French startup Kog is challenging the conventional wisdom that graphics processing units are unsuitable for advanced artificial intelligence tasks. The company claims that substantial performance gains can be extracted from standard datacenter GPUs through sophisticated software optimization, potentially reshaping how organizations approach AI inference workloads.
Kog gained significant attention in May after demonstrating that rapid language model processing is feasible on commercially available chips, specifically AMD and Nvidia’s latest datacenter processors. The company subsequently received over 200 business inquiries, indicating strong market interest in its approach. CEO Gaël Delalleau identified software engineering as the most promising initial application, targeting professionals frustrated by lengthy processing times for AI-assisted tasks.
The startup’s founder brings an unconventional background to the challenge, combining formal training in physics with cybersecurity expertise. This combination has shaped Kog’s deep-dive methodology for GPU engineering, though it comes with resource constraints. With an 11-person team, the company can only support a limited number of processors and models in the near term, though Delalleau anticipates a more automated approach will eventually expand this capability.
Kog faces a critical milestone ahead. The startup must demonstrate its optimization techniques work effectively with large language models, not just smaller experimental versions. Delalleau expects to showcase a tenfold speed improvement on a major model by September, a breakthrough that could substantially strengthen the company’s position for future funding rounds.
Originally reported by Techcrunch →
About the Author
