Kog Ignites AI Inference Speed on Existing GPUs

French startup Kog is challenging the AI hardware race by focusing on software optimization for existing datacenter GPUs. The company promises to unlock significant speed improvements for AI inference, aiming for a 30x acceleration on tasks that are currently bottlenecks. This approach targets enterprises already equipped with GPUs like AMD MI300X and NVIDIA H200, offering enhanced performance without requiring new hardware investments.

CEO Gaël Delalleau's unique background in solid-state physics and offensive cybersecurity informs Kog's deep-level GPU engineering strategy. This mindset involves reverse-engineering hardware to optimize performance beyond its intended design. The startup plans to achieve a 10x speed improvement on major large language models (LLMs) by September 2026, a milestone expected to attract significant customer traction and funding for their Series A round.

Kog's focus is on accelerating larger models, addressing market demand from sectors like software engineering and game development where processing speed directly impacts revenue. While the current hands-on methodology limits the number of chips they can optimize, the company envisions agent-based pipelines for broader future support. This ambition aligns with Europe's push for technological sovereignty, with Kog already receiving backing from French entities like Scaleway and Bpifrance.