Metaphor Systems (rebranded to Exa in January 2024)

Metaphor Systems, now operating as Exa, is an AI-powered search engine company that fundamentally reimagines how information is discovered and retrieved on the internet. Unlike traditional search engines built for human users, Exa is purpose-built for artificial intelligence systems—specifically large language models (LLMs) that power modern AI applications. The company provides a developer API that enables AI agents to search the web with unprecedented accuracy, speed, and contextual understanding.

At its core, Exa leverages neural search technology based on transformer models—the same architecture that powers systems like GPT-3 and Stable Diffusion. Instead of relying on keyword matching like Google, Exa's AI model predicts the most relevant links based on semantic understanding of queries. This approach allows developers to build AI applications that can access real-time, high-quality information from across the web, powering use cases from AI chatbots to enterprise research tools.

The company's mission is to organize the world's knowledge and make it accessible to AI systems. Exa envisions a future where AI agents can autonomously navigate the internet to find, process, and synthesize information—essentially serving as the "Google for the AI era." With backing from top-tier investors including Benchmark, Lightspeed, and NVIDIA, and adoption by leading companies like OpenAI, Cursor, Notion, and Databricks, Exa is positioning itself as critical infrastructure for the AI revolution.

Will Bryk — Co-Founder and CEO

Will Bryk grew up in New York City and studied Computer Science and Physics at Harvard University, where he conducted machine learning research and led the Harvard Undergraduate Robotics Club. Before founding Metaphor, he spent two years as a software engineer at Cresta (a conversational AI startup), where he was one of the first engineers and built real-time AI products. During his time at Cresta, Will also wrote a book about the history of civilization in his spare time. He realized that the future of civilization depends almost entirely on the quality of information we consume, which inspired him to leave Cresta in June 2021 to start Metaphor. Will is known for his deep technical expertise in embedding models and his creative problem-solving abilities.

Jeff Wang — Co-Founder

Jeff Wang studied Computer Science and Philosophy at Harvard University, where he became friends and collaborators with Will Bryk. After graduating, Jeff spent three years at Plaid, where he built data and web infrastructure and became known for his incredible engineering velocity and ability to find product-market fit for new projects. Jeff is recognized for his expertise in large-scale systems engineering and his ability to ship high-quality code rapidly. The co-founders were roommates at Harvard and worked together late into the night hacking on problem sets and technical projects, establishing the foundation for their future collaboration.

Note: There was initial confusion with Alex Gajewski, who was mentioned in some sources. However, Alex Gajewski is the co-founder of SF Compute, a separate company that provides GPU compute infrastructure. Alex previously worked on Metaphor's infrastructure during its early days but is not a co-founder of Metaphor/Exa.

The story of Metaphor begins in 2021, when Will Bryk and Jeff Wang—friends and collaborators from Harvard—witnessed the launch of GPT-3 and had a crucial insight. While Google had remained relatively unchanged for a decade, GPT-3 felt like a breakthrough that could actually understand the subtleties of human language. You could enter a paragraph into GPT-3, and it would understand exactly what you meant. Google, in contrast, relied on basic keyword algorithms and struggled with complex requests.

This observation led to a compelling question: what if we could make a search engine that felt as smart as GPT-3? The founders realized that pretraining (for large language models) and indexing (for search engines) feel remarkably similar—both involve code looking at all the text on the internet and trying to compress it into a better representation. GPT-3 itself wasn't a search engine, but it made them think about what a GPT-3-shaped search engine would look like.

The breakthrough came when they developed a new self-supervised learning objective called "next link prediction"—similar to how GPT-3 predicts the next word in a sequence, Metaphor's AI model would predict the next link. This approach requires self-supervised learning to generate essentially infinite training data, which is what makes generative models so powerful. Will left his position at Cresta in June 2021, and together with Jeff, they bought their own GPU cluster and spent approximately eight months iterating on model architectures. The "aha moment" was realizing they could train a neural network to predict what link would naturally follow a given text prompt, based on how people actually share links on the internet. They launched the first public beta in November 2022, entering Y Combinator's Winter 2022 batch to accelerate their progress.

Traditional search engines like Google were designed for humans in the late 1990s, optimized for displaying blue links that users would click. However, the emergence of AI agents and large language models has created a fundamentally different set of requirements. AI systems don't need visually appealing search result pages—they need structured, high-quality data that can be processed programmatically at scale. Google's keyword-based approach, while effective for human users, falls short when AI agents need to understand context, nuance, and semantic meaning.

The problem is particularly acute for developers building AI applications. When you search Google for "companies in SF building futuristic hardware," the results are dominated by SEO-optimized articles designed to attract human clicks, not the actual list of companies you're looking for. Google's algorithm prioritizes pages that game its ranking system rather than pages with the most relevant information. For AI applications that need to ground their responses in real-world data, this approach is fundamentally broken. The issue becomes even more severe when AI agents need to perform thousands of searches per second, requiring sub-500-millisecond latency and structured data formats that Google simply wasn't built to provide.

Moreover, the internet contains the collective knowledge output of mankind—millions of essays, hundreds of millions of research papers, billions of images and videos, and trillions of ideas across tweets, forums, and blog posts. Searching the internet should feel like navigating a grand library of knowledge, but instead, it feels like navigating a landfill. Knowledge on the internet is buried under an overwhelming amount of information. Many enterprises and developers need search capabilities that can cut through the noise to find specific, high-quality information—whether that's identifying qualified sales leads, finding relevant research papers, conducting competitive analysis, or discovering niche companies based on specific criteria.

The timing is critical: as AI models become increasingly sophisticated and autonomous, they need reliable access to up-to-date information from the web. While AI models are trained on vast amounts of data, they cannot perfectly memorize all information on the public web, and there's always a "last mile" of data—private data, alternative data sources, real-time information—that will never end up in training datasets. Without proper search capabilities, AI agents remain fundamentally limited in their ability to assist with complex, real-world tasks.

"Soon, AI will search the web more than humans. But search engines like Google were designed for humans, not AI. Whereas Google is optimized for human clicks, AI needs a search engine that's powerful and precise enough to retrieve thousands of results with the best information. That's where Exa comes in – we're the first search engine built for AI." — Will Bryk, CEO of Exa

Exa's solution is a complete reimagining of web search built from the ground up for AI systems. At its core, Exa trains embedding models using the same transformer technology behind ChatGPT to convert web pages into lists of numbers known as embeddings. These embeddings capture the semantic meaning of content, not just keywords. The result is a technology that packs the power of large language models into the search process itself, making search dramatically smarter than traditional keyword approaches.

The company's core product is a developer API that offers multiple search modes to fit different use cases. Neural search understands the meaning behind queries to deliver semantically relevant results. Keyword search provides traditional exact-match capabilities when needed. Auto-hybrid intelligently combines both approaches. The "Exa Fast" mode delivers results in under 350 milliseconds—making it the fastest search API in the world—while "Exa Deep" mode uses agentic search to find the highest quality information across the web, even if it takes longer. The API can return anywhere from 10 to 10,000 links with metadata, supporting real-time applications for AI agents.

Beyond basic search, Exa has developed several specialized products:

Websets: A powerful research agent that can build massive, structured datasets from the web. Companies use Websets to find qualified sales leads, identify research papers, conduct competitor analysis, or discover companies based on highly specific criteria (e.g., "700+ employee companies with seat-based pricing models that announced expansion into LATAM in the past 6 months"). Websets can crawl websites, blogs, press releases, LinkedIn pages, and news articles to find signals that would never appear in traditional databases.

Contents Endpoint: Retrieves raw text, highlights, or summaries from URLs, formatted specifically for LLM consumption. This eliminates the need for web scraping and provides clean, full-text content from any indexed page.

Highlights Feature: Instantly extracts any webpage content from each search result using a customizable embedding model, allowing developers to pull specific information without processing entire pages.

Answer Generation: Produces cited responses similar to Perplexity AI, with sources linked for verification.

Research Endpoint: Coordinates multi-agent systems to generate detailed research reports by breaking down complex questions into sub-queries.

What makes Exa fundamentally different is its infrastructure. The team built their own web crawling infrastructure for ingesting information from across the web, their own neural network models trained specifically for link prediction, and their own vector database written in Rust and designed to scale to billions of documents at low latency. This vertical integration allows Exa to offer capabilities that companies using Google's API or building on top of traditional search simply cannot match.

The company also addresses enterprise concerns with a Zero-Data-Retention (ZDR) mode, where all queries and data are automatically purged based on customer requirements. This is crucial for regulated industries handling sensitive information. Exa maintains SOC2 certification and provides comprehensive data processing agreements (DPAs), service level agreements (SLAs), and high-capacity rate limits for enterprise customers.

While Exa doesn't publicly disclose all technical details, based on available information, here's what is known about their technology architecture:

AI/ML:

Custom transformer-based foundation models trained on link prediction Neural embedding models for semantic search Contrastive learning approaches for training on billions of web pages Multiple model variants: Exa Fast (optimized for speed), Exa Deep (optimized for quality) Infrastructure:

Custom web crawling infrastructure for large-scale data ingestion Proprietary vector database written in Rust GPU clusters: 80 A100s initially, upgraded to 144 H200s ("Exacluster") Distributed systems architecture for handling billions of documents Backend:

Rust for performance-critical components (vector database) Likely Python for ML model training and serving Real-time indexing system Sub-425ms latency optimization for Fast mode Database:

Custom vector database designed for neural search Optimized for semantic similarity search at scale Supports billions of document embeddings API:

RESTful API for developers Multiple endpoints: search, contents, answer, research, websets Supports 100+ queries per second Sub-450 millisecond latency on average Key Integrations:

OpenAI API compatibility for seamless integration Claude Desktop integration Supports integration with various LLM frameworks MCP (Model Context Protocol) server for AI agents Deployment:

Cloud-based infrastructure Enterprise-grade reliability with SLAs Global availability Zero-data-retention mode available The company's technical approach is distinguished by its vertical integration—rather than relying on third-party crawling services or vector databases, Exa built every component from scratch to optimize for AI-first use cases.

Target Market Size:

The market opportunity for Exa sits at the intersection of several rapidly expanding sectors:

• Neural Network Market: Expected to grow from $45.43 billion in 2025 to $537.81 billion by 2034, at a CAGR of 31.60%. Neural search engines are a specialized subset driving significant growth within this market.

• Quantum-Enhanced Neural Search Engine Market: Projected to increase from $1.41 billion in 2024 to $1.80 billion in 2025 (27.4% CAGR), reaching $4.69 billion by 2029 (27.0% CAGR).

• AI Search Engine Market: Valued at $17.3 billion in 2024, projected to reach $73.7 billion by 2034, advancing at a CAGR of 15.6% between 2025 and 2034.

• General Search Engine Market: Expected to reach $252.5 billion in 2025 and grow to $440.60 billion by 2030 at a CAGR of 11.80%.

Exa's total addressable market (TAM) is substantial as they're targeting:

1. Every developer building AI applications that need web search capabilities
2. Enterprise companies implementing AI agents for research, sales, and operations
3. The emerging market of autonomous AI systems that will search the web more than humans

Current Traction:

While specific user and revenue numbers are not publicly disclosed, available data indicates strong growth:

• Thousands of developers and companies have adopted Exa's API
• Notable Enterprise Customers: OpenAI (uses Exa as default search tool in open-source gpt-oss models), Cursor (for technical web search), Notion (powers Notion AI), Databricks, Amazon Web Services, Vercel, and top private equity and consulting firms
• Team Size: Approximately 22 employees as of 2025
• Valuation Growth: From undisclosed seed valuation to $70 million (Series A, July 2024) to $700 million (Series B, September 2025) — a 10x increase in just over one year

Growth Metrics:

The company's growth trajectory suggests rapid adoption:

• Rebranded from Metaphor to Exa in January 2024 to better reflect their mission
• Raised $22M Series A in July 2024
• Raised $85M Series B in September 2025 (14 months later)
• 10x valuation increase within approximately 14 months
• Expanding from 7 employees at the time of rebrand (January 2024) to 22+ employees (2025)

Market Position:

Exa is strategically positioned in several ways:

1. First-mover advantage in AI-native search infrastructure
2. Technical differentiation through custom-built technology stack (crawling, models, vector database)
3. Performance leadership: Sub-350ms latency makes Exa Fast the fastest search API in the world
4. Quality leadership: Exa Deep provides higher quality results than competitors in benchmarks
5. Enterprise credibility: SOC2 certification, zero-data-retention, and adoption by major tech companies

Market Trends Supporting Growth:

• Explosive growth in AI agent development and deployment
• Increasing need for real-time information in LLM applications
• Enterprise adoption of AI tools requiring reliable search infrastructure
• Shift from human-first to AI-first internet services
• Growing recognition that AI systems need specialized infrastructure

The market timing appears optimal, as the company was founded in 2021—before the AI boom accelerated with ChatGPT in late 2022—giving them a significant head start in building AI-native search infrastructure.

Exa represents a fundamental shift in how information is organized and accessed on the internet. For decades, search has been optimized for human users clicking on blue links. But as AI systems become more sophisticated and autonomous, the infrastructure of the internet must evolve. Exa is building the critical data layer that will enable AI agents to become truly useful—allowing them to access, process, and synthesize information from across the web in real-time. This is not an incremental improvement; it's a paradigm shift in internet infrastructure.

The company's significance extends beyond technology. Exa is addressing what Will Bryk identified as one of civilization's most important challenges: the quality of information we consume. In an era of information overload, SEO spam, and AI-generated content flooding the web, having search systems that can accurately identify high-quality, relevant information becomes crucial. By organizing the world's knowledge in a way that both humans and AI can access effectively, Exa is positioning itself to be as foundational to the AI era as Google was to the internet era. Peter Fenton from Benchmark—who has taken seven companies to IPO including Elastic, the last major search company built—sees this potential clearly, choosing to lead Exa's Series B and join their board.

Moreover, Exa's success validates a critical thesis: the future of AI requires specialized infrastructure built from the ground up for AI systems, not retrofitted from human-centric tools. As more companies deploy AI agents for tasks ranging from customer service to scientific research, the demand for AI-native search will only accelerate. With major players like OpenAI, Cursor, and Notion already relying on Exa, the company is proving that specialized AI infrastructure can capture significant market share from incumbents. This matters because it demonstrates that startups with the right technical vision can still build foundational internet infrastructure, even in domains seemingly dominated by giants like Google.

Competition from Tech Giants: Exa faces formidable competition from well-resourced incumbents. Google is actively integrating AI into its search products with Gemini. Microsoft has Bing powered by OpenAI's models. Meta is reportedly building its own AI search engine. Perplexity AI has raised significant funding and achieved strong brand recognition in consumer AI search. OpenAI could potentially build its own search infrastructure. These competitors have massive distribution advantages, existing user bases, and virtually unlimited resources. While Exa's technical differentiation is strong, maintaining this edge as giants invest billions in AI search will be an ongoing challenge.

Scaling and Infrastructure Costs: Operating a search engine at internet scale is extraordinarily expensive. Exa must continuously crawl billions of web pages, train and serve large neural network models, maintain GPU clusters, and provide sub-500-millisecond latency globally. The operational costs are significantly higher than traditional search engines due to the computational intensity of neural models. While the company has raised $107 million, managing burn rate while scaling infrastructure to meet growing demand will be critical. The economics of AI search—with higher per-query costs than traditional search—requires either premium pricing (which may limit adoption) or achieving massive scale to make unit economics work.

Technical and Scaling Challenges: As Exa grows, several technical challenges loom. Maintaining search quality while scaling the index to cover more of the internet is non-trivial. Keeping information fresh requires continuous re-crawling and re-indexing. Handling adversarial content, spam, and malicious websites while maintaining accuracy becomes more difficult at scale. Latency guarantees become harder to maintain under high query volumes. The company must also navigate complex web technologies, handle sites that block crawlers, and deal with the increasing fragmentation of the internet across walled gardens and paywalls. Each of these challenges requires significant engineering resources and ongoing innovation.

Market and Business Model Risks: Exa's business model is primarily B2B, charging per query or subscription. This creates dependencies on developer and enterprise adoption, which can be slow and require significant sales and support infrastructure. The company must balance building for developers (who drive adoption) with serving enterprise customers (who provide stable revenue). There's also the risk of commoditization—if the technology becomes table stakes, larger players with more resources could replicate it. Finally, Exa must navigate potential regulatory challenges around data privacy, web scraping, copyright, and AI-generated content, which could impose additional constraints or costs on their business model.

Exa's publicly stated roadmap centers on achieving "perfect search over all the world's information" and enabling every company to harness it. In the near term, the company is focused on expanding its team—actively hiring ML research engineers, full stack engineers, and sales engineers—to build what they call "the Google for the new AI world." The $85 million Series B funding will be used to accelerate development of their core search engine, expand the Websets product for enterprise use cases, and potentially move beyond text to multimodal search capabilities (images, videos, audio).

The company's vision suggests several potential expansion areas. First, deeper vertical integration into specific industries—specialized search engines optimized for healthcare, legal, finance, or scientific research. Second, expanding from search API to a broader AI infrastructure platform, potentially including data enrichment, entity extraction, and knowledge graph services. Third, developing more sophisticated agentic capabilities where Exa's systems can autonomously orchestrate complex research tasks involving multiple queries, synthesis, and reporting. The company has hinted at building search that goes beyond "what's on the internet today" to include historical data, alternative data sources, and potentially real-time event streams.

Long-term, Exa is positioning itself to be the essential data layer for all AI applications. As AI agents become more sophisticated and autonomous—potentially handling everything from personal assistance to business operations to scientific research—they will need reliable, high-quality access to information. Exa aims to be the infrastructure that makes this possible. The company's trajectory suggests they could either remain an independent infrastructure provider (analogous to how AWS provides compute infrastructure) or become an acquisition target for major AI platforms. With Peter Fenton's guidance—having taken companies like Elastic public—the company appears to be building toward a potential IPO in the coming years. The key question is whether they can maintain their technical edge and achieve the scale necessary to become the definitive AI search platform before competition intensifies.

Startup: Metaphor Systems (now Exa)

Batch: YC W21 (Summer 2021)

Founded: 2021

Problem: Traditional search engines like Google were built for humans, not AI systems. AI agents need search that understands semantic meaning, delivers structured data at scale, and provides real-time information—capabilities Google's keyword-based approach cannot provide.

Solution: AI-native search engine built from scratch using neural embeddings and transformer models. Provides developer API with sub-350ms latency, semantic search, web crawling, and enterprise features like zero-data-retention. Powers AI applications at companies like OpenAI, Cursor, Notion, and Databricks.

Why It Matters: Building critical infrastructure for the AI era—the "Google for AI agents." Represents fundamental shift from human-first to AI-first internet services. Backed by top investors with proven track record in infrastructure companies.

Funding: $107M total (Seed + $22M Series A in July 2024 + $85M Series B in September 2025)

Valuation: $700 million (Series B, September 2025)

Stage: Series B / Growth Stage

Traction: Thousands of developers using API; enterprise customers include OpenAI, Cursor, Notion, Databricks, AWS, Vercel; 22 employees; 10x valuation increase in 14 months

Founders: Will Bryk (CEO, ex-Cresta engineer, Harvard CS/Physics) and Jeff Wang (co-founder, ex-Plaid, Harvard CS/Philosophy)