ATLAS

How to read this filing in an hour

Three stops, marked on the facing plate in the order to take them. They hold nearly everything that is unusual about this prospectus.

Hover any province, peak, sounding or settlement in this atlas for its value and where it sits in the filing.

Six quarters from $376M to $11.6B

Revenue compounded faster than compute cost, and the operating line crossed zero in Q2 2026. Below the operating line, mark‑to‑market losses on convertible notes dominate net loss and vanish at conversion.

In the plates of this chapter, zero is sea level. Revenue rises as land; losses are measured as depths below the surface.

Annual statement of operations

The gap between GAAP net loss and adjusted net income is almost entirely the change in fair value of convertible notes, derivative liabilities and SAFEs ($59.6B in Q2 2026 alone), driven by the rise in enterprise valuation.

Balance sheet, 30 June 2026 (actual)

Liabilities include the Amazon convertible notes carried at fair value; the preferred stock converts to Class A and Class N common at the offering, which is why the pro forma columns matter and are still empty.

Where the revenue comes from

Pay‑for‑intelligence (metered tokens) is four‑fifths of revenue; half of it is collected through cloud marketplaces; the ≥$100k customer base grew sixty‑fold in under three years.

• Customers with ≥ $100k11,900+
• Customers with ≥ $1M1,900+
• Customers with ≥ $10M~250

About 40 million further customers each below $100k. Over 50% of the Fortune 500 and over 30% of the Global 2000 are ≥$100k customers; revenue from Fortune 500 customers averages under 0.1% of their expenses (top 10: under 1%).

Concentration

No customer exceeded 10% for the nine months to 30 Sep 2026. Revenue outside the top five rose from ~64% (Q1 2026) to ~79% (Q3 2026). An estimated 5–7% of nine‑month revenue came from customers later found in violation of usage policy or commercial terms.

Usage dynamics the filing calls out

• Model versions released in the six months to 30 Sep 2026 produced over 80% of Q3 2026 revenue.
• QoQ growth accelerated from 48% (Q4 2025) to 132% (Q1 2026) after Opus 4.5 cut price and cost per task.
• Opus 5.5 lowered cost per task ~40% vs Opus 5; Fable 5.1 ~25% vs Fable 5 (up to 45% on agentic coding).
• Over 50% of ≥$100k customers use both the Claude Platform and a first‑party app (up from ~30% in Q4 2025).
• Top 5% of Claude Enterprise users on Claude Code spend ~20× those who don’t use it (under $75/month).
• Seasonality: customer spend fell ~20% in the two weeks to 2 Jan 2026 vs the prior two weeks.

What is sold, on what, through whom

Four model classes on four chip families, delivered through a platform, first‑party agents and three clouds. Compute is managed as one fungible pool and is shifting toward leased hardware in dedicated facilities (“Anthropic Direct”).

Master equipment leases for accelerators (the “2026 Lease Arrangement”), colocation with Fluidstack, direct data‑center leases in the US plus Australia and Canada, and power procured separately (grid or behind‑the‑meter). First owned‑footprint facility operational in 2026. Long‑term compute is typically contracted 12–18 months ahead; short‑term top‑ups cost more and are expected to rise.

• Responsible Scaling Policy: safety evaluations tied to capability thresholds.
• Claude’s Constitution, orienting the model around four priorities.
• Alignment, interpretability and safeguards research; model validation and training‑data controls.
• Trusted‑access programs (Cyber Verification Program, Project Glasswing) gate Mythos‑configuration models.
• Competition named: OpenAI, Google, large technology companies, open‑weight models incl. from state actors; and the “status quo” of entrenched enterprise systems.

Five classes of common stock, two of which run the company

A public benefit corporation whose board majority is chosen by a trust with no economic stake, and whose founders hold 50.1% of the general vote through a single share. Public Class A holders get one vote per share on what is left.

Capital structure after the offering

Sunset mechanics

After the Class T sunset, if Class F remains, Class A and Class F vote together for all directors, so the Founder LLC decides every general matter. After the Class F sunset, if Class T remains, the LTBT keeps electing the board majority and Class A elects the rest. Only once both are gone does Class A alone decide.

Public benefit mechanics

• 846,840,960 Class N shares (5% fully diluted at 30 Jun 2026, assuming Amazon notes convert) contributed to Anthropic Public Benefit Fund, LLC, a wholly owned subsidiary.
• Board may issue up to a further 2% to the Fund each time the share price rises 50% above the prior milestone, starting from the IPO price.
• Each co‑founder pledged 80% of personal equity to philanthropy; employee donation‑matching paid in Class N shares.
• Incorporated as a Delaware PBC in January 2021; LTBT created 2023; 10‑for‑1 forward split of Class A/N and 1‑for‑1,000 reverse split of Class T on 17 Sep 2026.

$141B raised, $77B of undrawn note facilities, and commitments still blank

The strategic partners are also the largest suppliers and distributors. Amazon’s notes convert at the IPO; Google, Amazon and Broadcom have agreed further note facilities. The compute commitment totals are placeholders in this draft.

• Aggregate capital raised as of 30 Jun 2026$141.1B
• Cash and cash equivalents, 30 Jun 2026, plus $19.3B short‑term investments$112.4B
• Amazon convertible notes outstanding ($M), converting into Class N$6,203
• Undrawn note purchase facilities ($M): Google, Amazon, Broadcom$77,000

Note purchase agreements · as of 30 Jun 2026

Amazon convertible notes

Converting at IPO, subject to Amazon’s ownership cap. Conversion is at the lower of the fixed price and the IPO price. Notes blocked by the ownership cap roll over and convert quarterly at the lowest of fixed price, IPO price and 10‑day VWAP. These notes are carried at fair value, which is the source of the large non‑operating losses on Plate 6.

Compute commitments

Purchase commitments and leases (including leases not yet commenced) entered as of and after 30 Sep 2026 have an aggregate committed spend of blank, substantially all compute and infrastructure: blank before 31 Dec 2030, the remainder in 2031–2045, and blank due within 12 months. The payments‑due‑by‑year table (operating leases, finance leases, purchase commitments) is entirely empty in this draft.

The filing says a significant portion relates to Anthropic Direct and expects compute to be “a significant use of net cash in operating and investing activities for the foreseeable future.”

Litigation and settlements

Bartz et al. v. Anthropic PBC: copyright class action by authors (filed Aug 2024) over books used in training. Settled for $1.5B in September 2025, approved by the district court in July 2026, expensed in Q3 2025. The filing notes a series of further copyright suits in US and international courts and that insurance may not cover such claims.

Use of proceeds

General corporate purposes incl. compute and infrastructure capex, hiring, possible acquisitions, and tax withholding on RSU settlement. No dividends intended.

94 pages, six territories

Risk factors are 29% of the prospectus by words. Business and industry risks take two‑thirds of that. Choose a territory on the map or below for the headline risks the company itself lists in its summary.

Seven co‑founders, three officers, four outside directors

Class A holdings before the offering, as disclosed; percentages, Class F and post‑offering columns are blank in the draft. Share counts reflect the September 2026 10‑for‑1 split.

Executive officers and directors

Other co‑founders (Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, Christopher Olah) appear as selling stockholders and Founder LLC members. Jay Kreps is listed as a former director ($200,000 fees).

Headcount and talent, as described

Approximately 5,300 employees at 30 Sep 2026, a majority in research, engineering and product. The team is credited in the filing with RLHF, Constitutional AI, GPT‑2 and GPT‑3, mechanistic interpretability, the first Responsible Scaling Policy and the Scaling Laws paper. Headquarters in San Francisco; offices in Europe and Asia‑Pacific. Some non‑US employees fall under industry‑wide collective bargaining agreements by law.