The database that decided the relationships between your data were more interesting than the data itself.
[ FILED: SAN MATEO, CA ] - The graph people, photographed mid-query, still looking for that lost napkin.
Somewhere right now, a bank is watching money move. Not one account - thousands, looping through shell companies in patterns no spreadsheet would ever flag. The query that catches the fraud ring does not ask "who paid whom." It asks "how is everything connected." That question runs on Neo4j, and increasingly, so does the answer.
Neo4j is the company behind the most widely deployed graph database on the planet. It does one stubborn thing differently: instead of storing data in rows and tables and treating the links between them as an afterthought, it stores the links as first-class citizens. Nodes and relationships. A person, and the fact that they know someone, weighted and queryable, sitting right next to each other on disk.
Most databases file their relationships in the back room. Neo4j put them on the front page.
The practical result is that roughly 84% of the Fortune 100 keep a graph somewhere in their stack, and a community north of 250,000 developers reach for Cypher - Neo4j's query language - when the shape of the question is a network rather than a list.
The relational database, for all its fifty-year reign, has a quiet weakness. Ask it a simple human question - "who are the friends of my friends, three hops out?" - and it sweats. Each hop is another JOIN, and JOINs multiply. By the fifth degree of separation the query is crawling, the server is warm, and the database administrator is updating their resume.
This is not a bug you patch. It is a consequence of the model. Tables are good at answering questions about things. They are bad at answering questions about how things connect. And as it turns out, most of the interesting questions - fraud, recommendations, supply chains, who-influences-whom - are connection questions wearing a trench coat.
A relational database treats a relationship as a foreign key. A graph database treats it as the whole point.- the entire pitch, in two sentences
The origin story is, regrettably, true. On a flight to Mumbai in 2000, Emil Eifrem sketched the property graph model on a napkin. He has spent the rest of his professional life building the thing on that napkin, and apologizing that the napkin no longer exists.
We conceived the idea for the first property graph database during a flight to Mumbai in 2000 - we sketched it on a napkin, one that I wish I still had but, alas, has since disappeared.- Emil Eifrem, Co-Founder & CEO
Eifrem, with co-founders Johan Svensson and Peter Neubauer, turned that sketch into an open-source project in 2007. The "4j" originally meant "for Java" - a humble footnote that the marketing department has been quietly grateful to leave ambiguous ever since. The bet was simple and, at the time, lonely: that the world would eventually need a database built for relationships, and would pay for one that was good at them.
Open source was the wedge. Give the engine away, let 250,000 developers learn it on their own time, and sell the enterprise edition to the companies those developers work for. It is a patient business model. It also worked.
A napkin grows up
At the center is the Neo4j graph database itself, queried with Cypher. Cypher's trick is that it reads like ASCII art: (a)-[:KNOWS]->(b) is a query that literally looks like the pattern it is hunting for. Influential enough, in fact, that it became the foundation of GQL - the first new ISO query-language standard since SQL in 1987.
Around that core, the catalog fans out for people who would rather not run a database themselves.
Fully managed graph database-as-a-service on AWS, Azure and Google Cloud. The database, minus the 3 a.m. pager.
65+ graph algorithms - centrality, community detection, pathfinding - plus machine learning on connected data.
Knowledge-graph patterns and integrations with LangChain and LlamaIndex that ground LLM answers in verified facts.
Run Neo4j's algorithms directly on data sitting in Snowflake or Databricks - no ETL into a separate store.
The query language looks like the diagram on the whiteboard. That is not a coincidence - it is the entire user-experience strategy.
The customer list reads like a roll call of organizations with complicated connections to untangle: NASA, UBS, Walmart, Klarna, IBM, Merck, EY, Daimler, Dun & Bradstreet. Some catch fraud. Some build recommendations. Some map supply chains that, post-2020, nobody takes for granted anymore.
USD, selected rounds · ~$565M total
Bars scaled to the largest disclosed round. "Earlier rounds" aggregates pre-2021 financing; figures approximate, drawn from public reporting.
The 2021 Series F - $325 million, led by Eurazeo and Alphabet's GV - was, at the time, described as the largest investment in database history. The interesting part came after the headline: in the three years that followed, the company doubled its recurring revenue and crossed $200 million, then took a further $50 million in 2024 that left the roughly $2 billion valuation intact.
Raising the biggest round in your industry is a press release. Doubling revenue afterward is a business.
For most of its life Neo4j sold a slightly academic idea: that relationships matter. Then generative AI arrived with a very public flaw - it makes things up, fluently and with great confidence. Suddenly "relationships matter" had a product name: GraphRAG.
The pitch is that a knowledge graph gives a language model a factual scaffold to stand on. Instead of guessing, the model retrieves real, connected facts - entities and the verified links between them - and reasons over those. The result is meant to be more accurate, more transparent, and more auditable. For a regulated bank or a hospital, "the AI can show its work" is not a nice-to-have. It is the difference between deployment and a meeting with legal.
The competition is real - Amazon Neptune, TigerGraph, ArangoDB, Memgraph, and the relational and vector databases that keep insisting they can do this too. Graph remains a category that still has to explain itself at dinner parties. But the trend lines that matter are bending toward connections: AI that needs grounding, fraud that hides in networks, supply chains that everyone suddenly wants to map, regulations that demand an audit trail.
Twenty-five years ago, "the relationships are the point" was a napkin sketch. Now it is most of the Fortune 100's homework.
So return to the bank, still watching money move. The fraud ring loops through its shell companies, confident that no single transaction looks wrong - and it is right. No single transaction does. But the shape of the whole thing, the pattern across thousands of hops, lights up like a constellation the moment someone asks the graph the right question. The data was always connected. Neo4j's contribution was deciding, in 2000, on a napkin, that the connections deserved a database of their own.