Kog Startup Aims to Deliver Lightning-Fast AI on Everyday GPUs
The race to speed up AI responses is heating up as companies seek cheaper and quicker ways to run large language models.
Kog a French startup is focusing on software tweaks to get far more performance from standard GPUs already in data centers.
Unlocking Hidden Power in Existing Hardware
This method could cut costs dramatically for businesses that rely on AI tools for coding and creative tasks.
The company demonstrated three thousand tokens per second on a small two billion parameter model using common chips like those from NVIDIA and AMD.
Its leader draws from a background in physics and low level security work to reverse engineer GPU operations in novel ways.
With only eleven team members the firm plans to expand support for more chips through future automated systems.
Broader Implications for the AI Industry
Success here might reduce the need for expensive specialized hardware and lower energy use across the sector.
European backers see this as a step toward greater technological independence in a field dominated by American players.
If the team hits its target of ten times faster speeds on big models by September it could attract serious new investment.
For regular users this means AI assistants that answer instantly instead of making people wait hours for results.
Overall the approach highlights how clever software can extend the life of current technology while the industry evolves.
