Kog Speeds AI Inference by Reworking GPU Software

Paris-based Kog, founded in 2023 by École Polytechnique graduate Gaël Delalleau, is betting that software rather than silicon is preventing conventional data-center GPUs from reaching their inference potential. The challenge is becoming more important as coding agents and other generative-AI systems execute long, sequential loops in which every planning, testing and revision step depends on the last. Faster single-request decoding could shorten response times and reduce the infrastructure cost of such professional workflows.

Kog released a technology preview of the Kog Inference Engine, or KIE, on May 28, 2026. The company said a 2B model produced 3,000 output tokens a second on eight AMD MI300X GPUs and 2,100 on eight NVIDIA H200 chips, using FP16 without speculative decoding. Kog has raised $5 million from Varsity VC and Bpifrance’s Deep Tech Program. Its pitch has drawn growing enterprise interest as it works to extend the system to larger third-party mixture-of-experts models.