Anthropic S-1IPO prospectus

Executive Summary

Anthropic sells Claude AI models, developer tools and applications to organizations and individual professionals, earning usage fees and subscriptions. It is developing Claude into an enterprise “Virtual Collaborator,” extending from coding into autonomous, multi-step knowledge work. H1 2026 revenue reached $16.3B, up 1,300% Y/Y, while its operating loss narrowed to $693M. The preliminary IPO combines company-issued shares and stockholder sales, but blank pricing and share counts prevent calculating the raise, valuation or buyers’ stake. Buyers would back extraordinary adoption against enormous infrastructure commitments, continued model-development demands and governance that gives public shareholders little control.

Business Model and Growth Drivers

Claude combines frontier models, a developer platform and first-party applications. Claude Code handles software development; Claude Cowork extends agentic work into broader professional tasks and began merging with Claude.ai in 9/2026. Platform tools provide memory, connectors, governance and coordination of multi-step workflows. Management reports Opus 5.5 reduced customer cost per task by ≈40% vs Opus 5, potentially making more workloads economical. Models launched in H1 2026 generated over 80% of Q3 revenue, demonstrating both rapid adoption and dependence on fresh releases.

• Revenue mix: H1 2026 consumption fees $13.1B, or 80%; subscriptions $3.22B, or 20%. Usage is metered in tokens, with prices varying by model. Enterprise minimum-spending contracts are increasing, but much revenue still depends on actual consumption.
• Enterprise expansion: More than 11,900 customers generated at least $0.1M of trailing-twelve-month revenue at 9/30/2026, vs ≈1,500 at FY2025 year-end; ≈250 exceeded $10M. The lower threshold is Anthropic’s enterprise-adoption proxy.
• Retention: ≈900% net revenue retention measures qualifying customers’ quarterly revenue against the corresponding quarter a year earlier, excluding Claude Pro and Max subscriptions. It reflects dramatic expansion within existing accounts, not simply customer retention.
• Distribution: Cloud marketplaces supplied 48% of H1 revenue. Amazon, Google and Microsoft are simultaneously infrastructure providers, distribution partners and competitors. The top five customers represented ≈25% of 9M 2026 revenue.

The Offering

• Size and valuation: Unpriced, with no price range or offering share count; gross proceeds, implied capitalization, revenue multiple, buyer ownership and per-share dilution cannot be calculated. A historical fully diluted base can be computed at 16.9B shares from the Public Benefit Fund’s stated 5% holding, assuming full Amazon-note conversion at 6/30/2026; this is not the post-IPO count.
• Structure: Primary and secondary Class A shares; secondary proceeds go to selling holders. Goldman Sachs and Morgan Stanley lead underwriting, with a 30-day additional-share option of unspecified size. Proposed listing: Nasdaq Global Select Market, ANT.
• Company proceeds: Compute infrastructure, operations, hiring and potential investments, plus tax withholding on restricted stock units settled around the IPO. This means some cash funds employee tax obligations rather than expansion.
• Capitalization bridge: Before-IPO preferred stock and convertible notes convert into common shares; public-benefit issuances and restricted stock unit settlement also change equity. “After IPO” additionally includes offering proceeds, expenses and settlement-related taxes; quantitative adjusted balances remain blank.
• Resale supply: Restrictions generally run up to 180 days, or 365 days for founders and specified holders, with earlier staged releases tied to trading days and earnings announcements. Separately, AMD conditionally committed to purchase up to $5.0B of Class N stock alongside a compute arrangement.

Key Financial Highlights

Historical figures; interim periods are unaudited. Dollar amounts in billions.

• Deficit explanation: $158B of redeemable preferred stock sits outside stockholders’ equity; accumulated losses total $133B. The note liability includes valuation increases, whereas Amazon-note principal outstanding was $6.20B—the deficit is not equivalent to borrowed cash.
• Accounting distortion: H1 2026 included $72.4B of noncash financing revaluation losses, largely because rising company valuation increased convertible claims’ measured value, plus $8.76B of other expense primarily related to note agreements.
• Operating improvement: Compute expense fell to 55% of H1 revenue from 201%. Q2 achieved $974M GAAP operating income, but near-term use of higher-cost compute could pressure margins.
• Compensation: H1 stock-based compensation was $2.19B. Offering-linked vesting could trigger substantial additional expense; cumulative restricted-stock-unit expense would have been $2.9B had performance conditions been probable at 6/30/2026.

Governance and Control

• Control: The Long-Term Benefit Trust elects four of seven authorized directors. Founder LLC holds fixed 50.1% general voting power; Class A carries one vote per share, while Class N is generally nonvoting.
• Public-benefit mandate: Directors balance investor returns with affected stakeholders and responsible AI development. Anthropic uses the controlled-company exemption permitting a nominating committee that is not entirely independent.
• Dilution: The Public Benefit Fund received 847M shares, or 5% of the stated historical fully diluted base. Further issuances may reach 2% of outstanding shares at each specified 50% share-price appreciation milestone.
• Employee equity: Outstanding options and restricted stock units totaled 2.85B at 6/30/2026, or 17% of that historical fully diluted base; recurring equity-plan reserve increases permit further dilution.

Risk Assessment

• Audit transition: Ernst & Young replaced Grant Thornton for FY2025; prior opinions were unqualified, with no reported disagreements or reportable events.
• Fixed commitments: ≈$321B of uncommenced infrastructure lease obligations dwarf current liquidity; variable customer usage could leave expensive capacity underused.
• Model leadership: Recent launches supplied most Q3 revenue, making research setbacks or cheaper competing models commercially consequential.
• Agent safety: Actual unauthorized-access incidents and cyber misuse show how autonomous capabilities can create liability or force restrictions.
• Regulatory and copyright exposure: Export restrictions temporarily disabled models; a government supply-chain designation threatens government-related business, while copyright claims continue after the $1.5B Bartz settlement.

What to Watch

• Enterprise expansion: Track the 11,900 large-spending customers and retention; continued expansion supports durable adoption, while contraction would expose usage sensitivity.
• Product economics: Watch whether Claude Code and Cowork adoption sustains revenue growth without reversing compute-cost efficiency.
• Capacity utilization: New infrastructure must support paying workloads; delays or idle capacity would turn commitments into financing pressure.
• Profit quality: Follow whether Q2 operating profitability persists after offering-linked compensation and infrastructure costs arrive.