OpenAI and Anthropic Now Capture 89% of AI Startup Revenue. Agent Builders Should Be Worried.

The numbers are now clear enough to stop debating and start planning around. A group of 34 AI startups generates nearly $80 billion in annualized revenue, up 112% in six months, according to an analysis by The Information. OpenAI and Anthropic capture 89% of that total. Six months ago, their combined share was 84.5%.

The concentration is accelerating, not stabilizing.

The Revenue Picture

Anthropic recently surpassed OpenAI in revenue, driven largely by AI coding tools and enterprise white-collar automation, The Decoder reported. Below the two leaders, Perplexity, ElevenLabs, and Cognition have each crossed $500 million in annualized sales since December, joining Cursor in the next tier, according to MediaPost’s analysis of The Information’s data.

The headline numbers come with caveats. Anthropic shares revenue with Amazon and Google, which resell Claude to their cloud customers. OpenAI shares 20% of revenue with Microsoft through 2030, capped at $38 billion in total payments. These arrangements inflate the reported totals but don’t change the structural picture: two companies control the model layer that everything else runs on.

Both companies also burn through more than $30 billion per year combined, mostly on training costs, per The Decoder.

Why This Matters for Agent Infrastructure

Every agent framework, every autonomous workflow platform, and every vertical AI application that calls Claude or GPT endpoints depends on pricing decisions made by two companies. That dependency has concrete recent precedents.

Anthropic introduced separate Agent SDK billing credits starting June 15, restricting third-party harness access to Claude subscriptions. OpenAI launched Guaranteed Capacity, a commit-based offering requiring 1-3 year compute reservations. Both moves shift the economics away from pay-as-you-go flexibility toward lock-in structures.

When 89% of the revenue in your supply chain goes to two vendors, those vendors set your margins. They decide what API access costs. They decide which use cases get preferential pricing and which get throttled. They decide, through model capability gates, what your agents can and cannot do.

The Sequoia Counterargument

Sequoia Capital has argued that the real value resides in the application layer, not the model layer, stating that “tools people actually use will win,” according to MediaPost. The thesis is that foundation models will commoditize over time, like cloud compute before them, and sustainable margins will accrue to the applications that embed them.

The revenue data complicates that thesis. If commoditization were happening, OpenAI and Anthropic’s share would be shrinking as alternatives gained traction. Instead, their share grew 4.5 percentage points in six months. Mistral, Cohere, AI21 Labs, and the rest of the non-leader cohort are growing in absolute terms but losing relative ground.

Microsoft generates at least $6 billion in actual revenue from Copilot-branded apps powered partly by OpenAI models, and a dozen enterprise software firms including Salesforce and ServiceNow collectively generate billions from AI features built on these two providers’ models, per MediaPost. The application layer is generating revenue. It’s also cementing the duopoly’s position as essential infrastructure.

The Pricing Power Test

Both OpenAI and Anthropic are approaching public listings. Public companies optimize for revenue growth and margin expansion. The most direct path to both is raising prices on the customer base that has already built their products around your API.

For agent builders, the question is whether to treat this concentration as temporary (open-source catches up, pricing competition emerges) or structural (two companies control the capability frontier and will use that position to extract value). The revenue trend over the last six months suggests structural.

The practical response is architectural: agent platforms that can swap model providers without rewriting their orchestration layer have optionality. Those hardcoded to a single provider’s API, tool-use format, or reasoning behavior are building on rented ground. At 89% concentration, the rent is going up.