Key points
- Galahad entered beta on October 1, 2026, allowing servers to store key-value cache on local NVMe drives.
- The system supports vLLM, SGLang, and Llama.cpp engines, using encryption for stored attention states.
- Corbenic AI attributes much of the claimed 14x speedup to a retrieval component called Blaise, not just storage.
The short version
- Corbenic AI released Galahad beta to persist AI model key-value caches to encrypted local NVMe storage.
- The tool aims to prevent GPUs from reprocessing identical text inputs after memory clears or server restarts.
- Competitors like Nvidia and LMCache already offer similar KV cache persistence to storage solutions.
The details
Corbenic AI launched Galahad into beta on October 1, 2026, introducing a mechanism to save key-value (KV) cache data to encrypted local NVMe drives. According to TechRadar, the Belgian startup designed Galahad to allow servers to skip rereading text after GPU memory is cleared or after a system restart. Founder and CEO Sietse Schelpe stated that the solution requires no special hardware, utilizing the NVMe drive already present in the server. The system integrates with serving engines such as vLLM, SGLang, and Llama.cpp to manage these cached attention states.
Galahad’s KV component, named Taliesin, writes each block of text’s state to disk under a fingerprint of the exact input bytes, the model, and the tenant. TechRadar reports that the stored copy only activates once GPU memory has been cleared or the server has restarted. Schelpe emphasized that the choice is not between SSD or RAM, but rather an extension of working memory. The company introduced Galahad with the tagline "the GPU never reads the same text twice," positioning the SSD as part of the AI's working memory.
The context
Corbenic AI claims a 14x speedup with Galahad, but its own paper credits most of this performance gain to Blaise, a retrieval component, rather than SSD loading. According to TechRadar, the startup is entering a crowded field where Nvidia, Phison, WEKA, and the open-source project LMCache already persist KV cache to storage. Nvidia Dynamo and CMX also offer similar capabilities. The timing aligns with data centers expected to absorb more than 70% of high-end memory chips in 2026. Galahad’s edge lies in exact, fail-closed reuse within a small library, contrasting with competitors that may handle cache eviction differently under memory pressure.
What to watch
- Monitor how Galahad handles cache eviction under high memory pressure compared to standard vLLM or SGLang implementations.
- Watch for updates on compatibility with other AI serving frameworks beyond vLLM, SGLang, and Llama.cpp.
- Track performance benchmarks to verify if the 14x speedup holds in production environments with varied workloads.
What it means for organisations
Teams evaluating AI infrastructure should assess whether persistent KV caching aligns with their data retention policies and encryption standards. Operations leaders need to determine if extending memory to NVMe storage introduces latency issues or complexity in multi-tenant environments. Security teams must verify that the encryption methods used for cached attention states meet organisational compliance requirements. Procurement should compare Galahad against existing solutions from Nvidia or open-source alternatives to ensure vendor lock-in risks are managed. Internal communication should clarify how this tool impacts hardware procurement and existing server configurations.
Background: Encryption
Encryption is the process of converting plain text into ciphertext using an algorithm and a key. It prevents unauthorized access to data at rest and in transit. Without the correct decryption key, the information remains unintelligible, protecting privacy and integrity against interception or theft.
Read the full guide: What Is Encryption: Meaning, Definition, and How It Works
What to do and how to stay safe: Corbenic AI
- Review current AI serving infrastructure to identify if GPU memory pressure causes frequent cache evictions.
- Evaluate whether existing NVMe storage capacity can accommodate additional KV cache data without impacting other workloads.
- Consult with security teams to ensure the encryption standards for stored attention states meet organisational policies.
- Monitor vendor updates for Galahad to confirm compatibility with your specific AI model versions and serving engines.
Step-by-step guide: Secure Home Routers for Small Business Networks
General security guidance from the TechSec Headlines newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does Galahad require special hardware to function?
No, founder Sietse Schelpe stated that Galahad needs nothing more exotic than the NVMe drive already sitting in the server.
Which AI serving engines does Galahad support?
According to TechRadar, Galahad promises exact, encrypted KV cache reuse for vLLM, SGLang, and Llama.cpp.
What is the main competitor to Galahad in this space?
Competitors include LMCache, Nvidia Dynamo, WEKA, and Phison, which already persist KV cache to storage.
