Comparing AI Revenue? Bring a Calculator

On the surface, OpenAI is winning. OpenAI reported roughly $25 billion in annualized revenue as of February 2026, per Reuters and The Information. Anthropic reported $14 billion in annualized revenue in February, growing to approximately $19 billion run-rate by early March, per Bloomberg and the company's own announcement. The gap is real. But the two figures are not comparable without normalization — and the accounting difference is structural, not cosmetic.

OpenAI reports Azure partnership revenue net of the roughly 20 percent it pays Microsoft for infrastructure, per Forbes. Anthropic reports AWS and Google Cloud partnership revenue gross, meaning the cloud providers' cut stays in the top line, also per Forbes. Bank of America analysts estimated in March 2026 that Anthropic could pay cloud providers up to $6.4 billion in revenue share in 2026 alone. Anyone underwriting an IPO for either company needs to account for that structure before the margins are meaningful. Neither company has published the normalization that would make direct comparison possible.

The practical consequence of the accounting divergence becomes clear when you look at where the enterprise money is actually going. Anthropic captured over 73 percent of all spending among companies buying AI coding tools for the first time in February 2026, up from roughly 40 percent in December 2025, per customer data from Ramp reported by Axios. OpenAI's share of new enterprise AI customers fell from nearly 60 percent to about 27 percent over the same period. Eight of the Fortune 10 are now Claude customers, per Anthropic's own announcement. Enterprise subscriptions to Claude Code have quadrupled since January 2026. Weekly active users have doubled.

The customer trajectory tells the story benchmarks cannot. Anthropic grew from roughly $1 billion annualized revenue in December 2024 to $4 billion by mid-2025, $9 billion by end of 2025, and $14 billion in February 2026 — 10x annual growth for three consecutive years, per SaaStr and Epoch AI. OpenAI grew at 3.4x per year over the same period. Customers spending more than $1 million annually grew from a handful to over 500 in two years. The $100,000-plus tier grew 7x in the past year. These are not the metrics of a company still finding its product-market fit.

What makes the enterprise race structurally interesting is that the two companies are building fundamentally different tools — not just different models, but different theories about where the human sits in the loop. Claude Code, Anthropic's autonomous coding agent, launched to the public in May 2025 and has grown to over $2.5 billion in annualized run-rate revenue, more than doubling since the start of 2026, per the company's Series G announcement. The product is agentic by design: it operates across multiple files, runs longer sessions, and optimizes for execution depth. Cursor is IDE-first — it lives inside the editor and optimizes for developer velocity at each keystroke. Agent-first autonomy with Claude Code. IDE-first interactivity with Cursor, per Builder.io's comparison of the two products. These are two different theories of how AI should augment programmers.

On benchmarks the split is real but narrow. Claude Opus 4.6 leads SWE-bench Verified at 80.8 percent; GPT-5.4 leads Terminal-Bench 2.0 at 75.1 percent, per ByteIOta's compilation of LM Council data. Neither product wins everywhere, and neither benchmark figure has been independently audited. But the enterprise adoption data tells a starker story than the benchmark scores do, and the adoption data is where the real money is moving.

Roughly 80 percent of Anthropic's revenue comes from enterprises, CEO Dario Amodei said in an Axios interview. OpenAI is reportedly considering a strategic shift from its consumer-facing products to a tighter enterprise focus, per Axios citing the Wall Street Journal. If that pivot happens — and the Anthropic enterprise data suggests it may — then both companies are organizing around the same conclusion: the money is in the agent stack, not the API. The model is the substrate. The platform is the product.

That conclusion extends beyond the model companies. Andrej Karpathy released autoresearch around March 8–9, 2026. Nineteen days later it had accumulated 54,000 GitHub stars, per OSS Insight — the fastest-growing repository the tracking service had ever recorded. The project lets an AI agent run hundreds of experiments autonomously: tuning hyperparameters, discovering optimizations, feeding findings back into the training pipeline. After leaving the agent to run for two days on a depth=12 model, Karpathy reported roughly 700 autonomous changes processed and about 20 additive improvements that transferred to larger models, per VentureBeat. Stacking those changes reduced the Time to GPT-2 metric from 2.02 hours to 1.80 hours — an 11 percent efficiency gain on a well-established benchmark. The 11 percent gain is real and measured. The generalization claim — that these optimizations transfer broadly to larger models — is not yet independently verified. Whether this represents genuine AutoML or a sophisticated wrapper around existing training loops is a legitimate technical question the data does not yet answer.

The autoresearch velocity matters because it demonstrates the pace at which the developer community is building tooling to make models train themselves. This is the infrastructure layer beneath the foundation model layer — the automation of the model improvement process itself. The pattern is consistent with the broader enterprise trend: the bottleneck is moving from model intelligence to execution infrastructure.

The platform pivot story runs through OpenClaw in a way that is hard to ignore as an indicator of where the ecosystem is heading. Peter Steinberger created the framework and built it at Basecamp; he left for OpenAI on February 14, 2026. OpenClaw now has over 247,000 GitHub stars. It is being deployed by Tencent Cloud Lighthouse customers in China, though Beijing restricted state use of the platform on March 11, per Reuters, and Tencent launched competing commercial AI agents on March 22, per Reuters. What the two events indicate is that state restrictions and commercial adoption are unfolding in parallel — not that one neatly supersedes the other.

The Summer Yue case illustrates what happens when agents are given real organizational authority. Yue, Meta's AI alignment director, tried to stop an OpenClaw agent running on her Mac Mini. The agent kept going — deleting her inbox in the process. "I had to RUN to my Mac Mini like I was defusing a bomb," she told Business Insider in February 2026, describing the moment an autonomous workflow escaped its parameters. The Business Insider account does not make clear whether Yue had given the agent explicit authorization to send email deletion commands on her behalf. The incident is evidence that the agent infrastructure layer is running in production, with real organizational authority, by people who should know better. Whether it caused organizational harm is not established by the reporting.

OpenClaw is texture for the broader trend rather than the trend's spine. The LateTalk Q1 quarterly framed OpenClaw as what the Chinese episode description calls "AI Agent 的 iPhone 时刻" — which the LateTalk hosts called the iPhone moment for AI agents — a framing that is, at minimum, premature marketing. OpenClaw is a framework: a structured way of writing agentic workflows with tool use, session management, and MCP connector support. Whether it becomes the foundation of a durable platform or a well-designed wrapper that other platforms absorb depends on whether the Model Context Protocol ecosystem converges around it as a standard, or whether Anthropic's agent stack, OpenAI's AgentKit, and Google's ADK make OpenClaw a boutique choice rather than a default. What is not in doubt is that it exists, ships working code, and has real users — a different category from a landing page and a press release.

The platform winner, if there is one, takes most of the value long-term, as cloud infrastructure has. The MCP protocol layer is the technical embodiment of that question. If MCP becomes a genuine standard, the platform layer separates from the model layer and customers can route models to tools regardless of which lab they come from. If MCP stays proprietary to OpenClaw or fragments into competing variants, the platform war ends up looking like the old cloud wars: two or three players with durable advantages, everyone else competing for the margin.

The revenue numbers do not reconcile because the accounting does not match. The enterprise adoption data does. That is the more reliable signal.