‘The GPU Never Reads the Same Text Twice’: Belgian Startup Corbenic Reimagines Memory for AI

Belgian startup Corbenic is proposing a novel way to squeeze more performance out of existing AI hardware. Its software platform, Galahad, moves part of a large language model’s working state—specifically the key‑value cache that stores attention information during inference—out of expensive GPU memory and onto local SSDs, then reloads it whenever the same text appears again.

In today’s chatbots and code assistants, models typically reread long histories from scratch, consuming vast amounts of memory bandwidth and compute time. Corbenic’s pitch is that “the GPU never reads the same text twice”: by checkpointing and reusing past context from encrypted storage, they claim up to 14‑fold throughput gains for repetitive workloads, along with meaningful reductions in infrastructure cost.

The approach comes as data centers face acute shortages of high‑bandwidth memory chips and spiraling AI infrastructure bills. If Galahad’s beta rollout proves reliable at scale, using commodity SSDs as a kind of “second‑tier AI memory” could offer enterprises a software‑centric alternative to constantly buying more accelerators—especially for knowledge management and customer‑support scenarios where similar queries recur again and again.