Serving Claude is cheap. Training the next one is not

On November 2, Dario Amodei and his co-founders filed an amended Form S-1/A with the Securities and Exchange Commission to list Anthropic under the ticker ANT. Most coverage stopped at two numbers: nine-month revenue of $33,242 million, up from $2,545 million a year earlier, and a quarterly personnel bill of $1,711 million across roughly 5,300 employees.

The more revealing number sits in a footnote on operating expenses by function. Anthropic splits its compute bill between serving the models it already has and building the ones it does not have yet.

FIG. 01 · OPERATING SPEND

Serving existing Claude models took $4,812M of Anthropic's $34,502M nine-month spend. R&D and growth took $29,690M.

Anthropic, PBC operating expenses by function, nine months ended Sep 30, 2026 (USD) · Total $34,502M

Serving customers (cost of revenue)$4,812M

• Compute and infrastructure for customer workloads$4,812MAnthropic Form S-1/A (SEC)

Training next models and growth$29,690M

• Research and development (model training and R&D staff)$19,240MAnthropic Form S-1/A (SEC)
• Sales and marketing (channel partners and marketing)$6,619MAnthropic Form S-1/A (SEC)
• General and administrative$3,831MAnthropic Form S-1/A (SEC)

• Nine-month revenue reached $33,242M, leaving a $1,260M operating loss. Anthropic Form S-1/A (SEC)

Sources: Anthropic Form S-1/A (SEC). Figures in millions from the operating expenses by function footnote in Anthropic, PBC Form S-1/A (filed Nov 2, 2026). Below the operating line, non-cash convertible note and SAFE remeasurements drove GAAP net loss.

Across the nine months ended September 30, 2026, cost of revenue, which the filing defines as substantially all compute and infrastructure for customer-serving workloads, was $4,812 million. Against $33,242 million in revenue, serving existing Claude models carried an implied gross margin of 85.5 percent. That is classic enterprise software profitability.

Yet Anthropic still posted a $1,260 million operating loss over those nine months (before a non-cash $94,324 million remeasurement on convertible notes, derivatives and SAFEs below the operating line). Research and development alone consumed $19,240 million, or 55.8 percent of total operating spend, followed by $6,619 million in sales and marketing and $3,831 million in general and administrative costs.

One profitable quarter, then back to the furnace

The quarterly table on page 160 shows how briefly a frontier lab gets to keep its margin. In the second quarter of 2026, on $11,555 million in revenue, Anthropic reported $974 million in operating income, an 8 percent operating margin. One quarter later, revenue jumped to $16,956 million, yet the company swung back to a $567 million operating loss as quarterly compute and infrastructure climbed to $9,863 million.

Why not slow the training runs and bank the cash? On page 39 of the prospectus, Anthropic writes the answer itself: open-weight models compete directly for “enterprise workloads that prioritize cost efficiency over intelligence.” Three days before the amended filing, an engineering lead from one of Europe’s largest fintechs walked onstage in New York and put exact numbers on that sentence.

Why Revolut stopped rationing tokens

At AI Engineer New York, David Abitbol, who leads AI operations at Revolut, showed a year of output data covering 2,000 engineers at the 80-million-customer digital bank. Early rollouts of metered frontier coding tools hit a billing wall.

“We were seeing engineers spend a thousand a month, ten thousand a month with not so much ROI. And so, when we launched it, we launched it with strict budget limits.”

Under strict monthly caps, engineers hoarded their allowance for end-of-quarter crunch work or only the hardest bugs. Starting in April 2026, Revolut rented 56 dedicated Nvidia B200 GPUs on US infrastructure with zero data retention and self-hosted open-weight GLM models, moving up to the 800-billion-parameter GLM-5.3 quantized to FP4.

Without adding a single GPU, tuning KV cache hits above 90 percent and cutting context windows brought time-to-first-token down from more than 10 seconds in April to 1.6 seconds in August, even as request volume rose 76 percent. Revolut’s blended token cost landed at $0.30 per million tokens, cutting its monthly coding invoice to one-tenth of what it had been paying Anthropic and OpenAI. With uncapped tokens, coding time improved 60 percent and feature lead time improved 70 percent, mostly because engineers used the model freely on routine inbox, Jira and documentation chores.

FIG. 02 · TWO LEDGERS

Anthropic earns an 85.5% margin serving today's models. Revolut cut its coding bill 90% by hosting open weights.

Frontier supplier vs large buyer, as reported Oct 29 to Nov 2, 2026

Metric	Anthropic (S-1/A)	Revolut (GLM-5.3)
Scale reported	$33,242M 9M revenueAnthropic Form S-1/A (SEC)	2,000 engineers on 56 B200sDavid Abitbol, AI Engineer NY
Serving cost	$4,812M (14.5% of revenue)derived	$0.30 per 1M tokens blendedDavid Abitbol, AI Engineer NY
R&D and model training	$19,240M in 9M 2026Anthropic Form S-1/A (SEC)	Open weights (800B params, FP4)David Abitbol, AI Engineer NY
Customer spend shift	~900% net revenue retentionAnthropic Form S-1/A (SEC)	1/10 of frontier API invoiceDavid Abitbol, AI Engineer NY
Role of frontier models	11,900+ $100K+ customersAnthropic Form S-1/A (SEC)	Kept as rationed exceptionsDavid Abitbol, AI Engineer NY

Sources: Anthropic Form S-1/A (SEC), David Abitbol, AI Engineer NY. Anthropic figures cover the nine months ended Sep 30, 2026 (Form S-1/A, Nov 2, 2026); customer count and retention are as of Sep 30, 2026. Revolut figures are from AI Operations Lead David Abitbol's AI Engineer New York talk (published Oct 29, 2026).

Revolut did not cancel Claude Code or Codex. Abitbol said his team keeps frontier models as rationed exceptions for the hardest problems while routing everyday volume to open weights. Even further down the stack, cheap coding tokens are compressing software pricing: on November 5, laid-off engineers at OccupySaaS shipped the first Apache-2.0 release candidate of ZeroAuthy, a clean-room Go replacement for Auth0 built with two isolated AI-assisted teams.

What I take from it

Anthropic enters the public market with $131.9 billion in cash and short-term investments, roughly 900 percent net revenue retention and more than 11,900 customers paying at least $100,000 a year. Demand for frontier intelligence is real, and serving yesterday’s weights is already cheap.

The catch is that yesterday’s weights do not stay proprietary for long. Once open weights clear the bar for 90 percent of daily engineering work at $0.30 per million tokens, a frontier lab can only charge frontier prices for the top decile of tasks that open models still botch. Anthropic’s 85.5 percent serving margin is real, and so is the $19,240 million R&D bill required to defend it for another six months.