Graphon AI Raises $8.3 Million Seed to Build Pre‑Model Data Layer for Enterprise AI
Key Points
- Graphon AI raises $8.3 million seed, led by Novera Ventures.
- Investors include Perplexity Fund, Samsung Next, Hitachi Ventures, GS Futures and others.
- Founders Arbaaz Khan, Deepak Mishra and Clark Zhang hail from Amazon, Meta, Google, Apple, NVIDIA and NASA.
- Technical advisors Jennifer Chayes and Christian Borgs co‑invented the graphon mathematics behind the platform.
- Product creates a pre‑model layer that maps relationships across documents, video, audio, images and logs.
- Early customer GS Group uses the technology for store‑traffic analytics and construction‑site safety.
- Graphon aims to overcome large language model context‑window limits by structuring data before ingestion.
Graphon AI emerged from stealth this week with an $8.3 million seed round led by Novera Ventures and backed by investors including Perplexity Fund, Samsung Next, and GS Futures. The San Francisco startup aims to create a “pre‑model intelligence layer” that automatically discovers relationships across multimodal enterprise data before it reaches large language models. Founded by Arbaaz Khan, Deepak Mishra and Clark Zhang, the company already counts South Korean conglomerate GS Group among its first customers, using the technology for store‑traffic analytics and construction‑site safety.
Graphon AI announced Wednesday that it has closed an $8.3 million seed round and is stepping out of stealth mode. The funding, led by Novera Ventures with participation from Perplexity Fund, Samsung Next, Hitachi Ventures, GS Futures and several other investors, will fund the development of a "pre‑model intelligence layer" designed to surface relational structure across an organization’s multimodal data before it reaches a foundation model.
The startup’s founders – Arbaaz Khan (chief executive officer), Deepak Mishra (chief operating officer) and Clark Zhang (chief technology officer) – bring experience from Amazon, Meta, Google, Apple, NVIDIA and NASA. Their technical advisors include UC Berkeley’s Jennifer Chayes, dean of the College of Computing, Data Science and Society, and computer‑science professor Christian Borgs, both of whom co‑invented the mathematical concept of a graphon that underpins the company’s approach.
Graphon’s product sits "before the model," not inside it. By applying graphon functions, the platform ingests documents, video, audio, images, logs and database entries, then automatically maps the relationships among them. The result is a persistent relational memory that any large language model or AI agent can query without being limited by the model’s context window, which typically caps at about one million tokens.
Current retrieval‑augmented generation (RAG) techniques can pull relevant snippets from a massive corpus, but they struggle to reason across data types that were never stored together. Graphon claims its layer can bridge that gap, enabling an LLM to answer questions that span a compliance log, a surveillance video and a customer database in a single, coherent response.
Early traction comes from GS Group, the South Korean conglomerate whose venture arm GS Futures also participated in the round. A vice president at GS, Ally Kim, said Graphon’s solution is already being used to analyze shopper movement in convenience stores and to improve safety on construction sites through CCTV analysis.
The investor roster reflects a broad belief that the problem Graphon tackles cuts across industries. Perplexity Fund, the venture arm of an AI search company, Samsung Next, the corporate venture unit of Samsung, and Hitachi Ventures all see value in a data‑layer that could make foundation models more useful for enterprise workloads.
While the company’s ambitions are sizable – it cites use cases ranging from enterprise content management to on‑device AI for phones, cameras and smart glasses – independent benchmarks are not yet public. Observers note that the AI infrastructure market has spent the past three years racing to build larger models with longer context windows. Graphon’s bet is that structuring data before it enters the model may be a more effective way to overcome the “context‑window ceiling.”
With seed capital secured and a roster of high‑profile advisors, Graphon AI now faces the classic startup test: turning a mathematically elegant concept into measurable performance gains at scale. The $8.3 million infusion gives the team runway to refine its platform, expand its customer base and demonstrate that a graphon‑based data layer can indeed close the gap between what large language models can theoretically understand and what they can practically deliver for complex, multimodal enterprise environments.