Anthropic Delays IPO to November, Targets $100B Revenue Run Rate

Full analysis

The Skeptic

"Gross margin above 80% when excluding revenue-sharing and model-training costs" is not a gross margin. It is a number built by removing the two costs that define the business. Every frontier lab can play this. Amazon and Google take their cut whether Anthropic likes it or not, and training the next model is not optional. Strip stock comp and training and call it "near breakeven," and you have described a subset, not a company. The $100 billion run rate needs enterprise contracts to close roughly 54% faster in one quarter than they did all summer. Nobody outside the roadshow has seen that pipeline convert. A clean "from burning cash to breakeven" arc in four quarters is too tidy for a business this compute-heavy.

The Safety Lens

A former Nasdaq executive said the IPO "went from being a no-brainer to not being a no-brainer" after Amodei's slowdown calls. Read that plainly: public markets are pricing Anthropic's core brand, responsible AI, as a cost. That is a real problem for anyone who buys the safety pitch. The more seriously a lab constrains its own deployment pace, the more the market discounts it. And Anthropic is already responding, quietly lining up new models to change the story back to capability right before the offering. The lesson for buyers who chose Claude partly for the governance posture: that posture bends under quarterly earnings pressure like everyone else's. Don't assume the safety-first identity is a fixed feature of the company you're building on.

The Researcher

Watch what's in the denominator. The $100 billion headline projects roughly 54% growth from July to December in a single quarter. That implies either a genuine acceleration in enterprise close rates or aggressive counting of committed-but-not-recognized pipeline. Either way, an outsider trying to model this lab's actual unit economics is working blind, because the "80% margin" figure excludes revenue-sharing with Amazon and Google plus training amortization, and those are structural, not one-offs. The one clean number is the cost trajectory: $2.30 per revenue dollar down to near breakeven. That is real. It just doesn't tell you what happens when the next training run lands.

The Compute Pragmatist

That cost improvement tracks the serving side getting efficient, better use of the hardware, cheaper inference per query as the H100 clusters matured. Good news, and it's real. The trap is assuming it extends to the whole business. Training the next frontier model still burns nine figures in a compressed window, and that cost is getting faster, not cheaper. Calling the margin "80% when excluding training" is a boast about serving efficiency that doubles as a concession on model development: one part of the business got lean, the other did not. This math only closes because Amazon and Google front the infrastructure. So the durable question is what breakeven looks like the quarter Anthropic trains Opus's successor. The efficiency curve on serving doesn't rescue you from a step-change in training spend.

The Builder

New Sonnet, Opus, and Fable in stealth testing, all landing close together before an offering, tells you to plan for version churn. When a lab needs a capability narrative for investors, the release calendar bends toward benchmark legibility and press-readiness, not toward the API reliability improvements you actually want. Expect prompt regressions, documentation lag, and possibly new enterprise pricing and context window changes bundled in. If you run production pipelines on Claude, budget re-evaluation cycles for Q4. The last thing you want is three new variants dropping the week your on-call engineer is already stretched.

Where they disagree

The Researcher and Compute Pragmatist accept the cost story as genuine progress; the Skeptic says "near breakeven after stripping the two costs that matter" is barely a story at all. Both are right about different things. Serving got cheaper. The company did not become profitable. The other split is Safety versus everyone: is the market's discount on Anthropic's safety brand a temporary roadshow problem or a permanent tax on governance-forward labs? If it's permanent, every lab learns to talk capability and mute safety before going public.

What this hinges on

One belief: does the $100 billion run rate convert, or is it pipeline counted generously to make November look good? If Q3 financials show recognized revenue near that pace, the delay was smart sequencing. If the number leans on committed-not-recognized bookings, the "80% margin" framing and the $100 billion claim are the same move, choosing the flattering denominator. For anyone building on Claude, the practical read is simpler: a capability push is coming, and it's timed for investors, so treat the Q4 model drops as a churn event to test around, not a gift to deploy on faith.

Prediction: Anthropic will complete its IPO no earlier than 2026-11-01, and its official S-1 filing will report a gross margin figure below the "above 80%" the roadshow floated, because the audited version has to include the revenue-sharing and training costs the pitch stripped out.

Confidence: Medium. SEC filings force full-cost reporting, and the flattering number is a private-pitch construct.

Why: The "80% margin when excluding revenue-sharing and model-training costs" number is built by removing exactly the costs a public filing cannot remove. Revenue-sharing with Amazon and Google and training amortization are real cash and real obligations, and an audited S-1 has to show them. Every frontier lab has this same gap between the roadshow contribution margin and the all-in gross margin, so the private number is systematically higher than the filed one. The opposite outcome, an audited 80%-plus gross margin on frontier model serving, would mean the training and partner costs are trivial, which contradicts the nine-figure training runs and the revenue-share structure everyone acknowledges.

Revisit by 2027-02-28: We're right if Anthropic's S-1 or first public financials report an all-in gross margin below 80%. We're wrong if the filed gross margin is 80% or higher, or if no filing appears by this date because the IPO slips again.

The IPO could slip past November again, which is why the wrong-if includes it. But the margin gap is the durable claim. The audited number and the roadshow number were never going to be the same.