Anthropic's Debut: The Order Book Is Loud, the Compute Bill Is Louder

Anthropic’s IPO showed strong demand. The harder test is how the 74/26 compute split, cash obligations and hyperscaler dependence affect the business model.

On 19 November 2026, Anthropic PBC began trading on Nasdaq under the ticker ANT, having priced its IPO at $83 a share — above the $70–78 range it had been marketing — in what Reuters described that morning as a record-setting listing. Nearly $500bn of orders were reportedly submitted, according to 247wallst, and Altimeter's $5bn cornerstone was confirmed on CNBC's Squawk Box the same morning. None of this is a disclosure. Order books are non-binding and padded, and a cornerstone tells you about conviction among the already-convinced. The more useful new information sits elsewhere: a compute-allocation disclosure that reframes the gross-margin argument, a Microsoft commitment far larger than previously understood, and a shift toward Anthropic operating its own data centres. Across the five debates that matter, this article asks one narrow question — does the new evidence resolve anything? Usage-driven growth: supported, and never the real question Demand is not in contention. Revenue for the nine months to 30 September 2026 was $33.24bn, up 1,206%, with third-quarter revenue of $16.96bn — up 1,127% year on year and 47% sequentially, per our 19 November 2026 report. Management's own run-rate disclosure, roughly $81bn annualised at 31 October from about $71bn at 30 September, is a consumption annualisation, not GAAP revenue and not a forecast. Enterprise customers spending at least $100,000 grew from roughly 950 in September 2025 to about 11,900 a year later. New product evidence reinforces it: Claude Fable 5.2 and Claude Mythos 5.2 shipped on 9 November 2026, following Opus 5.5 on 22 September and Haiku 5.5 on 7 October. But that sharpens the report's caveat rather than relaxing it. Revenue remains overwhelmingly consumption-based, and models launched within roughly six months of September generated more than 80% of third-quarter revenue. A booking curve that depends on shipping frontier capability every quarter is a different asset from a ratable seat business. Compute and margins: one number, two readings The report estimates third-quarter compute and infrastructure at about $9.86bn, or 58.2% of revenue, up from roughly 54.9% in the first half, and flags sustained readings above 70–80% for two or more quarters as a structural warning. The most consequential new datapoint — secondary reporting only, drawn from Fortune, CNN and Forbes coverage of the prospectus, and requiring primary-source verification — is that of roughly $18.8bn of nine-month compute spend, about 74% went to training, research, internal use and free-tier service, with only about 26% serving paying customers. If that holds, it explains why two apparently contradictory gross margins are both arithmetically right: the 38.3% trailing figure in market data and the roughly 85.5% functional margin Anthropic reports (about 85–86% in the third quarter, again per secondary reporting) describe different accounting geographies, not different economics. Anyone citing a sub-40% gross margin is not describing unit economics . But the same number cuts the other way. Seventy-four percent of a very large compute bill is not recovered through gross profit; it must be recovered through future pricing power or scale. The gross margin is defensible. The P&L is not. Both statements survive contact with the data, and neither should be smoothed away. Cash flow and the obligation stack The cash picture is intact. Nine-month 2026 operating cash flow was $6.57bn against $925m of capex — roughly $5.64bn of simple free cash flow , or about $3.69bn net of $1.955bn of infrastructure prepayments. The reported $36.5bn investing outflow is overwhelmingly the purchase of short-term investments, not an infrastructure bonfire. The obligation stack is the harder read. The report cites $319.99bn of contractual fixed and determinable obligations, a broader commitment aggregate above $900bn, about $70.7bn expected within twelve months, and $131.8bn of cash and short-term investments at 30 September. The $94.32bn of non-cash fair-value charges help explain a GAAP net loss of $104.04bn alongside $4.081bn of nine-month adjusted operating income — a gap that widened sharply in the third quarter and deserves more scrutiny than it has received. The genuinely under-appreciated point is concentration on the supply side. Three counterparties account for the bulk of that fixed-obligation figure: about $110bn minimum to Amazon, at least $121.3bn to Google for April 2026–July 2033, and — per secondary reporting pending confirmation in the filings — $47.2–49.2bn to Microsoft for November 2026–June 2033, alongside a $30bn Azure purchase commitment, a 1GW capacity option and roughly $5bn of Microsoft investment, with Claude now generally available in Microsoft Foundry. A deeper entanglement, not a diversification Cloud-provider marketplaces were 44% of nine-month revenue and 52% of net receivables , with the top five customers at roughly 25%. The new evidence makes that structure more explicit. Anthropic is now first-party inside three competing platforms — AWS Bedrock, Google Cloud Vertex AI and Microsoft Foundry — each of which runs its own frontier model, each of which is a supplier, and each of which now holds equity. That is not diversification ; it is a tripling-down on distribution through firms that are simultaneously customers, suppliers, shareholders and competitors, bound by multi-year minimums. The strategic upside is real — three channels, three custom-silicon programmes. The strategic cost is that Anthropic's largest customers, largest suppliers and largest shareholders are the same three balance sheets. Meanwhile, Anthropic Direct — a first-party data-centre strategy begun around 30 September 2026, leasing facilities from partners who hold the real estate and power, and described in secondary reporting as part of a compute plan of up to $567bn — is another item requiring primary-source verification. If it scales, compute cost migrates from purchased services to lease and depreciation, shifting cost between line items without necessarily improving unit economics. Any efficiency ratio built on the purchased-services line becomes a weaker tripwire. One small corroboration of the power constraint: the Wall Street Journal reported on 19 November that Anthropic is recruiting a natural-gas supply manager at around $400,000 a year. Valuation: flow, not fundamentals Everything published on pricing day about valuation is liquidity and sentiment. A non-binding order book . A cornerstone. A 190-point Dow open, with Micron up 4% and Western Digital up 5% on the AI read-through (Invezz, 247wallst, 19 November). And a cautionary precedent: SpaceX surged after listing, then fell roughly 50% from its peak, as the Motley Fool noted that morning. None of that moves next-twelve-month revenue, compute intensity or the commitment schedule. The framework that matters — forward price-to-sales , given negative GAAP earnings distorted by non-cash charges — has unchanged inputs. The one new valuation-relevant fact is structural: a capital mix shifting toward owned infrastructure arguably deserves a different multiple from either a pure software or a pure utility comparable. That is a question, not an answer. Where the sources disagree Two conflicts should be stated rather than resolved. Secondary coverage has described an "$80bn GAAP operating loss" for the first half; the report's evidence points to a nine-month GAAP operating loss of $(1.26)bn, suggesting the former conflates net loss with operating loss — treat the report figure as the working number pending the prospectus. And the 74/26 split, the Microsoft minimums and the Anthropic Direct plan all currently rest on secondary reporting; all three are load-bearing and none should be published as settled fact. What would settle the debate: the fourth-quarter print on 3 March 2027, the first quarter of GAAP operating profit with compute meaningfully below the flagged threshold, evidence that Claude Code and Cowork convert experimental spend into production spend, and hyperscaler capex commentary from Amazon, Google, Microsoft and Nvidia. The order book is loud. The compute bill is louder — and unlike the order book, it is binding. This article was produced with the assistance of artificial intelligence and is provided for educational and informational purposes only. It does not constitute investment advice, a recommendation, or an offer or solicitation to buy or sell any security, and it reflects information available as of 19 November 2026. Figures drawn from secondary reporting are labelled as such and have not been independently verified against primary filings. AI-assisted analysis can contain errors or omissions; readers should verify all figures against primary sources before relying on them, and consult a qualified financial adviser regarding their own circumstances.