AI Capital Cycle ② — The Model Duopoly & Claude Code

In Memo 1 we concluded — *the weakest node in the circuit is the first model company that can't raise its next round*. The circuit's self-validation is simple. Can the model company's revenue catch up to its compute cost?

OpenAI revenue: 2024 $3.7B → 2025 $13.1B → 2026 ARR $25B+. Explosive growth. But against compute costs — 2026 OpenAI loss ~$14B, cumulative loss to 2030 $44B+, HSBC estimates a $207B funding shortfall by 2030. *On current trajectory, profitability does not arrive even by 2030*.

Anthropic is more interesting. Revenue 2024 $1B → end 2025 $9B → April 2026 $30B run-rate. Dario Amodei in May: "We tried to plan for 10x growth. We got 80x. Too hard to handle." Compute spend 2025 $360M → 2027 $6.4B → cumulative ~$80B through 2029. But 2028 revenue $70B, FCF +$17B (leaked deck). *Profitability possible by 2027*.

Same industry, two model companies. One can't turn profitable; the other sees it in 2-3 years. What's different?

Nearly half of Anthropic's growth comes from one product. Claude Code.

Timeline:
- Feb 24, 2025 Public preview
- May 2025 General Availability
- November 2025 $1B ARR (6 months post GA)
- February 2026 $2.5B ARR (9 months post GA, disclosed at Series G)
- May 2026 — ~17-20% of Anthropic's total ARR; coding workloads overall ~50% of API revenue

Reaching $1B ARR in 6 months is *faster than ChatGPT consumer's ramp*. Revenue per user isn't $20-200/month consumer — it's *enterprise developers at $100-500/month*.

What is it? A CLI-native autonomous coding agent. A user says "fix this bug"; in the terminal, it reads the codebase, edits files, runs tests, commits to git. If Cursor is a chat/agent embedded in an IDE, Claude Code is *terminal-and-agent native*. If Devin is a cloud sandbox doing async PR-style work, Claude Code is *local and in-the-loop*.

The productivity shock: 70-90% of code at Anthropic is written by Claude (Dario, September 2025). Claude Code creator Boris Cherny: "Coding is now 'solved' for most use cases. I haven't written a single line by hand since November."

Claude Code's real significance isn't revenue. It is the first quantified proof that AI actually displaces labor.

SemiAnalysis: ~4% of all public GitHub commits authored by Claude Code (Feb 2026). Doubled in 60 days. +8% WoW. At the same pace, by end-2026 over 20%.

Pragmatic Engineer Feb 2026 survey, 15,000 developers: Claude Code #1 most-used AI coding tool; 46% "most loved."

Microsoft internal (Verge): "Mostly we use Claude code, because it's way out ahead of all the other AI models in that area." *Microsoft makes GitHub Copilot. Its engineers use Claude Code instead*. The strongest qualitative PMF evidence.

Claude Code's $2.5B ARR — *the first quantified proof that real end users pay real money and real work is substituted*. In nine months.

It's not only Claude Code. *Third-party tools built on Anthropic's model also explode independently*.

Cursor (Anysphere). Four MIT 2022 graduates. CEO Michael Truell, age 25. ARR: Aug 2024 $100M → June 2025 $500M → Nov 2025 $1B → Feb 2026 $2B. 20x in 13 months. Funding: Nov 13, 2025 Series D $2.3B at $29.3B → April 2026 in talks at $50B (a16z + Thrive) → April 22, 2026 xAI option to acquire at $60B later in 2026 (Fortune). Model mix Claude-weighted.

Cognition (Devin). May 27, 2026: $1B round at $26B post. ARR $492M, 50% MoM growth for 6 consecutive months. Q2 2026 ARR is 73x Q1 2025. Mercedes-Benz: 200,000-line COBOL modernization — human estimate 8 months → Devin 8 days. Customers: Goldman Sachs, Citi, US Army, NASA, Dell, Santander, BMW.

GitHub Copilot. Owned by Microsoft. 4.7M paid subscribers (Jan 2026, +75% YoY). Estimated ARR $1B+. Claude Sonnet 4.6 added on March 9, 2026. "Pick your agent" Agent HQ launched. Microsoft putting Claude into its own product means — for Anthropic, "our model is now in GitHub." For Microsoft, close to admission: "our users prefer that model."

Memo 1's circuit (NVDA→OpenAI→MSFT→NVDA) is just the first circuit. A second circuit — *between model companies* — has emerged.

On May 20, 2026, SpaceX's S-1 IPO filing revealed: Anthropic leases the entire output of xAI's Colossus 1 cluster (220K GPUs, 300MW) for $1.25B per month. Through May 2029. Total $40B+. Extending into Colossus 2. From May, $15B/yr to one counterparty: xAI.

Why does Anthropic buy from xAI? It already uses AWS Project Rainier (500K Trainium2) and Google TPU. But all combined are insufficient for demand Claude Code generates. xAI's Colossus was nearly the only place with "spare GPUs."

From xAI's side — Grok doesn't generate enough revenue. Colossus capex can't be recouped. Anthropic creates the *lease revenue* filling that capex. xAI uses it to buy more NVIDIA GPUs (Series E $20B at $230B, Jan 2026).

The second node:

*Anthropic revenue (Claude Code) → xAI lease → xAI capex → NVIDIA revenue → NVIDIA invests in OpenAI → OpenAI commits to Azure → MSFT buys from NVIDIA*

*Both circuits meet at NVIDIA*.

Implications: first, *as long as Anthropic's revenue is real enterprise demand*, the circuit is justified. Second, *if Anthropic revenue slows*, this node breaks first. Anthropic can't pay xAI lease; xAI cuts NVIDIA capex. The circuit unravels in reverse.

We return to the series-wide watch metric: *Anthropic's next-round valuation and revenue growth deceleration*.

The single proposition: AI's first PMF is coding. Claude Code is its first quantified proof.

OpenAI and Anthropic's unit economics split apart on revenue mix. OpenAI bears consumer ChatGPT's churn and CAC. Anthropic enjoys enterprise API's 88% retention and less-than-linear scaling. *The difference in 2027 profitability is not model quality but business model.*

And Claude Code — $2.5B ARR in 9 months — gives the first answer to the series's core question. *Can AI actually displace labor?* In coding, it has already happened. 4% of public GitHub commits, 70-90% of Anthropic's internal code, Mercedes' 8-month job in 8 days.

This is the first datum of the circuit's self-validation. In Memo 1 we asked: "When buyer and seller are the same person, what does the market price mean?" Claude Code's $2.5B is — *evidence that in at least one node, the buyer is not the seller*. Real enterprise customers pay real $100-500/month to substitute real coding labor.

But this is not the circuit's full validation. Outside coding — legal, accounting, marketing, customer service, design — the PMF data is not as clean. The final memo (Anthropic Economic Index) looks at all occupations.

*Claude Code is the first quantified proof. Whether the next occupation's data shows the same curve will decide the cycle's fate.*