While custom inference chips grab IPO headlines, a French startup named Kog is making a simpler bet: squeeze more speed out of the Nvidia and AMD GPUs companies already bought. TechCrunch reports Kog hit Hacker News in May with a tech preview showing very fast single-request decoding on datacenter cards such as the AMD MI300X and Nvidia H200. It does not magically speed up a laptop GPU, which disappointed some hobbyists, but enterprises heard a different message.
CEO and solo founder Gaël Delalleau told TechCrunch the preview produced about 200 real business leads. The first target use case is software engineering. Anyone who has parked a long Claude Code or agent job and waited hours knows speed is now a product feature, not a nice extra. Anthropic already charges more for Fast Mode. Kog wants the shops that refuse those delays, plus design partners that generate games and apps from a prompt and make more money when answers arrive sooner.
The product is the Kog Inference Engine, or KIE — software that sits on standard GPUs instead of a new rack of specialty silicon. That matters in a year when memory and accelerators are expensive and lead times are ugly. If Kog’s claims hold in production, CIOs get another lever besides “buy more H200s.” If they do not, it stays a clever demo. Either way, the inference race is no longer only a chip story.
Source: https://techcrunch.com/2026/08/14/kog-is-going-deeper-to-squeeze-more-inference-out-of-gpus/
