French startup Kog is betting that software optimization can unlock far more inference performance from conventional GPUs, targeting enterprises already running AMD MI300X and NVIDIA H200 hardware. The company attracted 200 business leads after a May tech preview that hit the front page of Hacker News, demonstrating 3,000 tokens per second (TPS) per request on its open-sourced Laneformer 2B model, a 2-billion-parameter model built specifically for the demo.

CEO Gaël Delalleau told TechCrunch that based on early feedback, software engineering will be the first use case — citing Claude Code users who sometimes wait hours for results. Kog also has design partners generating games and apps from prompts, where faster inference directly means more revenue. However, the startup learned that prospective customers are not prepared to fine-tune small models, so Kog has shifted focus entirely to accelerating larger models.

Delalleau comes from a background in solid-state physics and offensive cybersecurity, which he says shapes Kog‘s approach: understanding GPU hardware at the physical level and reverse-engineering it down to assembly language to use it for purposes it wasn’t originally designed for. This methodology is hands-on and time-consuming — Kog dedicates weeks or months to each new GPU. With an 11-person team, that limits the number of chips they can support for now, though they plan to eventually feed their techniques into agent-based pipelines to scale.

Kog‘s CEO dismisses the notion that GPUs are poorly suited for decoding, arguing that newer GPUs have memory bandwidth that software can unlock. The startup is backed by Scaleway, France’s Bpifrance, and French Tech 2030, and Delalleau expects to demonstrate a 10x speedup on a major LLM by September, which he says will be key to raising a Series A. The company is distinct from ZML, another French startup bypassing CUDA, positioning itself closer to Stanford’s Hazy Research lab but with an even deeper focus on GPU-level acceleration.