French startup Kog says your company's existing GPUs can run AI ten times faster. Here's how.
Kog claims software alone can dramatically speed up AI inference on the chips businesses already own. The September demo will be the real test.
Key points
- French startup Kog drew 200 business leads in May 2025 after a public demo showing 3,000 tokens per second on standard data-centre GPUs.
- Its small test model, Laneformer 2B, has been open-sourced, but Kog's real target is the large language models that power everyday AI tools.
- CEO Gaël Delalleau told TechCrunch a first major model running at 10x speed should be ready to demonstrate by September 2025, ahead of a planned Series A fundraise.
- Kog's seed round was co-led by Varsity VC and backed by French public investors Bpifrance and the French Tech 2030 programme.
When you run AI software, every answer you get requires a process called inference: the model reads your question and generates a reply, word by word. The faster that happens, the less you wait and the less it costs. That speed is a genuine bottleneck for many businesses right now.
Kog, a Paris-based startup, thinks the bottleneck isn't the hardware. It's the software running on top.
What is Kog actually claiming?
Kog says it can make AI inference dramatically faster without swapping out any chips. Its May demo, which caught wide attention on the programmer forum Hacker News, showed a small AI model producing 3,000 tokens (roughly 3,000 words or word-fragments) per second on AMD MI300X and NVIDIA H200 GPUs, the kind of chips large companies already run in their data centres.
That figure came from a purpose-built model called Laneformer 2B, which has about 2 billion parameters (think of parameters as the learned settings inside an AI model that shape its answers). Small by today's standards. The company has open-sourced it, but the hard commercial question is whether the same tricks work on the far larger models powering tools like Claude or ChatGPT.
Delalleau told TechCrunch he's confident they do, pointing to newer GPUs carrying more and more memory bandwidth that current software simply doesn't exploit. He expects to demonstrate a major large model running at ten times its normal speed by September 2025, a milestone he says will trigger the company's Series A raise.
It's worth comparing that target to what's already shipping elsewhere. Our earlier story on OpenAI's new Ultrafast mode making GPT-5.6 Sol run at 14 times normal speed shows the market Kog is entering is moving fast.
Who would actually benefit?
Practical gains would land first with software engineers and AI-heavy workflows. Users of Claude Code, Anthropic's coding assistant, sometimes wait hours for results. Anthropic already sells a Fast Mode at a premium price because speed has real value. Kog is pitching itself as a way to get that speed without the premium, using infrastructure a company already owns.
Design partners include platforms that let users generate games and apps from a text prompt. For those products, a faster result isn't just a nice feature; it directly affects how many users pay.
Kog isn't alone in this space. ZML, also French, released software that can bypass Nvidia's CUDA (the standard software layer that tells Nvidia chips how to work) to speed up inference across different hardware. Delalleau positions Kog differently, closer to Stanford's Hazy Research lab, with an even lower-level focus on GPU engineering, down to assembly language and binary code.
| Fact | Detail |
|---|---|
| Demo speed | 3,000 tokens per second, single request |
| Test model | Laneformer 2B (2 billion parameters, open-sourced) |
| Target milestone | 10x speed on a major large model, September 2025 |
| Business leads from demo | 200, as of May 2025 |
| Team size | 11 people |
| Seed backers | Varsity VC, Bpifrance, French Tech 2030, Scaleway |
That low-level work is slow. For every new GPU model, Kog spends weeks to months studying how the chip actually behaves before writing a single line of optimisation code. With 11 people, that limits how many chips the company can support at once.
Kog's real test isn't the demo; it's whether the company can prove the approach scales before a better-funded rival does.
Common questions
Can this speed up the AI tools I already use at work?
Not yet for most people. Kog sells to companies running their own AI infrastructure, not to individual users. If a company deploys Kog's software on its servers, you might notice faster results without knowing why.
What does "inference speed" actually mean for me?
Every time an AI tool answers a question or writes code, that's one inference. Faster inference means shorter waits and lower running costs for the company providing the tool, which can translate to lower prices or better products over time.
Is Kog's big claim proven yet?
Not on large models. The 3,000-token-per-second result used a small, specially built model. Delivering comparable gains on the much bigger models that power mainstream AI tools is the company's next, harder test.
