Anthropic S-1: Explosive Growth, Compute Contracts, and the Price of an IPO

As of October 31, 2026. B = billion US dollars; T = trillion US dollars. Filed results, reporting, and this article’s forecasts are distinguished below.

• Growth: H1 revenue $16.3B, 14× a year earlier. Reuters reports Q3 revenue of $16.8B.
• Profitability: H1 operating loss $0.69B. Most of the $80.75B net loss comes from financial-instrument remeasurement.
• Product: Opus 5.5 ranked first at 58 when AAII v5 launched. Carbon overtook it at 59 the next day.
• Price: Reported IPO expectations of $1.5–2T put it alongside Meta and Broadcom, each currently around $1.6T.

Original: Anthropic S-1 · Further reading: EO’s S-1 analysis

1. Growth: Customer spending is expanding even faster

• H1 $1.163B → $16.286B: Roughly 14× year over year.
• Q3 $16.8B: About 45% above Q2. Reuters reporting, rather than a finalized filed result.
• Year-end $110B: An analyst’s annualized revenue forecast. It extends the expected year-end pace over 12 months; it is not actual full-year 2026 revenue.
• Customers spending at least $100K annually: 950 → 11,900. September 2025 → September 2026. Of these, 1,900 customers exceed $1M, showing a widening large-customer base.
• September 2026 NRR: 900%. The same existing customer cohort spent 9× as much in the quarter as a year earlier. Excludes individual Pro and Max plans. New customers and deeper usage both drive growth.

Sources: S-1 customer metrics and NRR definition, Reuters, October 14.

2. Finance: An $80.7B net loss can coexist with an operating profit

What each number measures

• Why an operating loss? Operating expenses of $16.979B exceeded revenue of $16.286B. Research and development, classified by function, was $9.741B, about 57% of operating expenses.
• Why adjusted profit? Add back $2.185B of stock compensation and $0.126B of noncash charitable costs to reach +$1.618B. Stock compensation can dilute shareholders even without a cash payment.
• Split the periods: Q1 operating loss of −$1.667B → Q2 operating profit of +$0.974B. Operations improved, but H1 as a whole remained loss-making.

Why the net loss is so large

• Remeasurement losses: $72.388B + $8.720B. Changes in the value of notes, SAFEs, and other financial instruments. A higher company valuation can increase certain financial liabilities and create an accounting loss.
• Not $80.7B of cash burned: Operating cash flow was +$3.929B. Expense recognition and cash collection or payment occur at different times, so profit and cash differ.
• What to watch: Alongside remeasurement, check whether operating profit keeps improving, whether profit survives stock compensation, and whether cash remains after capital expenditure.
• Cash and short-term investments: $131.651B. Resources at the end of June. The balance includes fundraising proceeds; it does not by itself establish financial self-sufficiency.

Source: S-1 income statement, cash flow, and financial-instrument notes.

3. Demand: Engineers up 5×; other functions grow faster

• Engineers: weekly active users up 5×. Engineering still accounts for the largest absolute usage.
• Other functions: legal 108× · sales 41× · recruiting 41× · marketing 26×. Rapid adoption from smaller starting bases.
• The distinction: These figures measure demand for AI usage within each function. They do not mean job openings or employment grew by those multiples.
• Connection to Anthropic: The report also describes rapid Claude Code adoption in management, sales, and legal. The market is expanding from developer tools to work across the organization.
• The condition for returns: Inference costs at fixed performance fell 47% per quarter on average (slide 109). 80.2% of Anthropic revenue is usage-based. Usage and efficiency must outgrow unit-price declines to protect revenue and profit.

Sources: Our earlier State of AI article, report slides 99 and 109, S-1 revenue mix.

Show three more original business charts from our earlier article

4. AAII: Compare the original charts before and after the update

• Bottom → top: v4.3.2 (October 23) → v5 (October 27). The test mix changed; this is not a model upgrade.
• The redesign: Two new tests and three upgraded tests. 60% of the index uses private or held-out data.

Artificial Analysis original screenshot · Click to enlargeOriginal source ↗

Artificial Analysis original screenshot · Click to enlargeOriginal source ↗

• Opus’s strength: It retained the lead under the new evaluation. Its gap to Kimi was 3 points before, 12 after. Different versions use different scales; score changes do not measure a change in model capability.
• Subsequent ranking update: After these charts, limited-access Gemini 4 Carbon scored 59 on October 28 and took first place. Opus at 58 became second. Neither chart includes Carbon.
• Work tasks: Opus 5 → 5.5 improved in automation 57 → 70, terminal tasks 49 → 60, and scientific tool use 29 → 59. These are separate test scores.
• Performance has a price: Average v5 task cost: Opus $10.04, Astra $6.24, Sol $1.27. Premium pricing needs justification on the hardest work.
• Evaluation settings: Opus at max with default safety fallback. Kimi’s weights were not yet released. Ties in this article use the displayed rounded scores.

Sources: pre-update original, v5 original, Carbon update, Opus generation comparison.

5. Compute: Securing capacity ahead of future demand

• EO’s roughly $637B total: Nominal amounts across multiyear contracts. Not one year’s expense or a debt due all at once.
• Minimum purchases: Google $114.5B (through 2033), Amazon $110B (through 2036), Microsoft $44.9B (through 2033). Some obligations remain even if usage falls short.
• Different terms: Broadcom’s $161.2B is finance-lease rent excluding interest. SpaceXAI’s up to $120.5B has delivery and termination conditions. Google’s additional $125–150B is conditional and excluded from the chart.
• Interpretation: Capacity enables growth. The risk is falling revenue per unit alongside long-lived purchase obligations. Supply delays and weaker demand can both hurt profitability.

Sources: S-1 material contracts and commitments, EO analysis.

6. Competitors: China’s funding base is growing too

• Kimi versus Anthropic: $50B versus $965B, roughly a 19× difference in recent reported or announced valuations. Benchmark gaps and valuation gaps do not move in the same proportion.
• DeepSeek conversion: Using the report’s CNY 106B = $15.9B conversion implies over about $90B post-money. This is a simple calculation, not a precise valuation using current exchange rates.
• Takeaway: Chinese labs also command tens of billions in valuation and substantial funding. Private funding valuations, public market caps, revenue run rates, and net income measure different things.

Sources: Anthropic, OpenAI March·October talks·revenue reporting, Kimi funding report·Sacra revenue data, DeepSeek October funding, Mistral funding·revenue target, SpaceX 10-Q.

What does $1–2T buy? Compare companies already earning profits

• Basis: Market caps use October 30 closing-price data. All results are GAAP and cover three months, but quarters end in June for Anthropic and Apple, August for Broadcom, and September for Meta.
• Scale: A $1.5–2T IPO would price Anthropic alongside Meta and Broadcom. Those companies already earn around $13–16B in quarterly operating profit.
• Anthropic’s premium: Q2 operating profit was about $1B. A comparable valuation would pay much more for future growth and margin expansion than current earnings.
• Apple is a larger reference: Roughly $5T in market cap and $30B in quarterly net income. It shows the scale of already-established profits.
• One-off items: Meta’s operating profit includes a $9.89B legal-settlement-related expense. Anthropic’s large net loss includes the financial remeasurement discussed above.

Financial statements: Meta, Broadcom, Apple, Anthropic S-1. Market caps: CompaniesMarketCap’s Meta, Broadcom, and Apple.

7. Safety: Controls can constrain revenue growth

“catastrophic or existential risks to humanity”

— S-1 risk factors, excerpt

• Actual external contact: During an AISI evaluation, Mythos 5 tried to induce an external developer to approve a malicious change. The developer refused; no actual harm was reported. The key issue is an attempted attack outside the permitted scope. AISI report
• Controls can mask capability: In the company’s own cyber evaluation, out of 50 attempts per tier, general access had 0 successes versus 34 for Red Team. The other 16 Red Team attempts were incomplete. A limited test, not a real-world crime success rate. Program details
• Revenue tension: An estimated 5–7% of January–September revenue came from customers later judged to violate usage policies or commercial terms. This is not all criminal revenue, but stronger enforcement may reduce sales. S-1
• Shareholder control: The founders’ LLC holds 50.1% of ordinary voting power; the LTBT trust appoints 4 of 7 directors. Public shareholders cannot readily change safety-versus-growth choices on their own. S-1 governance

Market voices

8. IPO outlook: Connecting growth to profit

• Confirmed: A Class A offering is proposed under Nasdaq ticker ANT. Price, share count, and listing date remain unset. Official announcement
• Reported expectations: $1.5–2T valuation and up to $100B raised. Neither an offering-price range nor confirmed proceeds. Reuters, Axios
• This article’s base assumptions: $1.2–1.6T post-offering value and $60–90B in new shares. Calculated from $100B of 2027 annual revenue × 12–16. Editorial scenarios, not company guidance or consensus.

• Why $100B? A cautious assumption allowing for a miss against the year-end $110B run-rate forecast or lower prices. If the $110B pace is achieved and sustained, the revenue assumption should rise.
• What $2T requires: Continued high growth + usage gains that offset price declines + profits and cash sufficient to support compute commitments.
• IPO price per share: Post-offering value ÷ fully diluted share count. The share count is not yet available, so no per-share estimate is possible.
• Tradable supply: The S-1 describes 180- and 365-day lockups for certain holders, with early-release conditions. Missing amounts require an amended filing. S-1

• Watch the next filing: Confirmed Q3 revenue and earnings / post-offering share count and dilution / cash remaining after long-term commitments.
• Data: Original charts are available as PNG and SVG. Data CSV · Sources JSON.
• Original material: AA charts are screenshots of the source articles, preserving logos and figures. State of AI charts: Nathan Benaich / Air Street Capital, CC BY 4.0. Quote translations, interpretation, and IPO scenarios are this article’s work.