Kog Bets Software Can Unlock Stranded GPU Bandwidth for Inference

French startup Kog is betting that modern datacenter GPUs are substantially underutilized during inference — and that the gap is a software problem, not a silicon one. The company's May Hacker News tech preview demonstrated fast single-request decoding on AMD MI300X and Nvidia H200 hardware and generated 200 concrete business leads, according to CEO Gaël Delalleau — enough enterprise attention to anchor a seed round.

The urgency is legible to anyone building on frontier models. Claude Code users routinely wait hours for multi-step completions, and Anthropic charges a price premium for Claude's Fast Mode — an acknowledgment that speed has discrete dollar value. As agentic pipelines grow more complex, that latency penalty compounds across steps, making inference throughput a first-order product constraint.

What the Kog Inference Engine Demonstrated

Kog's demo centred on the Kog Inference Engine (KIE) running against Laneformer 2B, a purpose-built model with approximately 2 billion parameters that the company has since open-sourced. The result was 3,000 per-request tokens per second — the figure underpinning Kog's "30x faster LLM inference" claim. The critical caveat: that number was achieved with a small, purpose-built model. Kog has not yet demonstrated equivalent throughput against a full-scale LLM. Delalleau is targeting a first major model implementation at 10x speed in September, which he has tied directly to Series A readiness.

The methodology is deliberately low-level. Delalleau — an École Polytechnique solid-state physics graduate and four-time DEFCON CTF finalist — describes it in terms borrowed from offensive security: reverse-engineering hardware at the assembly and binary level to exploit behaviour the chip was not explicitly designed for. For each new GPU generation, Kog commits several weeks to months of engineering research before the KIE can target that hardware. With a team of 11, that per-chip cost puts a hard ceiling on how many architectures the company can support near-term.

The Bandwidth Argument

Delalleau's rebuttal to GPU skeptics centres on memory bandwidth trends. His position: successive GPU generations have shipped with increasing memory bandwidth that existing software stacks leave untapped during the decode phase of inference — the stage that dominates agentic workloads, where the model generates one token at a time. Decode is memory-bandwidth-bound, not compute-bound, which is why purpose-built inference chips such as those from Cerebras (which debuted in a May IPO) can look attractive on paper. Kog's counterargument is that software-level optimisation can close much of that gap on hardware enterprises already own.

Kog is not alone in this pursuit. French peer ZML released software that bypasses Nvidia's CUDA layer to enable fast inference across competing chips. Delalleau draws a different comparison — to Stanford's Hazy Research lab — and positions Kog's work as operating at an even deeper layer of GPU microarchitecture.

Target Market

Early demand has come primarily from software engineering workflows and from platforms that let users generate games and applications from prompts. Delalleau notes that prospective customers are not prepared to fine-tune small models, which explains why Kog pivoted post-launch toward accelerating larger model support rather than pushing customers toward Laneformer 2B directly. For generative app and game builders, faster outputs translate directly to higher user throughput.

Funding and the September Proof Point

Kog's seed round was co-led by Varsity VC — whose general partner Kamel Zeroual was Delalleau's co-founder at Stribe, a TechCrunch50 2009 participant — alongside Bpifrance and the French Tech 2030 program. Scaleway is listed as a supporting partner.

The September milestone is explicitly framed as the precondition for the Series A. "Once we've implemented our first major model at 10x speed, which I think will be in September, we'll be able to start demonstrating customer traction and from there, raise our Series A," Delalleau said. Without a live LLM running at meaningful speed multiples, Kog cannot demonstrate the customer traction required to raise the next round.

As inference costs accumulate across multi-step agentic pipelines, any credible software path to reclaiming stranded memory bandwidth on existing hardware carries real economic leverage. Whether Kog's September demonstration holds against a production-scale LLM will determine whether its physics-and-hacking methodology becomes a fundable infrastructure business.