French startup Kog claims up to 30x faster AI inference on conventional datacenter GPUs

French startup Kog, founded by Gaël Delalleau, claims to achieve up to 30x faster AI inference speeds on conventional datacenter GPUs (AMD MI300X, Nvidia H200), without specialized hardware. The company generated 200 qualified business leads following its Hacker News launch in May, with initial deployment expected in software engineering workflows. Kog developed an inference engine (KIE) that optimizes larger models through algorithmic improvements, with its demo showing 3,000 tokens per second. The company sees inference speed and costs as critical bottlenecks in the AI market, as customers seek alternatives to higher-cost fast modes offered by competing platforms.