Corbenic AI launches Galahad to store encrypted KV cache on local NVMe drives
Belgian startup Corbenic AI released Galahad beta on October 1, 2026, using server SSDs to retain attention states for repeated text inputs.
Key points
- Galahad saves key-value caches to encrypted local storage to avoid recomputing attention states for identical prompts.
- The solution supports vLLM, SGLang, and Llama.cpp without requiring specialized hardware beyond existing NVMe drives.
- Corbenic attributes its reported 14x speedup primarily to a retrieval component called Blaise, not just SSD loading.
Belgian startup Corbenic AI entered the beta phase of its Galahad product on October 1, 2026. The software allows servers to store large language model attention states on local NVMe drives, enabling GPUs to skip rereading text that has already been processed.
According to TechRadar, founder and CEO Sietse Schelpe stated that the system uses the solid-state drive already present in the server as part of the AI's working memory. The approach aims to reduce redundant computation by retaining key-value cache data across memory clears or server restarts.
Technical mechanics and performance claims
Galahad targets the prefill step in standard serving, where models recompute attention states from the first token for repeated documents. While engines like vLLM and SGLang reuse these states while they remain in GPU memory, Corbenic notes that this data is lost when memory pressure evicts it or when processes restart.
The system uses a component named Taliesin to write each text block's state to disk. This storage is keyed by a fingerprint of the exact input bytes, the specific model, and the tenant. The cached data is only loaded when those exact bytes reappear in a new request.
Corbenic’s accompanying paper claims a 14x speedup, but Schelpe told TechRadar Pro that this gain owes more to reading less text than to the speed of SSD loading. A retrieval component named Blaise is credited with the majority of this performance improvement, rather than the storage mechanism itself.
Market context and competition
The announcement arrives as data centers are projected to absorb more than 70% of high-end memory chips in 2026. Corbenic positions Galahad as a way to extract more work from existing hardware without needing exotic new equipment.
However, the company faces established competition in this niche. According to TechRadar, Nvidia Dynamo, WEKA, Phison, and the open-source project LMCache already offer solutions that persist key-value caches to storage. Corbenic differentiates Galahad by offering exact, fail-closed reuse through a single small library.
Schelpe emphasized that the stored copy only activates once GPU memory is cleared or the server restarts. He described the choice not as SSD versus RAM, but as a method to preserve state when volatile memory is unavailable or reset.
What should you do?
Security and infrastructure teams should evaluate how Galahad handles encrypted local storage in their specific environments. Since the product is in beta, organizations should test its integration with vLLM, SGLang, and Llama.cpp before deploying it in production workloads.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Corbenic AI
- Review the encryption standards used by Galahad for local NVMe storage to ensure compliance with data protection policies.
- Benchmark the 14x speedup claim in your specific environment, noting that Corbenic attributes most gains to the Blaise retrieval component.
- Compare Galahad's fail-closed reuse mechanism against existing solutions like LMCache or Nvidia Dynamo to determine the best fit for your stack.
- Monitor the beta release for updates on compatibility with tenant-specific fingerprinting and multi-model deployments.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Galahad require new hardware to function?
No, according to founder Sietse Schelpe, the system uses the NVMe drive already present in the server and does not require special hardware.
Which inference engines are supported by Galahad?
The product supports vLLM, SGLang, and Llama.cpp, as stated by Corbenic AI in its beta release announcement.
Where does the reported 14x speedup come from?
Corbenic's paper credits most of the 14x speedup to Blaise, a retrieval component, rather than the speed of loading data from the SSD.
