Recent Parallel and Exa funding

Why the hype with Parallel and Exa

Both companies are betting on the same macro thesis: AI agents need a fundamentally different web infrastructure than humans do — and neither Google nor legacy search was built for them. The capital raised reflects investor conviction that whoever wins this layer could become as foundational to AI as Google was to the early web.

How did they build the infrastructure?

Exa: Embedding-Based Neural Search

Exa raised an $85M Series B led by Benchmark at a $700M valuation, with participation from Lightspeed, Y Combinator, and NVentures, NVIDIA's venture capital arm.

Exa's software can run more than 100 queries per second with a latency of under 450 milliseconds. Developers can customize what data is retrieved from just URLs to full webpage content.

Exa also has side products like Websets: retrieval and data augmentation for GTM like Apollo.

Exa threw out the keyword/inverted-index model entirely. Unlike other search engines that preprocess documents into keywords, Exa trains neural models to preprocess each document into embeddings. Every webpage becomes a high-dimensional vector representing its meaning, not its words. Exa uses the same technology as ChatGPT to train an embedding model to convert web page content into vector representations. Exa's model training data includes shared links on web pages rather than simple text and sentences, this allows its search engine to better understand and predict the relevance of web page links. Exa's search engine doesn't just predict the next word, but the next related link. It's a rethinking of PageRank: instead of counting links as votes, they use the semantic relationship between linking and linked content as a training signal.

Exa exposes multiple search types: neural (pure embeddings), auto (intelligently combines neural and keyword), fast (streamlined), and deep (agentic: searches, processes results, and searches again until it finds high-quality information). The deep mode is essentially a mini agent loop inside the search itself.

Parallel: Excerpts for Agent Context Windows

Founded by former Twitter CEO Parag Agrawal, Parallel has raised $230 million in total, most recently a $100M Series B led by Sequoia Capital at a $2 billion valuation.

Parallel's approach is less about the embedding model and more about rethinking the output format for how agents consume information.Parallel is built so agents can specify declarative semantic objectives and it returns URLs and compressed excerpts based on token relevancy.

They're not returning a webpage for a human to read, but a compressed, relevance-ranked snippet designed to fit efficiently into an LLM's context window. Less noise, fewer tokens consumed, cheaper inference. When an AI agent conducts legal research, it needs not only search results but verifiable sources that can be cited and audited. For financial analysis, it requires up-to-date data rather than indexed snapshots. General-purpose search engines don't meet these needs and produce error rates unacceptable for professional use.

Parallel's differentiators over Exa in this framing are: index freshness, provenance tracking at scale (verifiable citations for regulated industries), and the publisher revenue-sharing model.

Future prediction

Both companies are racing to own the data retrieval layer of AI infrastructure — a position that becomes more valuable the more AI agents proliferate. Neither Google (optimized for human ad clicks) nor existing RAG/scraping tools (fragile, slow, unstructured) are good enough.

However, I’m not bullish because of the following reasons:

1. The horizontal nature of this market attracts bigger players to come in soon, unlike more vertical applications that big model providers won't encroach (such as GTM companies like Clay)
2. none of this technology really has a “hard tech moat”, unlike harder tasks like coding, and big companies with access to web data such as Google can easily reproduce it - they just haven't gotten around to it because they were busy with the model race.

The alternative

The fundamental problem of searching and ranking a large billions corpus is the same for traditional or agent search. Given none of this technology is new, traditional search still has an advantage if adopted to an agentic format since it already has a scalability infra headstart, e.g. layered retrieval using the same source corpora and traditional search at the bottom, and use embedding agentic search in the last layers .

I predict Google will still be the dominant player in the future.