Exa's AI-search bet: brand searches up 1,560% YoY, $250M Series C

Exa is an AI-native web search API (plus Websets for bulk retrieval) that returns ranked, full-content results designed for AI agents and LLM apps instead of a human-facing list of links.

That AI agents would search the web more than humans, so the winner is an API built for machines — full content, sub-200ms speed — not another Google for people.

~$5M seed (YC S21, 2022); $17M Series A (July 2024); $85M Series B at $700M (Sept 2025); $250M Series C at $2.2B (May 2026).

Will Bryk and Jeff Wang, friends from Harvard, started Exa (originally Metaphor) in 2021, betting Google could be beaten with neural search — queries like 'give me all ML engineers in NYC who have a blog, sorted by years of experience'. They bought a GPU cluster, built a large-scale index, and launched a consumer search engine in November 2022 — two weeks before ChatGPT. Within weeks AI builders were asking for API access, and Exa pivoted to search infrastructure for AI.

Exa announced the first web search API for AI in early 2023. After a ~$5M seed and a $17M Series A led by Lightspeed with NVentures and Y Combinator in July 2024, Benchmark led an $85M Series B at a $700M valuation in September 2025 (Peter Fenton joined the board), with Lightspeed, YC and NVIDIA participating. Cursor, Notion AI, and thousands of companies adopted the API, and US organic brand searches grew roughly 16x year-over-year by April 2026.

Still scaling: on May 19–20, 2026, a16z led a $250M Series C at a $2.2B valuation, with Exa reporting 400,000+ developers and 5,000+ companies. The company is training next-generation retrieval models, scaling to hundreds of thousands of searches per second, and building a global GTM team with LaunchDarkly's Marcus Holm as CRO.

Bryk and Wang recognized early that AI agents need a different search product — machine-readable full content, speed, and no SEO/ad incentives — rather than links designed for human eyeballs.

Exa built its own index, embedding database and crawler instead of wrapping Google or Bing, enabling zero data retention, custom classifiers and sub-200ms latency.

The timing wedge was accidental but decisive: the first engine shipped two weeks before ChatGPT, so Exa was already the obvious API when the first generation of AI apps needed web grounding.

Vertical index specialization (code, people, companies) made Exa the default search layer for coding and GTM agents, turning usage into brand search and press coverage.

When a platform shift changes who consumes a product (humans → agents), the infrastructure layer wins early: don't outrank Google for people, become the API the next 10,000 AI apps call.

As of 2026-09-02, Exa continues to scale as the search infrastructure layer for AI agents. After the May 2026 $250M Series C at a $2.2B valuation, it is training next-generation retrieval models, expanding its GPU cluster, scaling toward hundreds of thousands of searches per second, and hiring a global go-to-market organization (Marcus Holm, formerly LaunchDarkly president, joined as CRO). It reports over 400,000 developers and 5,000+ companies on the API. All customer and growth figures are company-reported.