ANT Offers Owners $1.26B of Half-Year Cash and 49.9% of the Vote

Revenue

H1 2026

$16.29B

Free cash flow margin

H1 2026

21.17% of revenue

Owner-cash margin

H1 2026

7.75% of revenue

Class A vote cap

Post-offering

49.90% of the vote

Summary and thesis

Anthropic filed its S-1 on Oct 26, 2026 to list Class A common stock on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley as representatives of 22 underwriters ^[1][7].

The price range, the number of shares offered, the net proceeds and the share count after the offering are all left open in the S-1 and are set at pricing. This is therefore a preliminary piece of research: it tells you what the business earns and what you would not control, and it leaves valuation to the day terms land.

The business case is strong on the figures that are filed. Revenue was $41M in 2023, $386M in 2024 and $4.59B in 2025, then $16.29B in the first half of 2026 alone, up 1,300.34% from $1.16B a year earlier ^[1].

Quarterly revenue rose from $376M in the first quarter of 2025 to $11.56B in the second quarter of 2026 ^[4].

Profit followed. Operating cash flow turned from -$2.64B in 2025 to $3.93B in the half ^[8].

Adjusted operating income was $1.62B, a 9.93% margin, and the GAAP operating loss shrank to $693M, a -4.26% margin against -183.58% a year earlier ^[1][4].

The control case is the opposite. Class A holders, which is what the public will buy, are capped at 49.90% of the vote, and a single Class F share held by an LLC of seven co-founders carries a fixed 50.10% ^[1]

The owner-cash staircase

The staircase starts from filed cash flow, so it needs no price. Operating cash flow was $3,929M for the six months to Jun 30, 2026. Subtract $482M of capital expenditure and free cash flow is $3,447M, a 21.17% margin on $16,286M of revenue (derived).

Stock-based compensation of $2,185M is a real cost to owners even though it uses no cash. Subtract it and owner cash is $1,262M, a 7.75% margin (derived) ^[4][8].

Two cautions keep that figure honest. First, $1,111M of interest income earned on the cash pile sits inside the half's results ^[8].

Strip it out, before tax, and owner cash is roughly $0.15B (derived and approximate). The operating business is cash positive, but the margin of safety per share is thin until the share count is known.

Second, this is six months, not trailing twelve months, and the quarterly path was steep: the second quarter had an 8.43% GAAP operating margin and a 19.34% adjusted margin, while the first quarter was -35.24% on a GAAP basis (derived from ^[4]).

Compute and infrastructure is the cost that matters. It was 159.78% of revenue in 2025 and 54.92% in the first half of 2026

Products and the token-consumption model

Anthropic sells access to its Claude models in four classes: Mythos, which includes Fable, plus Opus, Sonnet and Haiku ^[5]. They reach customers through the Claude Platform, a set of developer tools for building agents, and through first-party applications including Claude Code, Claude Cowork, Claude.ai, Claude Security, Claude Design and Claude Science ^[5].

Distribution runs direct and through AWS, Google Cloud and Azure marketplaces, which were 48% of first-half 2026 revenue, up from 11% in 2023 ^[2][8].

The revenue model is consumption. Of $16,286M of first-half revenue, $13,064M, or 80.22%, was consumption-based, meaning customers are charged by the token.

Subscription revenue was $3,222M, 19.78%, up 1,690.00% from $180M a year earlier ^[8]. Subscriptions run 12 months or less ^[4].

The consumption share fell 4.30 points from the first half of 2025 (derived), but the metered share is still the engine ^[8].

Metering by the token links revenue to cost per task. The company says Claude Opus 5.5 cut customer cost per task by about 40% against Claude Opus 5, and that Fable 5.1 cut typical workload cost by about 25% against Fable 5, up to 45%

Customers and concentration

Customer concentration is moderate and channel concentration is high. The table sets out what the S-1 says about each.

Remaining performance obligations of $10.8B are 66.31% of one half-year of revenue (derived), small next to the run rate, which fits a usage-driven model.

The concentration that matters is the cloud providers. Marketplaces were 48% of first-half revenue and 58% of net receivables ^[2][8].

The same companies are Anthropic's suppliers and investors: Amazon holds the convertible notes, Google is a preferred investor with an option on more, and Microsoft, Google and Amazon all sell compute to Anthropic under multi-year agreements ^[2][4]. The S-1 flags the conflicts of interest this creates as a risk factor.

The largest single compute arrangement in dollar terms after Broadcom's leases is not with a cloud provider. Between May and October 2026 Anthropic signed and amended several agreements with SpaceXAI, formerly xAI and a wholly owned subsidiary of Space Exploration Technologies Corp., for compute powered by NVIDIA GPUs.

It agreed to spend up to $120.5B over terms extending into 2029, about $3.1B a month once all capacity is delivered, which it expects in December 2026. After initial non-cancelable periods ending in January 2027 and March 2027, either party can cancel each agreement on 90 days' notice

The compute commitment stack

The S-1 gives compute commitments in several measures, and they do not add to one total. The MD&A contractual-obligations table leaves its aggregate and the range for agreements without a fixed dollar amount open for pricing ^[4].

The financial statement notes do give figures.

The fixed or minimum figure rose 4.87x from $54.56B at Dec 31, 2025 to $265.60B at Jun 30, 2026 (derived, different bases), and agreements signed since add about $245B more. Some of the $245B is subject to contractual conditions being met ^[8].

The filing also says payments under agreements with no fixed dollar amount are not yet determinable but are expected to be significant ^[8].

The company describes its largest arrangements individually, and those dollar figures are filled in. We laid them out below and summed the named dollar amounts ourselves.

The sum, about $637.00B, mixes stated minimums, up-to ceilings and finance lease rent over different terms, so it is a scale marker and not an S-1 total.

It excludes the AMD agreement, which the filing says should exceed $20.0B if the initial capacity is bought but has no fixed dollar commitment, and an estimated $125.0B to $150.0B of additional TPU compute for 2028 to 2031 that is not in the contractual table

Where the $80.75B net loss comes from

The net loss for the half was $80,748M. It starts with the $693M operating loss and adds $1,111M of interest income.

Then a $72,388M charge for the change in fair value of convertible notes, derivatives and SAFEs and $8,757M of other expense, net, followed by $21M of tax, reach the total ^[1][8].

The company says the fair value change was primarily driven by an increase in the fair value of its outstanding convertible promissory notes, and all of those notes are held by Amazon ^[1][2]. No cash left the company.

The loss is the rising value of Amazon's right to convert as Anthropic's enterprise value rose.

The balance sheet shows the same instrument. Convertible notes payable at fair value were $14,116M at Dec 31, 2024, $41,677M at Dec 31, 2025 and $108,566M at Jun 30, 2026, against $6,203M of principal outstanding, a ratio of 17.50x ^[4][8].

Total liabilities were $117,467M, redeemable convertible preferred stock was $157,510M and the stockholders' deficit was $(127,619)M ^[8]. The deficit follows from carrying the notes at fair value and holding the preferred stock outside permanent equity.

Debt note: never zero

Debt for any future enterprise value comes from the debt note, never from an open line. The outstanding debt is the Amazon convertible notes, which carry no interest ^[4][8].

The notes convert at the offering, subject to Amazon's 12% ownership cap, which can be waived, and any unconverted remainder stays as rollover notes ^[4]. The pro forma remaining balance is set at pricing, so we mark it unverified and do not read it as zero.

The method we will use when terms land: enterprise value is the share count after conversion times the offering price, plus the remaining rollover principal, minus cash, with leases excluded. We will not combine fair value carrying amount with post-conversion shares, which would count the same claim twice.

At $108,566M the carrying amount is 17.50x principal, and the choice between the two is the difference between a rounding error and a distorted multiple.

There is also contingent borrowing, and the S-1 asks readers to use remaining capacity, not the original amounts ^[4][8].

Governance, the Long-Term Benefit Trust and share classes

The Founder LLC has seven members: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah. Benjamin Mann is often called a co-founder but is not a member ^[6].

Class F ends if fewer than three continuing founders remain for 90 consecutive days, or if the LLC is redeemed or dissolved ^[6].

After the offering the board has seven authorized seats. Class T elects four, including Reed Hastings, Vas Narasimhan and Christopher Liddell as the Specified Class T Director, with one vacancy.

Classes A and F elect three: Dario Amodei, Daniela Amodei and Yasmin Razavi ^[6].

The Long-Term Benefit Trust has five seats, one of them vacant. Its trustees are Neil Shah, chair since June 2023, Richard Fontaine, Ben Bernanke and Christopher Stubbs ^[6].

The company paid the trust $2,716,750 in 2023, nothing in 2024, $906,089 in 2025 and $8,425,213 in the first nine months of 2026 ^[6].

Offering terms, use of proceeds and the last private marks

The company offers Class A common stock and selling stockholders include five co-founders: Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah, with share counts set at pricing ^[7]. Google holds an option to buy up to $10.0B of Class N in a concurrent private placement at the offering price ^[4].

Proceeds are for general corporate purposes, including compute and infrastructure capital expenditure, hiring and retention, possible acquisitions with no commitments today, and tax withholding on restricted stock unit settlements. No debt repayment is stated ^[3].

The company does not expect to pay cash dividends, and the credit agreement places limits on cash dividends ^[2][3].

Lock-up agreements run with Goldman Sachs and Morgan Stanley as representatives ^[2]; the lock-up periods were not located in the pages we reviewed.

The 22 underwriters are Goldman Sachs, Morgan Stanley, J.P. Morgan, Citigroup, Barclays, Wells Fargo, BofA Securities, Deutsche Bank, RBC, UBS, BMO, BNP Paribas, Credit Agricole, Jefferies, Mizuho, MUFG, Societe Generale, TD Securities, BTIG, Cantor, Capital One and Macquarie ^[7].

The latest private price is Series H, sold in the second quarter of 2026 at $589.01 per preferred share (the S-1 states it to four decimals), for a gross of about $65.2B across H-1 and H-2

Risk factors

• Frontier leadership and compute access: staying at the frontier depends on securing scarce compute inputs, including constraints on critical compute inputs ^[1].
• Key people: dependence on the co-founders and key technical staff ^[1].
• Competition in a fast-moving market for frontier models ^[1].
• Misuse, misalignment and liability from model behaviour, including the usage-policy revenue estimate of 5% to 7% ^[1][4].
• Channel concentration: marketplaces at 48% of first-half revenue and 58% of net receivables ^[2][8].
• Short operating history and growth that cannot repeat: 1,300.34% growth in a half is not a base rate ^[1].
• Regulation and politics, including export controls on advanced chips ^[2].
• Cloud providers that are also investors and customers, with conflicts of interest ^[1].
• Governance: a novel multiple-class structure, voting power not tied to economic interest, and public benefit duties that may conflict with financial returns ^[1].
• Take-or-pay compute: minimum purchase and take-or-pay terms, price escalators and uncapped fuel cost pass-throughs apply regardless of utilization

Peers

There is no listed pure-play superintelligence lab, so the comparison set is the companies Anthropic buys compute from and the enterprise software platforms that sell to the same buyers. Peer multiples come from our validated records of each company's latest 10-K or 10-Q, and enterprise values move with the market, so treat them as a snapshot ^{[11][12][13][14][15]}.

Amazon, which is both investor and supplier, is not in the table because we do not hold a validated record for it. Anthropic's column is six months; the peers are trailing twelve months.

Stock-based pay as a share of operating cash flow is the one ratio that needs no price.

The table shows two things. Anthropic's stock pay absorbs 55.61% of its operating cash flow, more than any peer here, the nearest being ServiceNow at 41.16%, so owner cash is much thinner than free cash flow.

And Anthropic's $10.80B of remaining obligations is small next to the multi-year backlogs of Microsoft, Alphabet and Oracle, which is what a metered model with 12-month subscriptions looks like. Oracle's owner cash is negative once capex is subtracted, which is the same compute-build pressure Anthropic's commitments represent from the other side.

Outside read one: Levie on Fortune 500 spend

Aaron Levie (@levie) reacted to the S-1 on Oct 26, 2026 by quoting a post from Ben Bajarin (@BenBajarin) that summarized the filing's Fortune 500 data ^[16][17]. These are Levie's views, not S-1 facts.

We quote them, then set each against the filing.

His headline: "This is the key stat. The average Fortune 500 customer of Anthropic spends less than .1% of their expenses on Anthropic."

The S-1 figure behind it is on page 142: for the three months to Sep 30, 2026, revenue from Fortune 500 customers averaged less than 0.1% of their total expenses, and the top 10 Fortune 500 customers averaged under 1%, over eight times the cohort average ^[4].

Our read has three parts. First, the filed stat supports "early": a broad base of $100K-plus customers spending a sliver of their budgets, with the top cohort only about eight times deeper

Outside read two: Tunguz on seven metrics

Tomasz Tunguz (@ttunguz) published "Anthropic S-1 Analysis: How 7 Key Metrics Stack Up" on X on Oct 26, 2026 ^[18]. The table sets his reading of each metric against the filing.

His derived numbers are his arithmetic, and we mark where we reproduced them.

Where we add to his read, the first point is gross margin. Note 8 supplies cost of revenue, so the derived margin is reproducible, but it is a supplemental functional presentation; the S-1 itself presents expenses by nature ^[8].

What proves the thesis wrong

• Compute and infrastructure expense rising back above 60% of revenue as growth slows, from 54.92% in the first half.
• The share of revenue from models under six months old holding above 80% only because new launches keep replacing old ones, while price per task keeps falling.
• Net Revenue Retention drifting toward 100% as the customer base broadens.
• The Class F and Long-Term Benefit Trust provisions being waived or loosened before pricing, which would change what Class A holders own.
• Owner cash turning negative once interest income is removed, or contingent notes at 8% to 9% turning a cash-rich balance sheet into a leveraged one.
• Fixed or minimum compute commitments ($265.6B at Jun 30, 2026, plus about $245B signed since) growing faster than revenue, with SpaceXAI capacity cancelled on 90 days' notice and replaced at a higher price.

What we'll update when terms land

• Price range, shares offered by the company and by selling stockholders, and the share count after the offering. These unlock enterprise value, EV/EBITDA, the owner-cash multiple and a first test of whether the Series H price of $589.01 is an anchor or a ceiling.
• Net proceeds and the capitalization and dilution tables ^[3], including how many Amazon notes convert under the 12% cap and the remaining rollover principal, which sets the debt input.
• Whether EV/EBITDA stays not meaningful: GAAP EBITDA was -$596M in the first half, so we will show it as n.m. unless EBITDA turns positive in the next filing.
• Third-quarter 2026 revenue and Run-Rate Revenue as of Oct 31, 2026, both left open today, which will test the 1,300.34% growth and the owner-cash margin.
• The MD&A aggregate compute commitment, the contractual obligations table and the range for agreements without a fixed dollar amount ^[4], to reconcile with Note 13 ($265.6B) and the $245B of later agreements, and to replace our derived $637.00B marker.
• Voting percentages for the Long-Term Benefit Trust and the Founder LLC, lock-up terms, any directed share program, and the Google concurrent private placement at the offering price ^[4].
• The Series G-2 versus Series H label on F-40 and F-42, if the filing is amended.

Unverified items

Left open in the S-1 and not estimated:

• The price range, shares offered, net proceeds and selling stockholders' share counts.
• The Long-Term Benefit Trust and Founder LLC voting percentages.
• Third-quarter revenue and Run-Rate Revenue as of Oct 31, 2026.
• The MD&A aggregate compute commitment and contractual obligations table.
• The pro forma remaining Amazon rollover notes.

Also unverified: the Series G-2 versus Series H label, the Max 20x and Team plan prices, lock-up terms, and the two-thirds non-cancelable split in Tunguz's analysis. We use no press-reported figures for other companies from his post.

We omit any implied company value and any trailing revenue estimate derived from revenue per employee, because neither appears in the filing. Nothing here is a recommendation or a price target; position limits and sizing are not addressed, and no order has been placed.