French startup Kog is trying to prove that the fastest path to faster AI inference runs through software, not new silicon. The company, founded by a solo founder with a background in offensive security, claims its optimization layer can deliver dramatically faster decoding on the standard data center GPUs enterprises already own.
Kog made the front page of Hacker News in May with a tech preview running on AMD MI300X and Nvidia H200 hardware. The demo hit roughly 3,000 tokens per second, but on a purpose-built small model of only about 2 billion parameters, and the open-source project that followed is now the foundation of its work. The startup’s bigger promise, a 30x speedup on large language models, remains unproven at scale.
Founder Delalleau studied solid-state physics at École Polytechnique before spending years in offensive cybersecurity, a background he says taught him to reverse-engineer systems down to assembly level. For every new GPU, the team of 11 plans to spend weeks or months digging into the hardware, a hands-on approach that is slow but thorough.
The first use case is software engineering, where developers wait hours for agentic coding tools to finish. Kog also has design partners generating games and apps from prompts, where faster output directly improves the product. Its seed round was co-led by Varsity VC, whose partner Kamel Zeroual worked with Delalleau at an earlier startup.
With Europe pushing for AI sovereignty, a French team squeezing more from existing chips fits the political moment. The question is whether Kog’s deep-hardware approach can scale beyond its founding model.
