Graphon AI gets $8.3 million to lighten the load on AI models

AI startup Graphon AI has secured $8.3 million in seed funding to tackle a core limitation of current artificial intelligence: the immense and costly computational load required to understand vast, interconnected datasets. The company’s ‘intelligence layer’ aims to map relationships across data outside of a large language model, a move that could reduce processing costs and unlock insights from previously inaccessible information.

“It’s a fundamental new technology as opposed to something that can make AI a bit more efficient,” Arvind Gupta of Novera Ventures, the round’s lead investor, said.

Even the most advanced LLMs are limited to processing millions of tokens at a time, while organizations “hold trillions of tokens across documents, videos, logs and databases,” the company said. Graphon’s $8.3 million seed round, with participation from Perplexity Fund, Samsung Next, GS Futures, Hitachi Ventures, and others, will fund the buildout of its AI infrastructure class designed to bridge this gap.

The technology’s success could impact the AI infrastructure landscape by offering a more efficient and scalable alternative to current methods. For companies sitting on massive, unstructured datasets, it presents a potential way to extract value more cheaply. This may influence future investment trends and affect the competitive positioning of companies reliant on large-scale LLMs.

The San Francisco-based company was founded by former Amazon senior applied scientist Arbaaz Khan, who serves as CEO. Khan says the idea is to create a relational map of an organization's entire data universe—from documents and videos to system logs—before it ever touches the LLM. This pre-processing is designed to be more efficient than having a massive model repeatedly analyze all the data to find connections.

From Robotics to Relational Data

Khan drew inspiration from his doctoral work in robotics at the University of Pennsylvania. A robot operating in a defined space, he explained, can use knowledge of that structure to reduce its computational needs. He applied a similar idea to data, using the mathematical concept of a graphon, which can identify and group disparate users or data points into "neighborhoods" based on shared relational properties. While the transformer technology in LLMs expends immense power figuring out which words are related, Khan’s intelligence layer does this work separately. “We’ll go build this big relational representation... and that is what is going to feed the model, instead of having the model do all of the heavy lifting,” Khan said. He argues this is a massive savings, stating it is "a lot more efficient to run this 200 million [parameter model] a thousand times than it is to try and run like a 5 trillion [parameter model] for one hour.”

Korean Conglomerate Puts Tech to the Test

South Korean conglomerate GS, an investor through its GS Futures arm, is already using the technology. Ally Kim, a vice president at GS who leads its 52g digital transformation initiative, said the team used Graphon to improve the analysis of closed-circuit television recordings that monitor construction sites for safety compliance. The company also used it to more efficiently analyze video of soccer players to scout for a GS-sponsored team, evaluating movements, strengths, and weaknesses. “We really need to expand our knowledge scope to multimodalities, like voice or video or other contexts. Graphon can be good support,” Kim said.

This article is for informational purposes only and does not constitute investment advice.