AI project failed again?
As of 2025, 65% of enterprise AI failures stem from context-related issues — not the model itself. Better LLMs, more RAG tuning — if your team keeps hitting the same wall, the culprit probably isn't the model.
RAG isn't always the answer
RAG (Retrieval-Augmented Generation) has become the de facto standard for AI adoption. Pull in your PDFs, build a vector DB, let the model work from the relevant chunks — that flow is sound. But it hits walls.
Specifically, when you need to:
- Connect PDF reports + Excel data + CCTV footage simultaneously
- Link "safety incident at this site → weather data that day → last week's construction schedule"
- Find relationships between different data sources, not just within a document
RAG specializes in retrieving relevant chunks from within documents. One analysis described it as "a librarian sprinting to grab three relevant books" — but the problem is that RAG can't know the relationships between those books. Without structured entity context, accuracy can sit around 60%.
Enterprise environments are more complex. Organizations hold trillions of tokens worth of data, while today's LLMs max out around 1–2 million token context windows. And models start degrading 30–40% before their advertised limits. Gartner predicted 60% of enterprises would abandon AI projects by 2026 — citing poor data foundations as the root cause.
What if there was a layer before the model?
In May 2026, researchers from Amazon, Meta, Google, and NASA emerged from stealth. What they built wasn't a better model. It was a layer that structures relationships in your data before the model ever sees it.
Graphon AI launched as the "pre-model intelligence layer" with $8.3M in seed funding. Lead investor Novera Ventures was joined by Perplexity Fund, Samsung Next, and Hitachi Ventures.
The name has a fascinating backstory. "Graphon" comes from a mathematical concept formalized in 2008 by Jennifer Chayes (Dean of Computing at UC Berkeley) and Christian Borgs — a continuous function representing relationship structures in increasingly large networks. The idea: any complex data network can be represented mathematically.
Here's how it works: before a model processes data, Graphon functions automatically discover relationships across video, audio, documents, images, and databases — building what the company calls "persistent relational memory". That structured memory connects to any foundation model or agent framework.
| RAG approach | Pre-model layer approach | |
|---|---|---|
| When it acts | Retrieves chunks at query time | Structures data at ingestion time |
| Relationship scope | Within-document only | Cross-source relationship discovery |
| Data types | Primarily text documents | Video, audio, databases, images |
| Accuracy | ~60% without structure | 90%+ with structured knowledge |
| Model dependency | Needs tuning per model | Works with any foundation model |
GS Group is already using it
This isn't theoretical. South Korean conglomerate GS Group is already running Graphon in production.
Two use cases have been made public. First: customer movement analysis in convenience stores — connecting CCTV footage, purchase history, and store layout to understand movement patterns. The kind of multimodal connection that text-only RAG can't handle.
Second: real-time construction site safety monitoring. CCTV + sensor data + safety regulation documents, connected live to detect violations automatically. GS VP Ally Kim said "Graphon has been an invaluable partner in GS Group's AI transformation journey."
How the investor sees this opportunity
Novera Ventures' Arvind Gupta: "Graphon changes where the intelligence happens. Most companies are trying to build ever-larger models." The strategy is shifting competition away from models and toward the data layer.
Diagnose your team's AI failures right now
-
Classify your context failure type
Are your AI errors "no relevant info found" or "wrong connections made"? The first is fixable with better retrieval. The second likely points to a data structure problem. -
Map your data source connections
List every data type your AI needs to reference. Text documents only? RAG is probably fine. Mix of video, images, databases, IoT data? You need cross-source relational structuring. -
Try GraphRAG first
Microsoft's GraphRAG is open source. Use it as a middle step to see if accuracy improves over RAG. Note: ontology design can take time to set up. - For multimodal cases, consider a pre-processing layer
If you're combining video, sensors, and documents, Graphon AI's pre-model approach fits. Industries with heavy multimodal industrial data (manufacturing, logistics, retail, construction) get the fastest ROI. - Audit your data governance first
No layer helps if your data isn't classified and permissioned. Only 37–40% of organizations have AI-ready data practices. Check this foundation before any tool adoption.
