Kog is a French startup building software that aims to speed up AI inference, the process of running a trained model to generate answers, on the graphics processing units that data centers already own. Its product, the Kog Inference Engine, targets fast response times for a single request rather than only high overall throughput across many requests at once.
The performance claims behind Kog
Kog says its software can extract up to thirty times faster large language model inference from standard datacenter GPUs such as the AMD MI300X and Nvidia H200, without requiring custom chips. In a public technology preview, the company demonstrated roughly three thousand tokens per second on a single request, though that figure came from a purpose-built two-billion-parameter model rather than a mainstream large language model, and the broader thirty-times claim remains unproven on production-scale models as of its 2026 preview.
Founder and team
Kog was founded by Gael Delalleau, who studied solid-state physics at the Ecole Polytechnique before working in offensive cybersecurity and competing as a four-time finalist at the DEFCON hacking competition. The company is based in France and employed around eleven people at the time of its 2026 technology preview.
Funding and backers
Kog raised a seed round co-led by Varsity VC, whose partner previously co-founded an earlier startup with Delalleau, with additional backing from Bpifrance and the French Tech 2030 program. The company plans to pursue a Series A funding round once it can show measurable customer traction from its technology preview.
Why GPU inference speed matters
As companies deploy more AI features, the cost and speed of running those models on GPU hardware has become a major expense and bottleneck, since data centers often cannot buy specialized AI chips fast enough to meet demand. Software that squeezes more performance out of GPUs already installed, rather than requiring new hardware, offers a potentially cheaper and faster path to scaling AI applications.
FAQs
What does Kog build? Kog builds software called the Kog Inference Engine that aims to make AI inference run faster on existing datacenter GPUs.
How much faster does Kog claim to make inference? Kog claims it can achieve up to thirty times faster large language model inference, though that figure was unproven on mainstream models as of its 2026 technology preview.
Who founded Kog? Kog was founded by Gael Delalleau, a former offensive security researcher and DEFCON finalist based in France.
Does Kog need special hardware? No, Kog is designed to run on GPUs that companies already own, such as the AMD MI300X and Nvidia H200, rather than requiring custom chips.
More in Artificial Intelligence
Key facts
- Founded by Gael Delalleau, a former offensive-security researcher and DEFCON finalist, based in France.1
- Builds the Kog Inference Engine, aimed at fast single-request AI inference on existing GPUs.1
- Claims up to 30x faster LLM inference on GPUs such as the AMD MI300X and Nvidia H200.1
- Seed round co-led by Varsity VC with backing from Bpifrance and French Tech 2030.1
Quick checkWhat problem is Kog trying to solve?Show answer
Kog is trying to make AI inference run much faster on the GPUs that data centers already own, instead of requiring companies to buy new specialized AI hardware.
Sources
- TechCrunch, Kog is going deeper to squeeze more inference out of GPUs, https://techcrunch.com/2026/08/14/kog-is-going-deeper-to-squeeze-more-inference-out-of-gpus/
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 9, 2026Researched and drafted with AI assistanceReport an issue