TigerGraph is a Native Parallel Graph (NPG) database, queried through GSQL, its own SQL-like, Turing-complete language, and the kind of storage-and-traversal engine a governed context layer like Atlan’s Enterprise Data Graph still has to sit above. TigerGraph no longer publishes throughput figures on its product page. The one performance statement it does make there is comparative and unquantified: in its own benchmark tests, it loads in an hour what it says another system takes a day to load. Below: what that architecture actually does, what TigerGraph’s Community Edition and Savanna platform each cost, and one fully sourced production account.
Most explainers of “what TigerGraph is used for,” including TigerGraph’s own site, cover the architecture and GSQL mechanics thoroughly, then repeat a customer-logo list, JPMorgan Chase, Ford, Intuit, UnitedHealth, without published depth behind any of them. Several also blur the free Community Edition into “open source,” which it is not. Grounding the use-case claims in the one production account with a full case study and a named quote, stating the licensing model plainly, and naming the question TigerGraph’s own marketing never poses, what a connection in the graph means to the business, and who is allowed to query it, is the gap worth closing here.
- Real architecture, not a marketing claim: a Native Parallel Graph engine split into a storage half and a processing half, both written in C++, running an MPP/BSP model that scales each half independently across a cluster
- Real production evidence: Jaguar Land Rover’s supply-chain graph, with a named quote and real before-and-after numbers, not an unlinked customer-logo list
- Real licensing clarity: Community Edition is free, including for production, and TigerGraph claims no open-source licence for it, despite how some secondary sources describe it
- The gap worth naming: TigerGraph’s own architecture answers what’s connected to what; a context graph still has to decide what a connection means to the business and who’s allowed to query it
- A related but different question: what a knowledge graph is and what a graph database like TigerGraph stores are related, but not identical, questions
Below: what TigerGraph is used for, how its architecture works, what Savanna and Community Edition each cost, real production use grounded in one fully sourced account, whether it can power GraphRAG, how it compares to Neo4j, and where a context layer fits above the graph.
| Field | Detail |
|---|---|
| What it is | A Native Parallel Graph (NPG) database queried in GSQL, with most of openCypher embedded inside GSQL and the path-pattern subset of ISO GQL |
| Architecture | Graph Storage Engine (GSE) plus Graph Processing Engine (GPE) on an MPP/BSP model, both implemented in C++ |
| Current platform | TigerGraph Savanna, launched January 2025; cloud-native, independent storage/compute scaling, nine prebuilt solution kits |
| Best for | Fraud detection, supply-chain analysis, recommendation engines, Customer 360, entity resolution; storage and traversal, not semantic reasoning. The query-language docs cover GSQL, openCypher and the GQL subset; SPARQL and OWL appear in none of them |
| Licensing / cost | Community Edition: free, production-eligible, capped at 300GB of combined graph and vector data and 16 CPUs, single-server, not licensed as open source. Savanna: $45 per GB/month standard, $126 Business Critical, storage billed separately at $0.025 per GB/month |
| Named production use | Jaguar Land Rover (sourced case study, named quote); JPMorgan Chase, Ford, Intuit, UnitedHealth cited in press and conference materials, without published case-study depth |
What is TigerGraph used for?
TigerGraph is a Native Parallel Graph database built for high-volume traversal, pattern matching, and graph machine learning, queried through GSQL, its own SQL-like, Turing-complete language. TigerGraph’s own description of the design is narrower than the superlative usually attached to it: the Native Parallel Graph “focuses on both storage and computation, supporting real-time graph updates and offering built-in parallel computation.” That points at the real architectural choice underneath, parallelizing graph storage and traversal across a distributed cluster instead of running them on a single, tightly coupled engine.
At a marketing level, TigerGraph frames its use cases around fraud detection, supply-chain analysis, recommendation engines, Customer 360, and cybersecurity, according to TigerGraph’s own use-case blog. Those categories hold up; what most explainers skip is which of them TigerGraph can actually document with a real customer, a real before-and-after number, and a name attached, the distinction the Jaguar Land Rover account below makes concrete.
It’s also worth being precise about what TigerGraph is not. TigerGraph’s query-language documentation covers GSQL, openCypher, and the path-pattern subset of ISO GQL. SPARQL and OWL appear in none of it, and TigerGraph does not offer ontology reasoning. That’s a scope statement, not a manufactured weakness: TigerGraph is a graph analytics engine, and a fast one, but storage and traversal are a different question from what a knowledge graph adds on top of a graph database, a gap a governed context layer is built to close.
How does TigerGraph’s architecture work?
TigerGraph’s speed claims trace to a specific architectural choice, not a general claim to take on faith: a Native Parallel Graph engine split into two components, running a bulk-synchronous, massively parallel processing (MPP/BSP) model across a distributed cluster.
The two components are the Graph Storage Engine (GSE), which handles native, disk-based storage of vertices and edges, and the Graph Processing Engine (GPE), which handles in-memory computation and traversal. Both are implemented in C++, according to TigerGraph’s own internal-architecture documentation. Splitting storage from processing is what lets TigerGraph scale each independently across machines in a cluster, rather than bottlenecking on a single engine doing both jobs at once.
TigerGraph no longer publishes throughput figures on its product page. The one performance statement it does make there is comparative and unquantified: in its own benchmark tests, it loads in an hour what it says another system takes a day to load. Anyone sizing a cluster should ask TigerGraph for numbers against their own data rather than working from figures that circulate without a live source behind them.
The architecture isn’t just marketing copy. Its technical grounding traces back to “TigerGraph: A Native MPP Graph Database”, a paper by TigerGraph’s own founder and colleagues, Alin Deutsch, Yu Xu, Mingxi Wu, and Victor Lee, which describes the Native Parallel Graph design in more detail than the marketing does.
GSQL and accumulators
GSQL is TigerGraph’s query language: SQL-like in syntax, Turing-complete, and built around accumulators, a mechanism that lets a query aggregate values across a traversal path without the join explosion a relational engine hits on the same multi-hop query. That mechanism is the real “why” behind Jaguar Land Rover’s production speedup, covered further below. GSQL is also no longer the only way in. TigerGraph supports most openCypher features embedded within GSQL, and the path-pattern-matching subset of the ISO GQL standard, so the proprietary-language objection that used to attach to TigerGraph has narrowed. Teams building a graph from scratch, rather than evaluating GSQL syntax in the abstract, are better served starting with how to build a knowledge graph for AI agents, which treats TigerGraph as one storage option among several.
A fast storage-and-traversal engine still leaves open what those vertices and edges mean to the business, and who’s allowed to query them. That’s the distinction knowledge graph construction for AI draws between building a graph and governing what gets built into it, and the gap graph database versus metadata layer names directly.
| Layer | What it decides | Who delivers it |
|---|---|---|
| Layer 1: Storage and traversal | What’s connected to what | TigerGraph’s GSE plus GPE, queried in GSQL |
| Layer 2: Semantic meaning | What a connection means to the business | Business definitions and ontology |
| Layer 3: Governed context | Who, or which agent, can query it | Atlan’s Enterprise Data Graph, delivered via MCP |
TigerGraph’s Native Parallel Graph engine handles Layer 1 only. Semantic meaning and governed, agent-safe delivery are separate layers built on top of it, not features TigerGraph’s own architecture attempts to provide.
What is TigerGraph Savanna, and how does it differ from the Community Edition?
TigerGraph ships two very different things under one brand name: a free, production-eligible Community Edition, and Savanna, its 2025 cloud-native commercial platform, and conflating the two is the most common licensing mistake in third-party coverage.
Community Edition is free, including for production use, capped at 300GB of combined graph and vector data, 16 CPUs, and a single server with no clustering, according to TigerGraph’s Community Edition page. What most secondary sources get wrong is calling it open source. TigerGraph does not license the Community Edition as open source, and claims no OSS licence for the product itself. The Open Source Notice on its site covers only the third-party open-source components bundled with TigerGraph, not the product’s own licensing, a distinction no third-party explainer surveyed here states cleanly.
Savanna, launched January 21, 2025, is the cloud-native commercial platform above it: independently scaling storage and compute, and nine prebuilt solution kits covering transaction fraud, application fraud, product recommendation, mule account detection, entity resolution (including a know-your-customer variant), Customer 360, supply-chain management, and network infrastructure, according to TigerGraph’s Savanna page. The speed claim TigerGraph attaches to Savanna is that it ingests data and runs queries “up to 100x faster than other graph platforms,” with no comparison basis stated, so read it as a vendor claim rather than a measurement.
Savanna’s pricing is published, which is unusual enough in this category to be worth saying: a Free Trial tier at $0, Savanna at $45 per GB/month for 24-7 service in US and Tier 1 regions, and Savanna Business Critical at $126 per GB/month, per TigerGraph’s pricing page. Storage is a separate line item at $0.025 per GB/month, and hourly instances run from $1.00 for two CPUs to $256.00 for 512. The service rate and the storage rate are added together on the same bill, not alternatives to each other.
The ROI case for the paid tier rests on a Forrester Total Economic Impact study TigerGraph commissioned in 2022. TigerGraph’s own press page for it no longer resolves; the URL now serves the homepage, so the figures cannot be checked against a live official source.
Who owns TigerGraph has changed since most coverage of it was written. In July 2025 the company took an undisclosed strategic investment from Cuadrilla Capital, and it is now led by CEO Rajeev Shrivastava rather than founder Yu Xu. TigerGraph’s newsroom has published nothing since that announcement, though its engineering blog and docs stay current into 2026.
One brand, two very different tiers
See how enterprises are governing context consistently across every graph, warehouse, and lakehouse a team runs, whatever licensing tier sits underneath.
Get the Context Layer EbookWhat are TigerGraph’s real-world use cases?
One TigerGraph production account has a fully sourced, named case study with real before-and-after numbers, worth leading with rather than an unlinked customer-logo list: Jaguar Land Rover.
Jaguar Land Rover integrated 12 data sources, equivalent to 23 relational tables, into a TigerGraph-powered supply-chain graph, cutting queries that previously took weeks, “if possible at all,” down to about 45 minutes, according to TigerGraph’s Jaguar Land Rover case study. The case study carries no publication date and references the COVID-19 pandemic in the present tense, so treat it as several years old. Harry Powell, Director of Data & Analytics at Jaguar Land Rover, described what that speedup looked like in practice: “We used the graph to re-sequence how our vehicle orders were to be built in our factory in response to a supplier failure. A process which in the past might have taken days was both modelled and evaluated in less time than it took to write the PowerPoint slide to present the idea.”
JLR’s story is one instance of the broader categories TigerGraph markets: fraud detection, supply-chain analysis, recommendation engines, Customer 360, and entity resolution, per TigerGraph’s use-case blog. Beyond JLR, JPMorgan Chase, Ford, Intuit, and UnitedHealth all appear as TigerGraph customers, but that evidence comes from press releases and conference-speaker announcements, not published case-study depth. Readers evaluating enterprise knowledge graph pitfalls before committing to a vendor will recognize the pattern: a named logo is not the same claim as a documented result.
Not sure how governed your graph data actually is?
Run a quick assessment on your current context layer before scaling agent workloads on top of any graph engine.
Calculate Your Context GapCan TigerGraph power GraphRAG and AI agent workflows?
Yes, and TigerGraph has built more of the surface than most coverage of it reflects. Its AI-assistant surface has been renamed at least once and is currently marketed as GraphRAG, under an “AI & Graph Intelligence” heading on TigerGraph’s own site. CoPilot, the previous name, still has live documentation but no longer appears in the product navigation, and TigerGraph has published no deprecation notice for it.
Two capabilities matter more than the naming. TigerGraph shipped vector search in version 4.2, first in alpha in March 2025 and in its first LTS release that September: a VECTOR attribute type, an HNSW index built automatically on load, approximate-nearest-neighbour similarity search, and hybrid graph-plus-vector queries written in GSQL. And TigerGraph publishes an MCP server, installable with pip install tigergraph-mcp, so an agent framework can reach the graph over the same protocol it uses for everything else. The caveat worth attaching: that server lives in a DevLabs GitHub organisation and has no entry in the product documentation.
Both capabilities are real, and neither answers the question above them. An MCP server exposes the graph in an instance; it does not carry what a connection means to the business, who owns the upstream asset, or which agent should be allowed to see it. A natural-language-to-GSQL feature makes the query language more approachable. It is a developer-productivity feature, not a policy one.
Readers looking at the broader GraphRAG and agent-memory landscape beyond TigerGraph specifically will find fuller treatment at what is GraphRAG, at what agent memory actually needs to persist, and at how MCP delivers business context to whatever agent framework a team runs.
How does TigerGraph compare to Neo4j?
TigerGraph and Neo4j are the two graph databases most often compared head-to-head, and most of the public comparison data available comes from vendors with a stake in the answer.
The benchmark figures circulating about TigerGraph and Neo4j almost all originate with TigerGraph. Its own benchmark page reports loading 1.8x to 58x faster and 2-hop path queries 40x to 337x faster than the systems it tested, from a 2018 report it ran itself. Competitor comparison posts restate those numbers; they do not reproduce them. TigerGraph’s own LDBC runs carry its own disclaimer: “The benchmarks on this page are not official LDBC benchmark results, as they have not been audited.”
The architectural difference underneath the benchmarks is the more durable point: TigerGraph’s Native Parallel Graph engine runs an MPP/BSP model built for distributed traversal across a cluster, while Neo4j uses index-free adjacency on a more tightly coupled engine. That’s a mechanism-level distinction, not a marketing scorecard, and it’s the reason the two products perform differently on different query shapes rather than one being categorically faster.
A full TigerGraph-versus-Atlan comparison for data governance is a separate question with its own dedicated answer. The third-party comparison readers are actually searching for is the one above; from here, knowledge graph tools compared covers the wider landscape, TigerGraph alongside Neo4j, Neptune, Stardog, and GraphDB, and Neo4j GraphRAG vs. LlamaIndex vs. LangChain Graph Transformer covers the construction-library side of Neo4j’s own GraphRAG stack. Teams evaluating Stardog or Ontotext GraphDB as semantic-first alternatives, or FalkorDB as a lighter-weight GraphRAG option, are asking a related but separate question again.
How Atlan approaches context above TigerGraph’s graph layer
TigerGraph sits at Layer 1 of the stack above: storage and traversal, done well. It is strong for fraud detection, recommendation engines, and supply-chain analysis. It is not a semantic knowledge graph tool, and its own query-language documentation is the evidence for that: GSQL, openCypher, and the GQL subset, with no SPARQL and no OWL anywhere in it. That is a scope statement about a fast graph analytics engine, not a knock on one.
What sits above Layer 1 doesn’t change with the storage engine underneath it, Neo4j, Neptune, TigerGraph, or none at all. Atlan’s Enterprise Data Graph adds business definitions, lineage, ownership, and policy context on top of that engine, the extension a metadata knowledge graph makes to a raw graph database and the semantic layer TigerGraph’s own architecture doesn’t attempt to provide. Atlan’s MCP server pattern then delivers that governed context to AI agents no matter which graph engine sits underneath, the infrastructure question behind how to implement an enterprise context layer for AI and what context engineering means in practice.
No customer has published a case study naming TigerGraph specifically alongside Atlan, so the argument here rests on the mechanism, not a named account. Governed context means something concrete: which team, or which agent, is allowed to see a given connection, and what that connection means in business terms rather than only in graph-traversal terms, the governance question that separates an active knowledge graph from a static one. A graph engine’s query language, GSQL included, doesn’t answer either question on its own, a distinction also covered in why AI agents need an enterprise context layer and in why AI agents fail in production once that governance layer goes missing. Teams weighing vector store versus graph database for agent memory, or comparing AI memory, RAG, and knowledge graphs more broadly, land on the same answer regardless of the storage pattern: what governs retrieval matters more than what stores it. AI readiness versus knowledge graphs runs the same test independent of which graph database sits underneath, worth doing before scaling past a single agent into a full AI agent harness.
See the context layer in action
Watch how Atlan governs the retrieval layer that graph databases like TigerGraph feed into, whatever query language sits underneath.
Watch the Demo SeriesWhat TigerGraph’s marketing doesn’t say about the layer above it
TigerGraph is production-proven infrastructure for storage and traversal. Jaguar Land Rover’s 45-minute supply-chain query, down from weeks, is real, sourced, and attributed to a named executive, not a marketing composite. What TigerGraph is not, on its current marketing surface, is easy to check: the homepage carries four percentage claims with no source attached, and three of the performance figures most often quoted about TigerGraph do not appear on the pages they are attributed to. The gap isn’t in what the engine does; it’s in the question its own marketing never poses.
A Native Parallel Graph engine answers what’s connected to what, at genuine scale. It doesn’t decide what a connection means to the business, and it doesn’t decide who, or which AI agent, is allowed to query it. Assuming Layer 1 performance is the whole answer to enterprise AI readiness is the mistake that shows up once a team scales past a proof of concept: the graph gets faster, but nothing new decides what an agent should be able to see inside it.
That question doesn’t have a settled, tidy answer, and it shouldn’t. It’s worth asking before the graph is already in production, not after.
FAQs about TigerGraph enterprise graph database
1. What is TigerGraph used for?
TigerGraph is a Native Parallel Graph database used for fraud detection, supply-chain analysis, recommendation engines, Customer 360, and entity resolution. Its most fully documented production account is Jaguar Land Rover, which cut supply-chain re-sequencing queries from weeks to about 45 minutes across 12 integrated data sources.
2. Is TigerGraph open source?
No. TigerGraph’s Community Edition is free, including for production use, but TigerGraph claims no open-source licence for the product itself. Its Open Source Notice covers only the third-party components bundled with it. The edition caps out at 300GB of combined graph and vector data and 16 CPUs on a single server, with no clustering.
3. What is GSQL?
GSQL is TigerGraph’s own query language: SQL-like in syntax and Turing-complete, with accumulators as the mechanism that makes multi-hop traversal fast without the join explosion a relational engine hits on the same query. GSQL is no longer the only way in. TigerGraph also supports most of openCypher, embedded within GSQL, and the path-pattern-matching subset of ISO GQL.
4. What is the difference between TigerGraph and Neo4j?
TigerGraph runs a Native Parallel Graph engine on an MPP/BSP model built for distributed, high-volume traversal; Neo4j uses index-free adjacency on a more tightly coupled engine. The benchmark figures circulating about the two almost all originate with TigerGraph. Its own benchmark page reports loading 1.8x to 58x faster and 2-hop path queries 40x to 337x faster than the systems it tested, from a 2018 report it ran itself.
5. What is TigerGraph Savanna?
Savanna is TigerGraph’s cloud-native platform, launched in January 2025, with independently scaling storage and compute and nine prebuilt solution kits spanning fraud, entity resolution, Customer 360, and supply-chain management. It’s the commercial tier above the free Community Edition.
6. Who uses TigerGraph?
Jaguar Land Rover is TigerGraph’s most fully sourced named account, with a published case study and a named quote. JPMorgan Chase, Ford, Intuit, and UnitedHealth appear in press releases and conference materials, but without the same case-study depth.
7. How much does TigerGraph cost?
Community Edition is free, capped at 300GB of combined graph and vector data and 16 CPUs on one server. TigerGraph publishes Savanna pricing: a Free Trial tier, Savanna at $45 per GB/month for 24-7 service in US and Tier 1 regions, and Savanna Business Critical at $126, with storage billed separately at $0.025 per GB/month and hourly instances from $1.00 to $256.00 depending on size. Self-managed Business Critical is contact-sales.
Sources
- TigerGraph: A Native MPP Graph Database, Deutsch, Xu, Wu, Lee, arXiv (TigerGraph-authored)
- TigerGraph DB product page, TigerGraph
- Internal Architecture, TigerGraph Server Docs
- Jaguar Land Rover case study, TigerGraph
- TigerGraph Savanna, TigerGraph
- TigerGraph pricing, TigerGraph
- TigerGraph Community Edition, TigerGraph
- Open Source Notice, TigerGraph
- 10 Graph Database Use Cases for Enterprise, TigerGraph blog
- TigerGraph benchmark page (vendor-run, 2018 report)
- Understanding Model Context Protocol (MCP), TigerGraph blog, 2025-06-06
- TigerGraph Accelerates Enterprise AI Infrastructure Innovation with Strategic Investment from Cuadrilla Capital, TigerGraph, 2025-07-15
- TigerGraph Server documentation home
