When Mel Morris, founder of Corpora.ai, ran a direct cost comparison between his platform and a leading frontier model earlier this year, the result was stark enough to stop the conversation.

“I can’t say the frontier model’s report wasn’t good; it was very good indeed,” Morris tells Capacity. “But we ran the same job using our system, and the job was 30% faster, it had 20% more citations, and the cost was five cents.”

The token count tells its own story. The frontier model consumed almost 10 million input tokens to complete the task. Corpora used just under 600,000 and produced a more expansive report.

A different architecture entirely

Corpora.ai is not a model, and Morris is emphatic on this point. The platform is a hybrid database architecture that combines the properties of a graph, vector, temporal, NoSQL, and geospatial database in a single structure, designed to ingest, decompose, and fully correlate documents at scale, then serve the net unique relevant content to whichever model handles summarisation.

At its core is a graph technology built entirely in-house, after Morris concluded that existing graph databases, he cites Neo4j and Dgraph as examples, could not handle the volumes Corpora required. “They would say, ‘Oh yes, we can handle three or four terabytes, maybe five.’ Our ambitions were significantly in excess of that. We wanted to look at hundreds of petabytes.”

The solution was to scale up rather than scale out, pushing a single server as far as it would go rather than distributing the graph across thousands of instances. “The more you can do in one instance, the less you have to scale out. And if you divide your data across 20 graphs, you never get a perfect structure, you lose the connections between them.”

Today, Corpora’s platform can process two million documents per second on a single enterprise-class server, roughly configured around two AMD Turin processors, four terabytes of RAM, and a large number of NVMe drives. The graph currently holds in excess of 200 petabytes of data, fully automated, requiring no manual tuning or sharding.

“From the moment a new document arrives in our service, within seconds it is fully represented within the graph,” Morris says.

Why conventional agentic search is expensive by design

The cost gap between Corpora and frontier model-based research stems from how each system retrieves information. A typical agentic research workflow dispatches 20 to 30 web searches, fetches and unpacks each result, trims the content, and passes a curated packet to the model for synthesis. The problem, Morris argues, is redundancy.

“By the time you get to the 30th document, the amount of net unique content represented across all 100 is actually quite small, maybe three times what the first document said. But you’ve had to look at 100 to get there.”

Corpora derives net unique relevant content directly from its graph, bypassing the fetch-and-process cycle entirely. “You end up serving results in milliseconds. A typical agentic web search might take 10 to 20 seconds per thread, and each of those web searches has a cost, and then processing each page has a cost, and that starts to stack up significantly.”

The company also owns its hardware outright, a deliberate choice Morris says was essential to achieving the performance optimisation the graph technology required. For summarisation, Corpora uses smaller open-weight models running on Nvidia RTX 6000 Max-Q, avoiding frontier model inference costs for tasks he argues don’t require them. “Our graph and all the underlying technology run without GPUs whatsoever. We only use GPUs for the AI summarisation layer.”

A pointed view on UK sovereign AI

Morris is supportive of the UK’s ambition to become a serious AI contender but pulls no punches on what he sees as the central blind spot in current strategy. “People tend to have this view that in order to be competitive, we have to have more GPUs than someone else. That’s how the race is defined today.”

The real constraint, in his assessment, is energy. “We have one of the highest energy costs in the world. That is a massive inhibition in terms of our ability to not only install GPUs, but actually to power them.” He points to OpenAI’s withdrawal from a planned northeast England facility as a signal that energy economics are already influencing where AI infrastructure lands.

His prescription is a more joined-up national conversation between chip designers, model providers, energy suppliers, and infrastructure operators, rather than funding each in separate buckets. “We have to start saying: what does the sum total capability look like? We can be ultra competitive when we put all those people together. But we don’t have those conversations.”

For now, Corpora is working selectively with universities and startups, with plans to turn its attention to the broader sovereign AI infrastructure opportunity in 2027. The technology, Morris says, makes a simple case for itself. “Don’t use GPUs for tasks that Corpora could do for a fraction of the cost, much faster and much better. If our energy costs are high, it makes sense to get more for less.”

RELATED STORIES

UK AI data centre investments: Empty promises or slow growth?

What is sovereign AI infrastructure, and why does it matter for telecoms?