French startup Kog raises seed funding to speed LLM inference on GPUs

👁 119 reads

French startup Kog has closed a seed round co-led by Varsity VC to develop software that squeezes faster inference out of conventional datacenter GPUs, after a May tech preview generated 200 business leads, according to CEO Gaël Delalleau.

The company’s Kog Inference Engine (KIE) targets the AMD MI300X and Nvidia H200 GPUs that enterprises already own, positioning itself as a software-only alternative to purpose-built inference chips. Its demo recorded 3,000 per-request tokens per second (TPS) on Laneformer 2B, a purpose-built 2-billion-parameter model that Kog has since open-sourced. The stated long-term target is 30x faster LLM inference on full-scale large models.

Delalleau acknowledged the gap between the demo and that goal. The demo ran on a small model, but he says the same approach scales to larger LLMs - pointing to newer GPUs’ growing memory bandwidth as the key unlockable resource. “GPUs have a bright future,” he said, calling the idea that they are poorly suited for decoding a misconception.

Early feedback shaped the product roadmap. Prospective customers were not prepared to fine-tune small models, pushing Kog to pivot toward accelerating larger ones. Software engineering emerged as the clearest first use case - veteran Claude Code users sometimes wait hours for results, and Anthropic itself charges a price premium for Claude’s Fast Mode. Kog also has design partners building game and app generation tools, for whom lower latency would translate directly into higher revenue.

The seed round was co-led by Varsity VC, whose founding partner Kamel Zeroual is Delalleau’s former co-founder from Stribe, a TechCrunch50 2009 company. Delalleau studied solid-state physics at École Polytechnique and worked in offensive cybersecurity before founding Kog, a background he says informs a “laws of the GPU” mindset on his engineering team.

Practitioner takeaway: Developers running long agentic coding sessions or multi-step AI workflows on hosted APIs face real wait times today. If Kog can deliver meaningful speedups on the H200s and MI300Xs that cloud providers already operate, latency-sensitive tasks - think overnight coding agents or real-time app generation - could become noticeably faster without a hardware swap. The caveat is that KIE’s headline numbers come from a small custom model; its performance on GPT-class models remains undemonstrated.

// read moreRiver AI Raises $1.1B Seed Round at Two Months OldMagazineAI code-testing startup Blacksmith raises $45M at $550M valuationMagazineAI startup Discovered Materials raises $9M for cooler semiconductorsMagazine

Kinetyra