A distributed graph database built for the kind of data that refuses to fit in rows. Open source. Apache 2.0. Built for trillions.
Photographed in pixels: the NebulaGraph mark, mid-orbit, somewhere between a particle accelerator diagram and a constellation map. Suitable for framing.
Somewhere inside a payment processor in Shenzhen, a query fires. It asks a deceptively simple question: who is this person connected to, and are any of those connections suspicious? The answer comes back in single-digit milliseconds. The user, none the wiser, completes their checkout. The graph that made that possible has roughly a trillion edges, and it runs on NebulaGraph.
This is the company's posture in 2026: largely invisible, deeply load-bearing. NebulaGraph - the open-source distributed graph database from Vesoft - is the kind of software you only notice when it stops working, which is rarely. It is used by Tencent, Meituan, JD Digits and Kuaishou. It also runs in dozens of less famous places where someone needed to ask a graph a hard question and could not wait for the answer.
For decades, the answer to "who knows whom" was a SQL join. Then another join. Then a third. Engineers became fluent in the art of pretending a network was a stack of tables, which works tolerably well until your dataset crosses some invisible line and the joins start measuring their runtime in coffee breaks.
The founding team at Vesoft - alumni from Meta, Alibaba, Huawei and IBM - had watched this happen often enough to be irritated by it. Graph problems were everywhere: fraud rings, social feeds, recommendation engines, supply chains, knowledge graphs. The infrastructure to answer them was either expensive, proprietary, hard to scale, or all three. Open source had not yet produced a serious contender that could handle the kind of scale a Chinese super-app demands. So they wrote one.
Caption: A graph problem, in plain English. "Find me all the accounts that share a device with an account that flagged for fraud in the last 90 days." In SQL, three joins and a prayer. In a graph database, one traversal.
Sherman Ye founded Vesoft in October 2018 with a small group of distributed-systems engineers who had collectively spent years tuning storage engines for other people's products. The thesis, as far as one can read it from the outside, was straightforward: graph databases were going mainstream, and the dominant options were not going to scale gracefully into the next decade. There was room - actually, there was urgent need - for a horizontally scalable, open-source alternative.
It was a bet that required patience. Open-sourcing infrastructure is a slow way to make money, in the same way planting an orchard is a slow way to make pies. The first commits landed in 2018. The first public release went on GitHub in May 2019 under the Apache 2.0 license. NebulaGraph 1.0 GA followed in June. NebulaGraph 2.0 GA arrived in March 2021. Each release was less an announcement than a hand-wave to the community already using it.
NebulaGraph is, at its core, a distributed storage engine wrapped around a query language. The engine separates compute from storage - a design choice that sounds dull until your graph outgrows a single machine, at which point it sounds like a religion. The query layer speaks nGQL, the native language, and increasingly supports OpenCypher, so engineers who learned graphs on Neo4j can move without rewriting their muscle memory.
Around the core sits the rest of the surface area: NebulaGraph Studio for browsing schemas and importing data, NebulaGraph Explorer for visual traversal, NebulaGraph Cloud for the people who would prefer not to think about Kubernetes, and an Enterprise edition that adds disaster recovery, role-based access control and the kind of support contracts that compliance teams require.
It runs on AWS and Azure. It plays well with Ubuntu and Nginx and Varnish. It exposes SDKs in enough languages that the question "does it support X" usually has a boring yes for an answer.
It is one thing to claim scale. It is another thing to be the database Tencent quietly reaches for when its product teams need to traverse a social graph that does not stop growing. Meituan uses NebulaGraph for the kind of recommendation logic that decides which restaurant appears at the top of your scroll. JD Digits and Kuaishou are in the customer list too. Most of these companies do not love giving public testimonials about their infrastructure. The fact that NebulaGraph appears in their stack at all is a quieter sort of endorsement.
Beyond the headline accounts, NebulaGraph has the kind of long tail open-source projects either earn or do not: government agencies, financial-services firms, knowledge-graph teams inside larger enterprises, and a growing population of developers building graph-augmented AI - retrieval-augmented generation that uses an actual graph to remember what is related to what. The phrase "graph RAG" did not exist when Vesoft started. Now it appears in the company's own keyword list, which is a polite way of saying the world caught up.
Surface fraud rings in real time by walking shared-attribute edges, not joining tables.
Power "people you may know" and "things you may want" at the scale of a national app.
Anchor LLM retrieval against a graph that actually knows how concepts relate.
Answer relationship questions in milliseconds, even when the dataset has trillions of edges.
Listen to the company long enough and a fairly modest ambition emerges. Vesoft does not particularly want graph databases to be exotic. The mission is closer to the opposite: to make them the kind of thing a backend engineer reaches for without ceremony, the way they would reach for Postgres or Redis. That requires three things to be true at once - the database must be open, the database must scale, and the database must not surprise you in production. NebulaGraph spends most of its release notes on the third.
The company joined the Linked Data Benchmark Council, which is the somewhat-Belgian organization responsible for keeping the graph database industry honest about performance numbers. It is the sort of move that gets noticed by exactly the right people and nobody else.
The funny thing about the AI boom is that it has quietly turned graphs into a load-bearing primitive. Retrieval-augmented generation works best when the retrieval has structure. Agents work best when they can ask the world a relationship question. Recommendation, fraud, drug discovery, supply chains - the headline use cases of applied AI in 2026 are all, at the bottom, graph problems wearing different costumes.
This is the bet NebulaGraph and Vesoft sit on top of, and it has aged well. The product is in production at scale. The license is permissive. The team is small enough to ship and large enough to support. The competitors are formidable but not obviously winning. The world keeps producing more connected data than it knows what to do with, and someone has to query it before lunch.
Back to the payment processor in Shenzhen. The query fires. The graph answers in milliseconds. The user finishes their checkout. Nothing dramatic happens, which is the whole point. Somewhere in the rack, NebulaGraph waits for the next question.
Official sites, source code, and video walkthroughs.