Exa, the AI search engine built by Harvard alumni specifically for AI agents, just closed a $250 million Series C round led by a16z. The valuation jumped from $700 million last September to a whopping $2.2 billion — nearly triple in a year. The fresh capital will go toward expanding its proprietary index, training next-gen retrieval models, and scaling throughput to hundreds of thousands of concurrent requests per second.
https://twitter.com/ExaAILabs/status/2057132080317042697
Exa’s big bet is that AI search needs a completely different stack than web search. The company claims it's built everything from scratch: its own crawler, a native index tracking over 500 billion URLs, custom embedding models, and a third-generation vector database. That's a deliberate departure from the common practice of wrapping Google's search results — a strategy Exa executives dismiss as "Google shell" solutions. The engine offers response times ranging from 200 milliseconds (Exa Instant) to multi-minute autonomous deep searches (Deep Max), which recently achieved state-of-the-art results on the DeepSearchQA benchmark. New research also suggests that using Exa instead of a traditional search engine for reinforcement learning training can cut token consumption by 69%, hinting that search is evolving from a reasoning tool into a training substrate.
Exa is now processing roughly 1 billion queries per month — a tenfold increase in a year — serving clients like Cursor and Cognition, two of the buzziest names in AI. In a delicious twist, the same startup that once erected billboards reading "10^18 > 10^100" (that's Exa greater than Google) has since partnered with Google Cloud Vertex AI as one of the first grounding providers for Gemini enterprise customers.