Exa Labs has closed a $250 million Series C led by Andreessen Horowitz at a $2.2 billion valuation, the San Francisco startup said Wednesday. The round more than triples the roughly $700 million valuation Exa carried after its $85 million Series B last fall — a raise that, per reporting, included Nvidia and Y Combinator — and brings total disclosed funding to around $357 million for the five-year-old company.
The pitch is narrow and infrastructural: a web search API built for machines, not humans. Exa operates its own independent index instead of reselling results from an incumbent. "Most other search providers actually wrap other search engines and therefore cannot compete on quality/latency/cost," CEO Will Bryk said. Co-founders Bryk and Jeffrey Wang trace the idea to an early transformer-inspired search engine they built at Harvard.
The retrieval stack underneath agents
Exa stores the web as embeddings in a custom vector database and queries billions of them in roughly a tenth of a second, reportedly using less memory than a high-end PC. A custom software layer, exa-d, parallelizes data processing across GPUs, and the embedding models are trained on an in-house Nvidia cluster. The product surface splits into Exa Instant — queries returned in under 180ms, which the company bills as the fastest of its kind — Contents for full webpage text retrieval, and Exa Agent for multi-step workflows.
That latency and freshness profile is the point. Agentic systems issue far more queries than human users, hit the index repeatedly inside a single task, and choke on the round-trip cost of a slow or rate-limited search call. A retrieval layer that returns clean page text in sub-200ms is a different primitive than a consumer SERP.
Traction and where the money goes
Exa reports more than 5,000 customer companies and over 400,000 developers, naming Cursor, Cognition, HubSpot, OpenRouter, and Monday.com among them, alongside Fortune 500 accounts. The new capital funds more AI infrastructure and hardware, training of new models, scaling toward "hundreds of thousands of searches per second," and go-to-market hiring.
Bryk frames the bet in volume terms: "As trillions of agents come online over the coming years, search needs will grow thousands of times beyond the total search volume of Google."
What it changes for builders
The raise is a vote that retrieval becomes a dedicated, paid infrastructure layer rather than a feature bolted onto an LLM. For teams shipping agents, the practical question shifts from "which model" to "which index" — whether to depend on a wrapper reselling Bing, a frontier lab's built-in browsing, or a native embeddings engine tuned for latency and recall. With a16z anchoring a $2.2 billion valuation, Exa is betting the answer increasingly favors purpose-built search, and that the economics of agentic query volume will reward whoever owns the index rather than rents it.
