Former Amazon and Meta researchers launch Graphon AI as enterprises confront AI’s growing memory problem
Pictured from left to right: Deepak Mishra, Co-Founder & COO; Arbaaz Khan, Founder & CEO; Clark Zhang, Co-Founder & CTO
Two years after generative artificial intelligence entered the enterprise mainstream, many organisations are discovering that large language models remain limited in how they understand relationships between fragmented pieces of operational information.
Chatbots can summarise documents, generate reports and answer questions fluently. Yet inside large organisations, artificial intelligence systems still struggle to maintain continuity across workflows, disconnected datasets and changing physical environments.
Former researchers from Amazon and Meta believe this has become one of the next major bottlenecks in enterprise AI infrastructure.
Their startup, Graphon AI, emerged from stealth in May 2026 with $8.3 million in seed funding and a claim that the future of enterprise AI may depend less on building larger models and more on helping systems understand relationships between information before inference begins.
The funding round was led by Khosla Ventures, with participation from Samsung NEXT and M Ventures. The company says its technology is already being tested in industrial and construction environments, including projects involving South Korean conglomerate GS Group.
Graphon AI describes its platform as a “pre-model intelligence layer”, terminology that reflects a broader shift taking place across the AI industry as companies increasingly focus on infrastructure surrounding AI models rather than models alone.
The AI industry is moving beyond “bigger models solve everything”
During the first phase of the generative AI boom, much of the sector’s attention centred on scale.
Larger models, larger datasets and larger context windows were widely presented as the path towards more capable AI systems. Companies raced to increase the amount of information models could process in a single prompt, with some systems now capable of ingesting lengthy videos, extensive documentation and large codebases.
Inside enterprises, however, organisations have encountered more practical limitations.
AI systems may retrieve relevant documents without understanding how those documents relate to operational processes. They may identify information without recognising historical dependencies between systems, departments or events. Long prompts can also increase computational costs while making outputs less predictable.
Those constraints are becoming increasingly visible as companies attempt to move AI beyond experimental chatbots into infrastructure, logistics, manufacturing, healthcare and industrial operations.
In those environments, conversational fluency alone is not enough.
An AI system monitoring a construction site, factory floor or logistics network must maintain continuity across changing workflows, sensor feeds, maintenance histories and physical environments. Understanding relationships between events becomes as important as generating language.
That is the problem Graphon AI says it wants to address.
From retrieval-augmented generation to relational AI systems
Much of today’s enterprise AI infrastructure relies on retrieval-augmented generation, commonly known as RAG. In these systems, external information is retrieved from documents or databases before being passed into a large language model.
The approach became widely adopted because it allowed organisations to connect generative AI systems to internal company knowledge without retraining foundation models themselves.
However, retrieval alone does not necessarily provide contextual understanding.
Graphon AI argues that enterprise systems increasingly require structured relational reasoning rather than simple document retrieval. Its platform attempts to organise enterprise data into interconnected knowledge structures before information reaches the underlying AI model.
The company describes this as a “pre-model intelligence layer”, although the underlying concepts are not entirely new. Knowledge graphs, relational databases and semantic mapping systems have existed for years inside search engines, enterprise software and recommendation engines.
What has changed is the rise of multimodal AI systems capable of processing text, images, video and operational data simultaneously.
That shift is forcing AI infrastructure companies to rethink how information is organised and prioritised before inference takes place.
“Our goal is to give AI systems a structured understanding of reality before inference begins,” Graphon AI said in its launch announcement.
The wording may sound ambitious, but the broader challenge is increasingly recognised across the enterprise AI market.
AI’s memory problem is becoming an infrastructure business
The rise of companies focused on vector databases, AI memory systems, orchestration platforms and graph architectures suggests the AI market is entering a new phase.
The first phase of the generative AI boom focused primarily on model capability. The next phase increasingly revolves around operational reliability.
That includes:
- reducing hallucinations,
- maintaining persistent context,
- improving reasoning consistency,
- lowering inference costs,
- and integrating AI into operational workflows.
In practice, many enterprise AI systems still lose coherence when tasks become long-running, operational or dependent on fragmented historical information.
This has become particularly important in industrial environments where decisions may depend on maintenance histories, equipment relationships, workflow dependencies or evolving site conditions.
Graphon AI is positioning itself within that transition.
The company says its system can connect data from multiple sources, including video feeds, operational documentation and enterprise systems, into relational structures that help AI systems identify meaningful connections over time.
Construction environments have become one early testing ground.
Why industrial environments expose AI’s weaknesses
Graphon AI says its platform is being deployed with GS Group in construction environments where AI systems analyse video streams, operational records and workflow data together.
Construction sites are highly dynamic environments involving contractors, machinery, safety protocols, scheduling systems and changing physical conditions. Information is distributed across cameras, reports, maintenance logs and planning tools.
Traditional conversational AI systems struggle in such environments because they are not naturally designed to maintain operational continuity across fragmented systems.
That challenge is not unique to construction.
Manufacturing facilities, logistics networks and critical infrastructure environments increasingly expose the limits of AI systems built primarily for conversational tasks.
The problem becomes more pronounced as enterprises attempt to deploy AI agents capable of autonomous or semi-autonomous decision-making.
An AI agent that forgets previous workflow steps, misunderstands operational dependencies or loses contextual awareness can introduce business risks rather than operational efficiencies.
That is one reason AI infrastructure startups have become one of the fastest-growing segments of the broader AI market.
The economics of AI are changing
Another factor driving interest in relational AI systems is cost.
Processing extremely large context windows requires substantial computing resources, particularly in enterprise environments where multiple AI systems may operate simultaneously across large datasets.
The AI industry’s rapid expansion has already intensified concerns around infrastructure costs, energy consumption and scalability.
Companies are therefore increasingly looking for ways to make AI systems more contextually efficient rather than simply larger.
Graphon AI claims its relational approach reduces unnecessary computation by helping models focus on relevant information instead of processing entire datasets indiscriminately.
Those claims remain difficult to independently verify. Like many early-stage AI startups, the company has released limited public benchmarking data.
Still, the broader direction across the industry is becoming clearer.
Rather than relying exclusively on larger models, enterprises are increasingly investing in AI infrastructure layers designed to improve how models access, organise and interpret information.
The enterprise AI stack is fragmenting
The emergence of companies such as Graphon AI reflects a broader fragmentation taking place across the AI ecosystem.
Rather than converging around a single dominant architecture, enterprise AI is evolving into a growing stack of specialised infrastructure layers, including vector databases, orchestration systems, AI observability tools, memory platforms, governance software and graph-based reasoning architectures.
Many of these companies are attempting to solve problems that became visible only after generative AI moved from demonstrations into operational environments.
That transition is changing how investors and enterprises think about the sector.
For much of the past two years, AI value was concentrated primarily around foundation model developers. Increasingly, however, the market is recognising that deploying AI reliably inside complex organisations requires substantial supporting infrastructure.
Graphon AI is one of several startups attempting to define what that infrastructure layer may look like.
Whether “pre-model intelligence layers” become an enduring category remains unclear. The terminology itself may evolve as enterprises experiment with different approaches to memory, orchestration and contextual reasoning.
What appears more certain is that many organisations are discovering that larger models alone do not automatically produce operationally intelligent systems.
The next phase of enterprise AI may therefore depend less on generating convincing language and more on helping machines maintain structured understanding across increasingly complex environments.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.