revenue cited in the October piece, quarterly compute expense can be split out:

Q3 compute expense rose 77.2% quarter over quarter, revenue 46.6%. Q1 is backed out of the nine-month total of $18.808B by subtracting Q2 and Q3.

The commitments: about $900B in total, of which about $451.6B is expected by the end of 2030 and about $70.7B within the twelve months after September 30, 2026. Part of it is conditional or terminable.

Placing the $70.7B against the Q3 cost structure:

Q3 annualized compute          $9.884B × 4    ≈  $39.54B
Next-12-month commitments      $70.7B         ≈  1.79x Q3 annualized
Hold the Q3 revenue/compute    $70.7B × 1.72  ≈  $121.6B of revenue
October run-rate               $81.0B          →  50% more needed

This is a sensitivity, not a forecast. It assumes the whole $70.7B is expensed over the next twelve months and the ratio stays at the Q3 1.72. If efficiency returns to the Q2 2.08, the same spend supports more revenue and the pressure eases a lot; if some commitments are capitalized facilities, the expense spreads over a longer period. But it frames the question: over the next year, revenue has to grow half again beyond the October run-rate for the new compute not to compress margins.

6. The strategic placement: suppliers are also shareholders

Alongside the IPO the company plans to sell $26B of Class N stock to Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC Partners. Class N is generally non-voting.

The seven sit across chips, memory, foundry and cloud — where the main recipients of the $900B in commitments are. $26B is 2.9% of the commitment total. Suppliers taking a small equity stake in exchange for certainty on a large long-term order book is not unusual, but it raises three things to watch: whether procurement is priced at arm's length, supplier concentration, and money flowing from shareholders to the company and back to the same shareholders.

7. Customers and case studies: the edge moves from the model to the workflow

As of September 30, 2026:

A 900% NRR cannot be read as a SaaS renewal rate. It compares eligible customers' latest three months with the same three months a year earlier, and excludes fixed subscriptions such as Claude Pro, Max and Ultra. It says enterprise usage is expanding very fast; it does not say every existing customer will keep that pace.

Five case studies, with their numbers and the limits on reading them:

The common direction is that AI is moving from answering questions to completing tasks, and embedding into enterprises' existing permissions, audit and compliance systems. Permissions, audit and human review — things outside the model — are becoming part of the competitive edge. Anthropic sells model, platform and applications at once, which is the advantage; the risk is that large customers build their own model routing across those layers and turn it into one replaceable supplier among several.

8. Governance and the founders letter: which decisions may depart from profit maximization

The founders letter makes two commitments that may cost near-term revenue: no advertising in Claude, and possibly delaying releases when facing high-risk capabilities, citing the restricted release of Claude Mythos Preview. The founders have each pledged 80% of their personal Anthropic equity to charity, and the company is setting up a public-benefit fund equal to 5% of its equity.

The letter sorts AI risk into four kinds: misuse, loss of control, authoritarian abuse, and economic disruption.

For a Class A buyer both of these hold at once: the special governance may keep long-term research and safety decisions from being held hostage to quarterly results, and public shareholders' influence on the board and strategy is genuinely limited. The governance structure itself belongs in the valuation — as a control discount, not just a statement of values.

9. Valuation: one price, 12x or 27x

Excludes the over-allotment, future dilution, and the gap between equity value and enterprise value.

At the same price, realized revenue gives 27.4x and the latest month extrapolated gives 12.0x. The more-than-double gap is what the market is paying for growth holding up.

Three scenarios:

1. Bull: enterprise agent usage keeps climbing, model prices fall more slowly than unit costs, and operating margin widens.
2. Middle: revenue keeps growing but the pace normalizes, and the compute investment wave makes margins swing.
3. Bear: price competition cuts unit prices fast, long-term compute commitments go underused, and cash flow and valuation come under pressure together.

Current disclosures are not enough to assign probabilities to these.

10. Three places the draft does not reconcile

1. Growth multiples mix two conventions. The draft says 2024 "grew about 8.4 times" and 2025 "grew about 11.9 times". By the revenue figures, 2024 is 9.4x 2023 (growth of 8.4x) and 2025 is 11.9x 2024 (growth of 10.9x) — the first is stated as growth, the second as a multiple. The October piece says 2025 was "about 11.9 times 2024", which is right. Section 3 here uses "N times the prior year" throughout.

2. Compute commitments are given twice. Section 4.1 of the draft says "about $900B"; its final section says "$900.9B". Both may come from different places in the filing, but they cannot both be exact; this piece uses about $900B. A smaller one of the same kind: the midpoint valuation is $968.5B in the draft's section 4 and $968.4B in its section 9; 13.087B × $74 = $968.44B, so the latter.

3. The roughly $5B of short-term compute cannot all sit in Q3. If it did, Q3 compute without it would be $9.884B − $5B = $4.884B, 12.4% below Q2's roughly $5.577B — while revenue rose 46.6% in the same quarter. That does not hold together. More likely the $5B is a multi-period contract amount or includes prepayment, with only part of it in Q3 expense. The draft calls it "related expense", which invites reading it as a single-quarter cost.

11. What this piece cannot compute

• The SEC filing was not reopened. All financial figures are as the draft states them; this round only rechecked arithmetic and consistency across documents.
• The final price is not set. Pricing is on November 18 and the share count may move with the over-allotment; every valuation here is a static calculation.
• No present value or cancellable share for the commitments. $900B is a nominal total; the draft does not say how much is conditional or terminable.
• No breakdown of the adjustments. The draft does not split the 18 points between -3% and +15%.
• The quarterly compute split rests on the October piece for Q2 revenue. About $11.6B comes from a press report of investor communication, not from this amendment; the Q1 1.39 is derived from it.
• The roadshow video and full slides were not reviewed. The draft says that part was rebuilt from the official agenda and fact sheet.
• The onchain pre-IPO products named after Anthropic are not ANT stock. They carry no shareholder rights; see this piece. The price range only gives them a reference anchor.

Conclusion. Anthropic has shown with realized revenue that frontier AI is a real and fast-growing market: $33.242B of revenue in nine months, positive operating cash flow, an $81.0B annualized run-rate. The roadshow price is also far more restrained than October's talk — 12.0x run-rate at the midpoint.

But commercial success and return on capital are separate questions. About $900B of compute commitments is already close to the company's entire equity value; the next year's slice requires revenue half again above the October run-rate to avoid compressing margins; and the counterparties on those commitments are, through Class N, becoming shareholders too. What sets long-term value is not how capable the models are, but whether cash flow can steadily outrun the compute bill already signed. Even if AI creates enormous social value, a foundation-model company need not capture most of the economic surplus.

Sources: the primary materials listed in the draft are the Anthropic IPO site, IPO fact sheet, founders letter, prospectus, second amended S-1 on EDGAR, roadshow announcement, retail roadshow, and the Legora, Lilly, Rakuten, Salesforce and Walmart case studies. Snapshot 2026-11-11, before pricing; these files were not reopened for this piece. Research notes only, not investment advice. DYOR.