Geo News
Community curated by people like you
LatestAICryptoHealthWorld AffairsUS Politics
French startup Kog pursues AI inference optimization amid GPU chip competition
00

French startup Kog pursues AI inference optimization amid GPU chip competition

Aug 14, 2026

French startup Kog, led by solo founder Gaël Delalleau, is developing low-level software optimizations to accelerate AI inference on standard datacenter GPUs like the Nvidia H200 and AMD MI300X. Following a May 2026 tech preview that achieved 3,000 tokens per second on a 2B parameter model, Kog is shifting focus to larger models to satisfy demand from 200 business leads. Backed by Varsity VC and Bpifrance, the 11-person team aims to achieve 10x speed on a major model by September 2026 to kick off its Series A round.

Kog inference optimization technology

  • ▪The Kog tech preview demonstrated 3,000 per-request tokens per second using Laneformer 2B, an open-source model with approximately 2 billion parameters.
  • ▪French startup Kog is developing software optimization technology to increase the speed of artificial intelligence inference on conventional datacenter GPUs.
  • ▪Kog released a tech preview in May 2026 demonstrating real-time large language model inference on AMD MI300X and Nvidia H200 GPUs.

AI inference market demand

  • ▪Kog CEO Gaël Delalleau reported that the startup received 200 tangible business leads following its May 2026 technology preview.
  • ▪Kog is targeting professional users of AI workflows, such as software engineers, who experience delays with existing tools like Claude Code.
  • ▪Kog CEO Gaël Delalleau stated that prospective customers are not prepared to fine-tune small models, prompting the startup to focus on accelerating larger models.

GPU acceleration methodology

  • ▪Kog CEO Gaël Delalleau utilizes a low-level reverse-engineering approach inspired by his background in solid-state physics and offensive cybersecurity.
  • ▪Kog dedicates several weeks or months of engineering research to optimize its software for each new GPU model.
  • ▪Kog employs a team of 11 people, which limits the number of GPU chips the startup can optimize for the foreseeable future.

Kog funding timeline

  • ▪Kog raised a seed funding round co-led by Varsity VC, a firm where CEO Gaël Delalleau's former Stribe co-founder Kamel Zeroual is a venture capitalist.
  • ▪Kog is supported by Scaleway, backed by Bpifrance, and is part of the French Tech 2030 program.
  • ▪Kog plans to raise a Series A funding round after demonstrating customer traction with its first major model running at 10x speed.

Development roadmap milestones

  • ▪Kog aims to implement its first major model at 10x speed in September 2026 to demonstrate customer traction.
  • ▪Kog plans to eventually feed its optimization methodology into agent-based pipelines to support a wider range of chips and models.

1 source

Techcrunch
Kog is going deeper to squeeze more inference out of GPUs
View source article

Featured stories

View more in Compute, chips & AI infrastructure

Accelevation raises $540 million in US IPO

Sep 29, 2026 · 2 sources

AMD acquires AI startup World Labs founded by Fei-Fei Li for $8.2 billion

Sep 28, 2026 · 7 sources

OpenAI and Synopsys partner to develop AI model for chip design

Sep 30, 2026 · 3 sources

FTC opens investigation into OpenAI and Anthropic over consumer protection

Sep 30, 2026 · 7 sources

Story comments

Loading comments…

Related entities

France

Related Projects

Cerebras

Topics

Compute, chips & AI infrastructureAI inference (scaling)AI startups

Featured stories

View more in Compute, chips & AI infrastructure

Accelevation raises $540 million in US IPO

Sep 29, 2026 · 2 sources

AMD acquires AI startup World Labs founded by Fei-Fei Li for $8.2 billion

Sep 28, 2026 · 7 sources

OpenAI and Synopsys partner to develop AI model for chip design

Sep 30, 2026 · 3 sources

FTC opens investigation into OpenAI and Anthropic over consumer protection

Sep 30, 2026 · 7 sources