Back in 2021, two Harvard graduates — Will Bryk and Jeffrey Wang — made what seemed a somewhat forward-thinking call at the time: AI needed its own dedicated search engine. ChatGPT had not yet launched, and large language models were still little more than academic curiosities confined to research circles. They invested $5 million to build a GPU cluster and built a search engine from the ground up, one designed to let AI truly comprehend the entire web.
This vision was validated two years later. Early 2023 saw an explosion of AI applications, all of which required pulling real-time information from the internet. Yet Google Search was built for human users: humans want a list of ten blue hyperlinks. AI systems have entirely different requirements: they demand full, high-quality page content free of SEO spam, exhibit extreme sensitivity to latency, and require zero persistent data retention.
That’s when Exa officially launched the first web search API built exclusively for AI workloads.
It’s Not Just a Wrapper for Third-Party Search Engines
Most search API providers act as mere middlemen, simply wrapping existing public search engines under a thin abstraction layer. Exa took a fundamentally different path: it spent multiple years building a full-scale web search engine from scratch.
Its crawler indexes over 500 billion unique URLs. Its research team trains custom embedding models on proprietary GPU hardware, and it built a purpose-built vector database optimized for the high-concurrency workloads demanded by AI agents. Exa itself puts it succinctly: “There are more space programs than independent web search engines — and there’s good reason for that.”
Building the entire stack in-house grants full, granular control over three critical variables: output quality, latency, and operational cost. Six months ago, Exa lagged behind Google for code-specific search use cases; today, nearly all mainstream coding AI agents rely on Exa’s APIs.
How It Enables AI to Truly “Understand” the Web
Exa’s core innovation is neural semantic search. Instead of matching raw keyword strings, it interprets the underlying semantic intent behind each query.
For example, a keyword-based search for “fintech companies in Switzerland” will only return pages that contain those exact words. Exa grasps your underlying intent: you want a curated list of Swiss fintech firms, returning only genuinely relevant results rather than generic pages stuffed with matching keywords.
On top of its semantic search backbone, Exa built an exceptionally intelligent text extraction model named Highlights. This model extracts the most contextually relevant paragraphs from any webpage in under 100 milliseconds. Benchmarked on the SimpleQA dataset, a 500-character Highlights summary achieves the same accuracy as parsing the full 8,000-character original page, while consuming just 1/16 of the total tokens.
This means AI models can generate equally accurate answers while processing far less raw text — a massive direct cost reduction for all LLM applications priced by token consumption.
One Unified API, Four Distinct Search Modes
Exa exposes a single unified search API, with four internally optimized search modes tailored to different latency/quality tradeoffs:
- Fast: The world’s fastest search API, with P50 latency under 350ms. Ideal for latency-critical use cases including AI voice assistants and real-time live chatbots.
- Auto: Balances latency and result quality, set as the default search mode.
- Deep: Runs agentic iterative search — executing follow-up lookups and re-searching until it surfaces the highest-quality information; P50 latency hovers around 3.5 seconds. Optimized for deep research and complex multi-layered queries.
- Deep-Reasoning: Deep search augmented with native logical reasoning capabilities, built for queries requiring multi-step logical inference.
Beyond core search, Exa also exposes supplementary specialized APIs:
- Contents API: Fetch full cleaned raw webpage content
- Monitors API: Track real-time web content updates and new pages
- Agent API: Run asynchronous long-running deep research tasks
Who Uses Exa?
Exa now serves over 5,000 enterprise clients and 400,000 independent developers. Its customer roster includes:
- Cursor: Leverages Exa to pull up-to-date documentation and repository code data, delivering far more accurate AI code generation
- Cognition, OpenRouter, HubSpot, Monday.com, and dozens more industry leaders
Its total monthly query volume surged tenfold year-over-year, jumping from roughly 100 million requests in April 2025 to 1 billion requests by April 2026.
Funding & Team Background
In May 2026, Exa announced a $250 million Series C funding round led by a16z, pushing its post-money valuation to $2.2 billion. Prior financing rounds included a $22 million Series A (2024) and an $85 million Series B (2025).
Team Profile
Both co-founders are Gen Z born in the 1990s: CEO Will Bryk (29) and Jeffrey Wang (28). The core team totals 26 full-time staff, 7 of whom are of Chinese descent — including Hubert Yuan, a graduate of Tsinghua University’s prestigious Yao Class. The company has recently added senior engineering leaders: former head of retrieval infrastructure at Meta, former search backend lead at Yandex, and an entire research team poached from Google.
Straightforward Core Thesis
At its core, Exa is answering a foundational industry question: What should search look like when AI Agents become the primary consumers of web information?
Google Search was built for human end users: humans scan ten links, click through pages, and manually judge relevance themselves. AI systems do not need hyperlinks — they require pre-extracted, structured information ready for immediate reasoning. Additionally, AI agents will generate search volumes orders of magnitude higher than human users: over the next several years, LLM-driven search traffic is projected to hit 1,000 times the scale of today’s total Google search volume.
This unprecedented scale demands an entirely redesigned underlying search infrastructure — and Exa aims to become that foundational backbone.