French Startup 

Cerebras proved in May that Wall Street will pay up for purpose-built AI chips, taking its IPO debut to a warm reception. Paris-based Kog is working the opposite side of the same problem. Instead of building new silicon, the eleven-person startup wants to prove that the AMD and Nvidia GPUs already sitting in enterprise datacenters have far more inference speed left in them than anyone is currently pulling out.

A Hacker News debut that turned into 200 leads

Kog's public introduction came in May, when a technical preview claiming "extremely fast single-request decoding" on standard datacenter GPUs hit the front page of Hacker News. The demo ran on AMD MI300X and Nvidia H200 hardware, the kind already deployed across enterprise clusters rather than anything exotic.

Some commenters were let down that the approach didn't extend to consumer laptop GPUs. Others saw the commercial angle immediately. With inference speed now functioning as a hard cost and latency bottleneck across the industry, a pitch to unlock more performance from hardware companies already own found an audience fast. "We had 200 tangible business leads," CEO Gaël Delalleau told TechCrunch.

Software engineering looks like the first real use case, based on the early response Kog has tracked. Anyone who spends serious time in Claude Code knows the wait for a response can stretch into hours on complex tasks, and Anthropic has already priced that pain point into its product by charging extra for a Fast Mode tier. Kog is aiming squarely at customers frustrated by exactly that kind of delay.

The company isn't only chasing coding workflows. Delalleau also pointed to design partners building prompt-to-game and prompt-to-app products, where a faster generation loop translates directly into more revenue per user session.

The gap between the demo and the promise

Here's where Kog's pitch gets harder to take at face value without scrutiny. The company advertises "30x faster LLM inference," and its public demo backs that up with a genuinely striking number: 3,000 tokens per second on a single request. But that figure came from Laneformer 2B, a purpose-built model Kog designed and trained itself, and it holds roughly 2 billion parameters. That's a small model by current standards, built from the ground up around Kog's own architecture rather than a large commercial LLM someone would actually deploy for hard reasoning tasks.

Kog has since open-sourced Laneformer 2B, and the company reports it clears 45.1% on HumanEval+ and 51.6% on MBPP+ under greedy decoding, respectable results for a coding-focused model in its size class but not benchmark-leading against larger general models.

Turning that same speed advantage into a technique that holds up on the much bigger models enterprises actually run in production is a different engineering problem entirely, and one Kog hasn't demonstrated in public yet. Delalleau doesn't dispute that gap. Since the May launch, he said, the team has shifted its full focus toward scaling the approach up to larger models rather than continuing to refine performance on small ones, after learning that prospective customers had little appetite for fine-tuning small custom models themselves.

Delalleau remains confident the underlying method transfers. "GPUs have a bright future," he said, pushing back on the idea that memory bandwidth limits make them a poor fit for decoding. He argues that framing has become outdated as newer GPU generations ship with more bandwidth than current software knows how to exploit.

Reverse-engineering the chip instead of replacing it

Kog isn't the only company betting that software, not new hardware, is the unlock here. French peer ZML has released hardware-agnostic tooling that bypasses Nvidia's CUDA stack entirely to run fast inference across competing chip architectures. Delalleau draws a different comparison for his own work, positioning Kog closer to Stanford's Hazy Research lab, with an even narrower focus on GPU-level acceleration specifically.

That focus traces back to Delalleau's own background rather than a typical research pedigree. He studied solid-state physics at École Polytechnique before moving into offensive cybersecurity, working as a white-hat hacker and reaching the finals of DEFCON's CTF tournament four times. He says that history shapes how he pushes his team to work now, treating a GPU's constraints the way a physicist treats a law of nature rather than a specification to accept at face value.

The hacking background maps just as directly onto the engineering approach. Reverse-engineering something down to assembly and binary code, he said, teaches you to understand a system well enough to use it for a purpose it wasn't built for. That's a fair description of what Kog is attempting with commodity GPUs never marketed as inference specialists in the first place.

This method doesn't scale cheaply. Delalleau said the team commits several weeks to months of dedicated low-level engineering research for every new GPU it wants to support. With a headcount of 11, that puts a real ceiling on how many chip families Kog can cover in the near term, a constraint the company hopes to loosen eventually by feeding its own methodology into agent-based pipelines that could extend support to more models and hardware automatically.

Kog isn't operating without institutional backing while it works through that constraint. The startup is supported by cloud provider Scaleway and has backing from Bpifrance and the French Tech 2030 program, giving it a sovereignty angle as Europe pushes to build independent AI infrastructure capability on both chips and models.

What comes next

Kog's near-term roadmap now hinges on a single proof point. Delalleau expects the team to hit 10x speed on its first major large-scale model implementation by September. That milestone, he said, is what unlocks the next phase: demonstrating real customer traction and using it to raise a Series A.

Until that happens, Kog's 30x claim rests on a small model most enterprise customers wouldn't deploy on its own. Whether the same low-level GPU engineering that produced 3,000 tokens per second on Laneformer 2B holds up at the scale of an actual production LLM is the test the company still has to pass in public.