Corbenic Galahad: SSD-Based KV Cache Reuse for AI Inference

What happened

Belgian startup Corbenic has unveiled Galahad, a product that aims to let AI inference servers reuse stored key-value (KV) cache data from solid-state drives instead of recalculating it on the GPU. The company claims a 14x speedup. According to TechRadar, the pitch centres on exact, encrypted KV cache reuse across three widely used inference engines: vLLM, SGLang and Llama.cpp.

The reporting places Galahad in a crowded field. TechRadar notes that LMCache, Nvidia and Phison have already moved into this space and got there first, so Corbenic is entering as a challenger rather than a pioneer.

The headline also questions where the speedup comes from, suggesting that much of the gain owes to the system reading less rather than to raw compute. That is an interpretive framing of the company's performance claim. The coverage available here does not include independent benchmark results, so the 14x figure should be read as the company's own claim.

Why it matters

For organisations running their own large language models, the bottleneck is often not model quality but how much work the hardware repeats. When a model processes a prompt, it generates intermediate attention data. Reusing that data for repeated or shared context avoids redoing the same calculation. Moving that cache onto SSDs, rather than holding it all in scarce GPU memory, could let operators serve more context and more users from the same hardware. This is a infrastructure and operating-model question as much as a performance one.

The experience angle is direct. Response time is a customer experience variable, and so is consistency when many users share long documents or conversations. Encryption matters too, because cached data can contain sensitive prompts. Any product that claims to speed up inference while protecting that data touches the trust layer of an AI-enabled service.

By the numbers

• 14x speedup claimed by Corbenic for Galahad, as reported in the headline coverage
• Three inference engines named as supported: vLLM, SGLang and Llama.cpp
• Three competitors named in the reporting: LMCache, Nvidia and Phison

The Renascence take

Most readers will fixate on the 14x. The more useful question is what the system no longer has to do, and whether that matters to the people on the other end of the service.

A speedup only becomes an experience improvement when it lands on a moment a customer notices. The principle here is that latency is designed behaviour: every avoided recalculation is valuable only if it reduces waiting at the journey step where patience runs out. Customer-obsessed operators should stop asking whether a vendor's benchmark is impressive and start measuring time-to-first-response on their highest-stakes interactions, under real concurrency and with encryption switched on. If the gain holds there, it is worth pursuing; if it doesn't, the 14x is just a number.