French startup Kog hit the front page of Hacker News in May with a bold claim: standard datacenter GPUs like the AMD MI300X and Nvidia H200 can deliver dramatically faster inference with the right software. The company’s demo showed 3,000 tokens per second — but there’s a catch. That number came from a purpose-built 2 billion parameter model, not a full-size LLM. The gap between demo and promise has drawn skeptics, though CEO Gaël Delalleau isn’t backing down.
The demand signal is real. Kog says it pulled in 200 tangible business leads after the launch, with software engineering emerging as the first target market. Delalleau points to Claude Code users who wait hours for results as exactly the pain point Kog wants to eliminate. Anthropic already charges a premium for Claude’s Fast Mode, which tells you speed sells. Kog is also working with design partners who generate games and apps from prompts, where faster inference translates directly into revenue.
Delalleau’s background is unusual for an AI infrastructure founder. He studied solid-state physics at École Polytechnique, then spent years in offensive cybersecurity — including four DEFCON CTF finals. That combination, he argues, gives his team a unique lens: understanding the laws of the GPU like physics, then reverse-engineering down to assembly language to squeeze out performance the hardware was never designed to deliver. The tradeoff is brutal, though. Each new GPU architecture requires weeks or months of deep engineering work, and with only 11 people, Kog can realistically support a limited number of chips.
That constraint matters because Kog competes in a space where Cerebras just went public selling the opposite thesis: purpose-built inference chips. Kog is betting enterprises would rather unlock performance from hardware they already own than rip and replace. The company has French government backing through Bpifrance and Scaleway, and plans to feed its methodology into agent-based pipelines to scale chip support. The real test comes in September, when Delalleau says Kog will implement its first major model at 10x speed — and if that lands, a Series A follows.
