CHAPTERS
Exa’s $85M Series B and the company’s core mission
Nicolas Dessaigne opens with Exa’s Series B announcement and frames the conversation around what Exa is trying to achieve. Will Bryk positions the company as tackling a foundational shift: the web now needs to be searchable by AI systems, not just humans.
- •$85M Series B at a $700M valuation sets the context and urgency
- •Exa’s mission: organize the world’s knowledge “for real”
- •Premise: AI is a new “user” of the web with different needs than humans
- •Exa is positioned as infrastructure rather than a consumer destination
What it means to build “search for AIs” (not search for humans)
Will explains how optimizing for AI changes the entire search engine design space. Instead of short keyword queries and UI-driven link lists, AI systems issue complex prompts, scan large result sets, and require customization and quality.
- •AI queries can be long, specific, and structured—search must support that
- •AIs can evaluate far more results; returning large result sets matters
- •Quality and relevance outweigh UI and ad-driven ranking incentives
- •Customization knobs (filters, domain scopes, result volume) are essential for AI customers
Exa’s under-the-hood business model: search infrastructure for companies
Rather than competing as a consumer search portal, Exa sells to enterprises and startups that embed AI into products. Exa becomes the backend web-search layer powering other applications users interact with daily.
- •Exa serves companies building AI products rather than end consumers
- •Search happens “under the hood” inside other apps and agents
- •This model enables product choices Google/Bing wouldn’t prioritize
- •Explains why fewer visible competitors exist in this specific niche
Two product modes: fast API search vs. slow, high-quality “Websites”
Exa provides a spectrum of latency options similar to model tiers in AI (fast vs. deliberative). Fast search is synchronous, while “Websites” can take minutes (or longer) to produce deeper, higher-quality results.
- •Fast search: synchronous, optimized for low latency
- •“Websites”: slower pipeline aimed at maximum quality and depth
- •AI workloads require multiple latency profiles depending on task
- •Agents and backends can trade time for improved result quality
Origins before ChatGPT: building a better Google for ‘nerds’
Exa began in 2021, inspired by GPT-3’s deeper language understanding and frustration with stagnant, SEO-heavy search. The early product targeted humans who want high-quality knowledge without ads and SEO noise.
- •Founded ~4 years prior to the interview; before ChatGPT
- •Motivation: GPT-3 showed “deep understanding” while Google felt unchanged
- •Goal: reduce SEO/ad influence and elevate knowledge quality
- •Target user archetype: “nerds” seeking precise, high-signal information
The post-ChatGPT pivot: realizing the product fit for AI systems
After launching and getting attention on Twitter, ChatGPT’s release changed the landscape and triggered inbound API demand. Will describes the ‘click’ moment: what they built for knowledge-hungry humans was ideal for AIs, so they rapidly built and priced an API.
- •ChatGPT shifted demand toward programmatic search for AI apps
- •Early API requests were initially dismissed, then recognized as signal
- •Key insight: “nerds and AIs” both optimize for high-quality knowledge
- •Fast pivot: API became a wrapper over the existing engine plus pricing
- •Being first to say “search engine for AI” felt contrarian at the time
Contrarian early bets and YC lessons: deep tech first, users later
Will highlights decisions that went against conventional startup playbooks, including heavy investment in compute and extended R&D before intense user feedback loops. The team focused on first-principles research to create a new kind of search engine.
- •Early plan: raise after YC and spend heavily on a GPU cluster
- •Belief in the “bitter lesson” applied to search: scale compute + learning
- •Long R&D phase (~1.5 years) with less emphasis on user interviews
- •Strategy: build foundational capability first, then productize for demand
Owning the full stack: crawling the web instead of wrapping Google
Exa chose to crawl and index the web itself, enabling control, customization, and enterprise-grade guarantees. Will argues that many customer requirements (domain scoping, large result counts, zero retention) require owning the entire search stack.
- •Full-stack ownership enables deep customization for customers
- •Enterprise use cases: limit search to specific domain sets
- •Need to return far more than typical consumer search result caps
- •Compliance/security: true zero data retention requires independence
- •“Control your destiny” is positioned as a key strategic advantage
Why Exa trains its own search model (and why off-the-shelf embeddings fail)
Will explains that generic open-source embedding models don’t work well at web scale and chaos. Exa trains its own model and continuously looks for ways to add compute to improve search quality across billions of pages.
- •Open-source text generation is decent; open-source search models lag
- •Web search differs: massive scale, noisy/chaotic documents, long tail queries
- •Off-the-shelf embeddings degrade when moving from millions to web-scale docs
- •Exa’s approach: continuous training and compute scaling for search-specific performance
Infrastructure scale: GPU clusters, web indexing, and reliability goals
Will describes Exa’s growing compute footprint and how the Series B will expand research velocity, crawling, and processing. He frames web scale as increasingly “manageable” and focuses on scaling quality and reliability rather than indexing everything immediately.
- •Current hardware: ~$5M GPU cluster (144 H200s, 18 nodes)
- •Series B funds: more compute, faster research cycles, broader crawling
- •Pragmatism: you don’t necessarily need to index the entire web at once
- •Scaling challenge includes reliability and production readiness for enterprise use
Why LLMs still need search: edge knowledge and a changing web
Will argues from information theory and practical limitations that LLM weights can’t contain the full, evolving web. Search remains essential for fresh facts and long-tail (“edge”) knowledge that is often most valuable.
- •LLMs compress data into weights; they can’t “memorize” the entire web
- •Long-tail facts and niche knowledge often aren’t captured reliably
- •The web changes constantly; static model snapshots go stale
- •Search complements LLMs by retrieving up-to-date, high-coverage knowledge
Evaluating search quality: no standard benchmarks, building in-house evals
The conversation turns to measurement: search evaluation is hard and lacks standardized benchmarks. Exa is developing diverse internal evals and plans to publish methodologies to help establish a broader standard for AI-search evaluation.
- •Search evals are difficult; no single metric captures all needs
- •No widely accepted benchmark exists for Exa’s AI-focused web search
- •Exa builds a suite of internal evals with “secret sauce” components
- •Intent to publish papers/benchmarks to drive industry-wide progress
Agents as the primary users: designing for 100 searches per request
Will predicts most AI products will evolve toward agentic systems, changing performance requirements. Agents may run many searches per task, making speed and latency reductions multiplicatively important.
- •Market maturity varies: single LLM calls today → more agents over time
- •Agents enable longer, asynchronous workflows but increase search volume
- •Design target: an agent doing ~100 searches per user request
- •Latency improvements (1s → 100ms) can reduce end-to-end agent time dramatically
- •Exa’s philosophy: build for the world that’s coming, not today’s snapshot
The post-Google search world: ‘perfect information’ as a civilizational upgrade
Will paints a vision of search becoming frictionless and comprehensive—removing information blockers across industries. He uses recruiting and sales as examples where perfect, continuously updated knowledge would transform coordination and decision-making.
- •Goal: immediate access to needed information with minimal blockers
- •Recruiting as search: find all best-fit candidates, instantly
- •Sales as search: discover all relevant companies, including new ones
- •Societal framing: we’ll look back amazed we operated without perfect info
- •Exa claims to already support parts of these use cases, aiming for “perfect”
Team building, creative hiring, and founder mindset in a high-stakes market
Will describes the kind of people Exa wants—high-hunger builders across engineering, research, and go-to-market—and shares a creative puzzle-poster recruiting tactic. He closes with lessons on enduring constant problems, focusing on execution over competitors, and staying motivated by mission and “the game.”
- •Hiring focus: intelligence, hunger, engineers to scale to trillions of pages
- •Researchers to run experiments, create training data, and build evals
- •Go-to-market roles due to a large and expanding market
- •Puzzle posters generated ~100 applicants and at least one hire
- •Founder lesson: constant fires are normal; startups are sequences of problems
- •Competitive mindset: don’t over-worry—execution speed and velocity matter
