Every AI Agent Needs to Search the Web. Exa Just Raised $250 Million to Be the One They Call.

At a glance: Exa Labs. Founded 2021, San Francisco. YC-backed. CEO Will Bryk (co-founded with Dan McArdle and Jeffrey Wang). Product: neural search API built for AI agents and developer pipelines, not human end-users. Funding: $250 million Series C, announced May 20, 2026, led by Andreessen Horowitz. Valuation: $2.2 billion (post-money). Previous round: $85 million at $700 million valuation (fall 2025 — valuation tripled in roughly six months). Total raised: ~$361 million. Customers: 5,000+ companies including Cursor, Cognition, HubSpot, OpenRouter, and Monday.com; 400,000+ developers. Team: 121 employees as of January 2026. Part of our AI Models & Companies reviews.

Google was not built for AI agents. Neither was Bing, nor any traditional search engine. They were built for human beings — people who type a phrase into a box, scan a list of blue links, and click the ones that look promising. Every design decision, every ranking signal, every snippet of text displayed on the results page, was optimized for that human-mediated workflow.

AI agents don’t work that way. When a coding assistant like Cursor searches the web for documentation, or when an AI research tool looks up background on a company, or when an autonomous workflow agent retrieves live data to make a decision, none of those agents are reading a list of blue links. They’re consuming structured data programmatically — often hundreds of queries per session, at speeds no human could match. They need different things: precise API access, content extracted from pages rather than links to pages, understanding of what a query means rather than what words it contains.

This is the gap Exa was built to fill. On May 20, 2026, the company raised $250 million at a $2.2 billion valuation — a round that tripled its valuation from six months earlier — on the thesis that search infrastructure for AI agents is a fundamentally different problem from search for humans, and that solving it well is worth a great deal of money.

The investors leading the round are Andreessen Horowitz, with Nvidia and Thrive Capital also participating. The bet they’re making is that every AI product needs web search, and that the search layer will be worth as much as any other piece of AI infrastructure.

The Keyword Problem

Traditional search engines — Google, Bing, and every search API built on their indexes — use some version of the same basic architecture: inverted indexes that map words to documents. When you search for “startup building developer tools for LLM observability,” a keyword-based search engine finds pages containing those exact terms and ranks them by signals like backlink authority and click-through rate.

This works adequately for humans who can evaluate what they get back, click through a few results, and reformulate their query if needed. It works poorly for AI agents.

An AI agent running a research workflow might need to find companies working on AI debugging tools without knowing that the documents describing those companies use the phrase “observability” instead of “debugging.” Or it might need to find recent analysis of a CEO’s leadership style from sources that never say “leadership style” explicitly. Keyword search surfaces documents containing the right words; it may or may not surface documents containing the right ideas.

Exa’s approach is different. Instead of building an inverted keyword index, Exa trains neural embedding models — transformer-based, the same class of architecture underlying large language models — that convert web pages into high-dimensional vectors representing semantic meaning. A query isn’t matched against words; it’s matched against meaning. Documents that express the same ideas with different vocabulary surface anyway.

The company describes itself as building “the first web-scale neural search engine.” The claim is specific: search filtered by meaning, not by keyword, at the scale of the full web index. The technical challenge isn’t the embedding model itself — those are well-understood — it’s running them at web scale in real time.

What the API Actually Delivers

Exa’s product is not a search engine in the consumer sense. There is no search box to type into, no results page to browse. It’s an API — a set of endpoints that developers call from within their applications, agents, or data pipelines.

The core offering has several layers:

Neural search: The foundational query endpoint, where semantic embedding matching replaces keyword lookup. Ask for companies working on drug discovery using AI, and Exa returns relevant companies even when the documents describing them don’t use those exact phrases.

Contents: Exa returns not just URLs but the actual text content of matched web pages. For an AI agent building a research workflow, this is the difference between a list of links (useless) and source material the model can reason over (useful).

Query-dependent highlights: Rather than returning full page text, Exa can extract the specific passages most relevant to a query. The company reports this lets AI applications fit four to five times more sources into the same token budget — a meaningful efficiency gain for applications running up against context window limits.

Exa Instant: A latency-optimized variant for applications where search speed is the constraint.

Exa Agent: A multi-step search workflow that handles iterative research tasks autonomously rather than requiring the application to chain individual queries.

The cumulative effect is a search API designed from the ground up for consumption by software, not by people — where the output format, the relevance model, and the latency profile are all calibrated for programmatic use in AI pipelines.

The Traction

Exa was founded in 2021 and went through Y Combinator. The early years were quiet — the company was building infrastructure, not selling a consumer product, and AI agent development was still nascent. The explosion in agent-based AI development in 2024 and 2025 changed that.

By May 2026, Exa has more than 5,000 company customers and over 400,000 developers using its API. The customer list is a who’s who of AI development: Cursor (the AI coding tool that reached $2B ARR by February 2026), Cognition (makers of Devin, the AI software engineer), HubSpot, OpenRouter, and Monday.com. These are not experimental accounts — these are production integrations inside tools that millions of people use daily.

The fall 2025 funding round valued Exa at $700 million on $85 million raised. Six months later, the valuation is $2.2 billion on a new $250 million raise. That trajectory — valuation tripling in roughly half a year — reflects something real in the growth numbers, even if Exa has not disclosed specific revenue figures publicly.

The $12 million revenue figure that has circulated in some databases appears to be significantly outdated. Companies with 400,000+ developer accounts and 5,000+ enterprise customers whose products collectively serve tens of millions of end users are generating substantially more than that in API revenue. The most reliable proxy is valuation, and $2.2 billion at Andreessen Horowitz’s expected return multiples implies Exa’s revenue run rate is well into eight figures and growing.

The Team

Will Bryk is the CEO. He grew up in New York City, studied computer science and physics at Harvard (where he also did machine learning research), and worked as a software engineer at Cresta, an AI startup building real-time coaching tools for customer service agents. He left Cresta to start Exa when, as he has described it, he concluded that the quality of information people and systems can access shapes the quality of decisions civilization makes — and that search was the wrong tool for the coming era.

Co-founders Dan McArdle and Jeffrey Wang round out the founding team. The company has grown to 121 employees as of early 2026, with the new $250 million available to accelerate hiring.

Andreessen Horowitz has been watching AI infrastructure closely — they backed Cursor, which is Exa’s largest public customer. The same fund investing in a search API and a coding tool that depends on that search API is a coherent thesis: a16z is building exposure to the full AI developer toolchain, from model inference to code editing to the search infrastructure that ties it together.

The Competitive Landscape

Exa’s nearest competitor in the developer-facing AI search API category is Tavily. Both target roughly the same customer: developers building AI agents and pipelines who need programmatic web search. Tavily took a different architectural approach — closer to a managed search service with curated outputs rather than Exa’s full neural index — but competes for the same API wallet share.

Perplexity operates at the consumer/prosumer level and has published an API, but its primary business is answering questions for human users, not serving as infrastructure for other AI products. You.com has a similar consumer-first orientation. Brave Search offers a developer API focused on privacy and independence from Google’s index rather than semantic embedding quality.

On a published benchmark evaluation of 500 real-world queries conducted by Thinking Machines, Exa scored 64.8% versus Perplexity’s 60.1% — a meaningful gap in a category where fractions of a point of relevance quality accumulate into substantially better AI product outputs over millions of queries.

The broader threat comes from the large platforms. Google has invested heavily in its Gemini API, which includes access to Google Search results. If Google decides to aggressively compete for the AI search API market — pricing down its search API, matching the neural relevance features Exa has spent years building — it would be competing with enormous structural advantages: the most comprehensive web index in the world, the lowest latency at scale, and a brand that every developer already trusts.

Exa’s response to this risk is implicit in its positioning: Google’s index was built to serve human search. Exa’s was built to serve AI. Whether that architectural distinction holds as an advantage over time, or whether Google converges, is a genuine open question.

What $250 Million Buys

The announced use of funds centers on two things: training the next generation of embedding models, and scaling infrastructure to support hundreds of thousands of searches per second.

The compute requirements for training a new embedding model on web-scale data are substantial — on the order of what a mid-size model training run costs, multiplied by the need to do it iteratively as the web changes. Exa’s embedding models are not a one-time artifact; they need to be retrained and updated as new content appears and as the company learns which semantic distinctions matter most for its customers.

Infrastructure for hundreds of thousands of searches per second — the publicly stated target — is a significant step up from current operating levels. AI products at Cursor’s scale (millions of active users, each potentially triggering multiple search calls per session) generate search volumes that stress any infrastructure not designed for that load. Exa is investing in the capacity to serve the AI products of 2027 and 2028, not just 2026.

The hiring component — 121 employees is still small for a company targeting this scale — will likely focus on ML research (model quality), infrastructure engineering (scale and latency), and enterprise sales (the 5,000-company customer base is early relative to the addressable market).

The Honest Assessment

The funding round and the valuation trajectory are real signals. Three billion dollars of implied market cap on a company that most people outside AI development have never heard of reflects genuine market validation: the people building AI products at scale — Cursor, Cognition, the others in Exa’s customer list — chose Exa over Google’s API, over Perplexity’s API, over Tavily, and are paying for it in production at meaningful volumes.

The risks are real too. Google’s structural advantages in web indexing are not going away. The AI search API market is early enough that pricing pressure from well-capitalized incumbents is a constant possibility. And the neural embedding advantage that justifies Exa’s premium positioning requires continued research investment just to maintain — the gap can close if competitors train better models or adopt similar architectures.

The more interesting question, which won’t be answered until the AI agent market matures further, is whether “search built for AI agents” is a distinct market with durable pricing power, or a feature that gets absorbed into the broader model provider ecosystem. If OpenAI, Anthropic, and Google each eventually offer native search integration that’s good enough for their respective model ecosystems, the independent search API category could compress significantly.

Exa’s bet is that the answer is “distinct and durable.” The $250M Series C is the market betting alongside them.

The Bottom Line

Exa Labs is building infrastructure that most users of AI products will never see or think about — but that underlies the quality of information those products can access. The growth from $700 million to $2.2 billion in six months, at a time when AI agent development is moving from experimental to production at thousands of companies, is a credible sign that the need is real and Exa is meeting it.

For developers building AI products: Exa’s search API is worth evaluating seriously, especially if your application needs to understand meaning rather than just match keywords. We’ve covered the Exa MCP server separately for those integrating via Model Context Protocol.

For observers of the AI infrastructure space: this funding round is part of a broader pattern. The tools that AI products depend on — search, data pipelines, memory stores, evaluation frameworks — are becoming significant businesses in their own right, as the explosion in AI product development creates sustained demand for every layer of the stack below the model. Exa is one of the cleaner bets on that layer.

The search engine was never designed for this moment. Exa was.