For the past three years, the generative AI arms race has been defined by a single, obsessive metric: the context window. Whether it’s 128k or a million tokens, the industry has been trying to shove more data into the “brain” of the LLM at the moment of inference.
But a new startup, comprised of alumni from the AI labs of Meta, Amazon, and Google, argues that we’re looking at the problem backward. Instead of building a bigger bucket, they want to change how the water is organized before it ever hits the model.
Graphon AI emerged from stealth today with $8.3 million in seed funding to build what it calls the world’s first “pre-model intelligence layer.” The round was led by Arvind Gupta at Novera Ventures, with a notable cap table that includes Perplexity Fund, Samsung Next, Hitachi Ventures, and Aurum Partners.
Moving Beyond the “RAG-Industrial Complex”
The current standard for enterprise AI is Retrieval-Augmented Generation (RAG). It’s a bit like a librarian sprinting to a shelf, grabbing three relevant books, and handing them to the AI. It works for simple queries, but it fails when you ask the AI to understand the relationship between a sensor log in a factory, a CCTV feed from the warehouse, and a PDF of safety protocols.
Graphon’s thesis is that the world isn’t a sequence of tokens; it’s a web of relationships.
“AI has spent the last decade learning to mimic language,” Graphon Founder and CEO Arbaaz Khan said in a press release. “But the world isn’t made of tokens, it’s made of relationships. By preserving that structure, we make foundation models more accurate and more useful at enterprise scale.”
Rather than waiting for a user to ask a question, Graphon’s layer sits between the raw data (video, audio, logs, databases) and the model. It uses graphon functions, a mathematical framework for representing massive, complex networks, to build a persistent, relational memory of the organization.
Worldclass Pedigree
The team’s technical bench press is heavy. The leadership includes researchers and engineers who have clocked time at MIT, NASA, NVIDIA, and Rivian. Perhaps more importantly, they’ve secured the backing of the literal architects of the field.
Jennifer Chayes, Dean of Computing at UC Berkeley, and Christian Borgs, the professor who coined the term “graphon,” are both serving as technical advisors. Their involvement suggests that this isn’t just another wrapper startup, but a serious play at applying high-level graph theory to the unstructured data problem that has plagued the enterprise.
Real-World Physics
While the math is dense, the applications are grounded. Graphon is already working with GS Group, the South Korean conglomerate. The use cases range from the mundane to the mission-critical:
- Convenience stores: Analyzing customer movement patterns across video feeds.
- Construction sites: Using CCTV and sensor data to spot safety violations or process gaps in real-time.
- Agentic Workflows: Giving AI agents a “map” of how an enterprise functions so they can make decisions without hallucinating a step in the process.
“Graphon changes where the intelligence happens,” says Arvind Gupta, Managing Director at Novera Ventures. “Most companies are trying to build ever-larger models. Graphon is improving the layer between raw enterprise data and the model itself.”
The Competitive Landscape
Graphon enters a crowded market of AI infrastructure players, but its “pre-model” positioning is a clever pivot. By claiming to be model agnostic, they aren’t competing with OpenAI, Anthropic, or Meta. Instead, they are positioning themselves as the essential plumbing that makes those models actually usable for a Fortune 500 company that has trillions of tokens of data but no way to link them.
With $8.3M in the bank and an all-star team, Graphon is betting that the future of AI isn’t just about how much a model can remember, it’s about how much it understands before the prompt is even typed.
Companies can learn more at https://www.graphon.ai/.
