CHAPTERS
- 0:05 – 0:19
Exa’s $85M Series B: what it enables and why it matters
Nicolas welcomes Will Bryk and opens with Exa’s $85M Series B at a $700M valuation. The conversation frames the raise as fuel for a long-term bet: rebuilding search for a world where AI systems are primary “users.”
- •Exa announces a major Series B: $85M at $700M valuation
- •Positioning: this is about building foundational search tech, not a small feature
- •Sets up the core theme: search is changing because AI needs it
- 0:19 – 1:18
A search engine built for AI (not humans)
Will explains Exa’s core idea: traditional search engines were optimized for human UI workflows, while AIs behave differently and demand different system design. Exa is built to handle richer queries, many more results, and deeper configurability.
- •AI search behavior differs from human keyword-and-click workflows
- •Need to support complex, paragraph-length queries and higher recall
- •AI can scan massive result sets; quality and filtering become central
- •Search must be customizable for the company deploying the AI
- 1:18 – 1:52
Infrastructure, not a consumer destination: how Exa fits into products
Rather than competing as a consumer search portal, Exa sells to companies via an API layer that powers AI features inside other applications. The end user interacts with an AI product; Exa operates behind the scenes.
- •Exa serves enterprises/startups building AI into their products
- •Positioned as “under-the-hood” search infrastructure
- •Value comes from reliability, quality, and integration—not branding or UI
- •Different business model than Google/Bing consumer search
- 1:52 – 3:15
Why few companies do this—and Exa’s two-speed product strategy
Will argues the space is sparse because building search from scratch is extremely hard and most teams avoid it. He then describes Exa’s spectrum of latency products: fast synchronous search vs a slower, highest-quality mode.
- •Search-engine-from-scratch is a multi-year R&D challenge
- •Infra business model enables choices incumbents won’t make
- •Two main offerings: fast synchronous search and slower “Websites” mode
- •AI workloads require a range of latency/quality tradeoffs
- 3:15 – 4:34
Before ChatGPT: trying to build a better Google for “nerds”
Exa started before ChatGPT, inspired by GPT-3’s ability to understand rich text while Google felt stagnant. The team initially built for humans who care about high-quality knowledge—minimizing SEO and ads.
- •Origin point: GPT-3 made deep query understanding feel possible
- •Goal: higher-quality knowledge retrieval without SEO/ads incentives
- •Built for demanding human users (“nerds”) who want precision and depth
- •This foundation later mapped well to AI needs
- 4:34 – 6:04
The pivot: realizing ‘search engine for AIs’ and shipping an API
After ChatGPT’s release, Exa began receiving API access requests and gradually realized the product fit was stronger for AI systems. Once the phrase “search engine for AIs” clicked, the company quickly built and commercialized an API wrapper.
- •ChatGPT shifted attention and created demand for programmatic search
- •Early API requests were initially dismissed as non-core
- •Key insight moment: explicitly framing Exa as AI-first search
- •Rapid execution: API + pricing layered onto existing search tech
- 6:04 – 7:20
Contrarian early bets: GPUs, research-first execution, and YC lessons
Will describes Exa as deliberately contrarian, including the decision to invest heavily in compute early and pursue a research-heavy path. The team prioritized building a new system over classic early-stage playbooks like frequent user interviews.
- •Applied the “bitter lesson” mindset: scale compute to win
- •Plan to spend a large share of funding on a GPU cluster was unusual
- •Long build phase focused on foundational tech over immediate feedback loops
- •Takeaway: some non-standard YC paths can work when the problem demands it
- 7:20 – 8:15
Owning the full stack: crawling the web and controlling customization
Exa chose to crawl and index the web itself rather than wrapping existing search engines. Will argues full-stack ownership is essential for enterprise needs like domain-level filtering, high result counts, and strict data retention guarantees.
- •Full-stack control enables deep customization (e.g., search only specific domains)
- •Higher recall (e.g., thousands of results) requires owning indexing/ranking
- •Enterprise requirements: true zero data retention and policy guarantees
- •Strategic benefit: controlling destiny rather than depending on incumbents
- 8:15 – 10:59
Training Exa’s own search model + compute scale + the evals problem
Will explains why off-the-shelf embedding models fail at web-scale retrieval and why Exa trains its own model. He details their dedicated GPU cluster and highlights a major challenge: search evaluation lacks standardized benchmarks, forcing Exa to build in-house evals and plan publications.
- •Off-the-shelf models don’t handle chaotic, massive web corpora well
- •Exa continuously looks to ‘pour more compute’ into improving retrieval
- •Dedicated hardware: a multi-million dollar GPU cluster (144 H200s)
- •LLMs can’t memorize the whole web; search remains necessary and dynamic
- •No standard search evals exist for this; Exa is building and intends to publish benchmarks
- 10:59 – 12:19
Agentic customers and latency at scale: ‘100 searches per request’
As customers evolve from single LLM calls to full agents, Exa anticipates much higher search volume per user request. That pushes product priorities toward speed and reliability, since small latency reductions compound across many tool calls.
- •Market trend: companies shifting toward agentic systems over time
- •Agents often operate asynchronously and can run many searches per task
- •Latency becomes a multiplier (e.g., 100 searches makes speed critical)
- •Exa designs for the future state of AI products, not today’s snapshot
- 12:19 – 13:07
What the Series B will fund: scaling research, crawling, and hiring
Will explains the rationale for raising $85M now: scaling every layer of the system to accelerate progress. The focus is increasing compute for faster R&D, expanding crawling/processing, and recruiting top talent.
- •Capital goes to scaling GPU compute for faster iteration and research
- •Scaling crawling and web processing to broaden/refresh coverage
- •Pragmatic approach: you don’t necessarily need to index the entire web
- •Hiring is a core use of funds to build the best team possible
- 13:07 – 18:39
The broader mission: ‘organizing the world’s knowledge for real’ + founder mindset
Will outlines a five-year vision where information blockers (recruiting, sales discovery, coordination) largely disappear because search becomes dramatically more complete and timely. He closes with founder reflections: startups are constant fires, motivation comes from mission and the ‘game,’ and competitors matter less than execution speed.
- •Vision: instant, comprehensive access to knowledge—no ‘wish I knew that earlier’ moments
- •Concrete examples: recruiting and sales as solvable information problems
- •Hiring philosophy: prioritize intelligence and hunger; creative recruiting tactics (puzzle posters)
- •Founder lesson: challenges never stop; progress is solving a sequence of hard problems
- •Competitive stance: focus on velocity and execution rather than fear of incumbents
