Optimizing AI Workloads Through Local Storage
Belgian startup Corbenic AI has launched Galahad, a solution aimed at preventing GPUs from redundantly processing identical documents. By utilizing the NVMe drives already present in enterprise servers, the technology seeks to turn existing storage into an extension of an AI model's working memory. As data centers face increasing pressure to maximize the efficiency of hardware amid a global shortage of high-end memory chips, such solutions have become highly relevant.
Galahad, which entered beta on October 1, 2026, functions by saving a model's Key-Value (KV) cache—the attention state generated while a model reads a prompt—to encrypted local storage. When the same text is processed again, the system reloads this state, bypassing the need for computationally expensive re-prefilling.
The Role of Retrieval in Performance Gains
While Corbenic founder Sietse Schelpe highlights a significant performance boost—citing a 14x speedup in some scenarios—the company’s internal documentation suggests that much of this efficiency stems from a retrieval component called Blaise rather than just SSD loading. Blaise optimizes performance by providing the model only with the specific sections of a document required to answer a query, significantly reducing the total amount of text processed.
Regarding the technology's mechanics, Schelpe explained that the choice is not between RAM and SSD, but between utilizing storage or forcing the GPU to recompute the data. «So the choice is not SSD or RAM. It is SSD or computing it all again on the GPU,» he noted, adding that the reload occurs only once before the model generates its response.
Competitive Landscape and Industry Context
The field of KV cache persistence is becoming increasingly crowded. Major players, including Nvidia with its Dynamo and CMX frameworks, as well as projects like LMCache, WEKA, and Phison, are already offering methods to move cache data to storage. Corbenic differentiates Galahad by emphasizing "exact, fail-closed reuse," ensuring that if a stored record fails a consistency check against input bytes and model parameters, the system defaults to recomputing the data to avoid errors.
Security is another pillar of the Galahad pitch. The library encrypts the stored state and mandates tenant identity verification to prevent data leakage—a response to research indicating that shared KV caches could potentially be exploited to reconstruct user prompts.
Challenges and Future Outlook
Despite the potential benefits, Corbenic faces hurdles. As a small team with self-published research, the startup must compete with well-funded open-source projects and industry giants. Furthermore, independent benchmarks are currently lacking. For the industry to fully embrace the technology, experts suggest that Galahad must undergo head-to-head testing against established solutions like LMCache on identical hardware.
Ultimately, Corbenic’s value proposition aligns with a growing industry sentiment: software-based optimization is a critical alternative to simply purchasing more GPU capacity. Whether Galahad can establish itself as a standard layer in the AI stack remains to be seen, but its focus on making existing server infrastructure work harder is a trend that is unlikely to fade.