Exa Raises $250M as AI Agents Threaten to Replace Human Web Search

Exa just closed a $250M Series C at a $2.2B valuation, led by Andreessen Horowitz. The timing was almost theatrical: Exa announced its raise on May 20, the morning after Google declared its own search box obsolete at I/O 2026. Whether that was coordinated or coincidence, the message landed clearly: the era of search built for humans is giving way to search built for machines.

What Exa actually is

The core difference from Google, Bing, or any traditional search engine is that Exa uses neural embeddings to understand query meaning, not keyword matching. The company was founded in 2021 by Will Bryk and Jeff Wang, two Harvard roommates who watched GPT-3 launch and saw a gap: GPT-3 could understand natural language at a level that Google's keyword algorithms could not match, but search was still built on term frequency and link graphs.

Bryk and Wang started as Metaphor Systems (YC W22), rebranded to Exa in January 2024, and raised $85M at a $700M valuation in September 2025. The Series C more than triples the five-year-old startup's valuation since that raise just last fall. The speed of that re-rating tells you everything about how hot this category has become.

The stack they built from scratch

Most search APIs you've heard of are wrappers. There's a reason there are more space programs than independent search engines , most other search providers actually wrap other search engines and therefore cannot compete on quality, latency, or cost. Exa went the hard route.

• Crawlers that track over 500 billion URLs, research teams that train special embedding models on a GPU cluster they assembled, and a new vector database built for the extremely high queries-per-second that agents need.
• The infrastructure is built in Rust, running a custom vector database with Matryoshka embeddings (a technique that stores embeddings at multiple resolutions for flexible precision tradeoffs), document clustering, binary compression, and assembly-level SIMD optimizations. The embedding model was trained for over a month on a 144-GPU H200 cluster.
• They built the fastest search API in the world at sub-200ms, and fast text extraction models that reduce LLM token counts by over 20x.
• The resulting database can search billions of vectors in under 100ms with high recall, while enabling keyword and metadata filters, and using less memory than a gaming PC.

The key insight behind the architecture is that AI agents have different needs from humans: the optimal response to a search query coming from an AI is not necessarily the same as one from a human. AI agents don't need to see ads, they need to see results. They don't want just the first few pages , they want all the relevant results to take advantage of growing context windows.

The numbers

The Series C round was led by Andreessen Horowitz, with participation from existing backers Benchmark, Lightspeed Venture Partners, and Y Combinator, alongside individual AI leaders including Scott Wu, Igor Babushkin, and Tal Broda. In conjunction with the funding, Sarah Wang of Andreessen Horowitz joined Exa's board of directors.

Sacra estimates that Exa hit $10M ARR in September 2025, up roughly 1,010% year-over-year from an estimated ~$0.9M the prior year. Revenue has scaled rapidly since launching its commercial API, driven by growing demand for AI-native web search as LLM-based applications move into production. The company monetizes primarily through usage-based API pricing rather than seats. Exa charges a flat rate per request: $5 per 1,000 searches, $1 per 1,000 content retrievals.

Who's already using it

Exa already powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com, and over 400,000 developers. The Cognition team, makers of the Devin coding agent, put it simply: "Exa is how Devin sees the web." That's not a marketing line , it describes a real architectural dependency. When Devin needs to look something up, it calls Exa.

Exa's enterprise client roster includes names like Databricks, AWS, Vercel, and Anysphere, the maker of AI coding tool Cursor, validating that demand is not speculative. The GTM story is also compelling: thousands of sales and marketing teams use Exa for CRM enrichment, lead discovery, and web monitoring , use cases that generate high, recurring query volume.

The market moment

The timing of this raise reflects a broader structural shift: AI agents are now expected to surpass humans in web search queries by volume in 2026, according to Exa's own announcement. In the next few years, the number of searches from LLMs will be 1000x more than Google searches today. That's not a rounding error , it's a complete inversion of who the customer for search infrastructure actually is.

The competitive landscape is consolidating fast. Tavily, purpose-built for agents with over one million developers, was acquired by AI cloud company Nebius in February 2026 for up to $400M, turning an independent infrastructure vendor into a proprietary component of a larger platform. Developers who do not want their search layer tied to a cloud provider now have one fewer independent option. Perplexity, valued at $20B after a $500M Series E in May 2025, occupies a different lane: its Sonar API bundles LLM synthesis with retrieval, which adds cost and reduces control for developers who already have a reasoning model and need only the underlying data.

The retrieval and grounding layer is emerging as a distinct infrastructure category separate from models, and this round is the clearest signal yet that investors are pricing it accordingly.

Who wins, who loses

For developers building agent stacks, Exa's traction is a signal that search infrastructure is not a commodity to be solved with a quick Bing API key. The sources describe three technical constraints that distinguish AI-targeted search from consumer search: latency requirements for real-time agents, high query volume because agents search far more than humans, and the need to surface long-tail, context-rich results rather than ranking optimized for clicks.

The losers in this story are the wrapper APIs , companies that resell Google or Bing results with a thin AI veneer. As Exa scales up infrastructure and model training, the gap between Exa and wrappers will become clearer. Six months ago Exa was worse than Google at code search; now they're used by nearly every coding agent. That trajectory is hard to argue with.

Google itself is a complex case. The round reflects a surge of investor interest in the future of online search, a market long dominated by Alphabet's Google. But Google's own pivot to AI-native search at I/O 2026 validates the premise Exa was built on. The risk for Google is that as it rebuilds search for AI, it may inadvertently commoditize the keyword-ranking moat it spent 25 years building , and companies like Exa are ready to fill the gap at the infrastructure layer.

What comes next

The $250M goes toward three things, according to Exa's announcement: training next-generation search and retrieval models, expanding infrastructure capacity to support hundreds of thousands of searches per second, and scaling its global go-to-market operations.

On the talent side, recent hires include the head of retrieval infrastructure from Meta, head of search backend at Yandex, and a research team out of Google. Marcus Holm, President of LaunchDarkly, is joining as Exa's CRO to lead the global GTM organization. The company is clearly shifting from pure research mode into a scaled commercial operation.

The deeper second-order question is what happens when search becomes a background primitive that no human ever directly touches. Exa's growth depends on the expansion of AI applications requiring web search capabilities. If foundation model providers integrate high-quality search directly into their platforms, or if AI applications increasingly rely on proprietary data, demand for independent search APIs could decline materially. That's the existential risk worth watching. For now though, with 5,000+ companies and 400,000+ developers already on the platform, Exa has enough momentum to make that a tomorrow problem.