Executive Overview
As the artificial intelligence industry accelerates past the training phase and enters the high-stakes era of widespread deployment, the race for faster AI inference has become the defining technological bottleneck of the decade. While hardware startups and specialized chip manufacturers—exemplified by Cerebras and its purpose-built architecture drawing massive market enthusiasm during its initial public offering—are banking on novel hardware to slash processing times, a quiet revolution is taking shape in Europe.
Enter Kog, an ambitious French deep-tech startup betting that the true limits of standard, off-the-shelf datacenter GPUs have barely been scratched. Eschewing the need for exotic, purpose-built hardware, Kog is pioneering a radical software-driven optimization strategy. By treating traditional datacenter accelerators like the AMD MI300X and Nvidia H200 not as static endpoints, but as programmable canvases of untapped memory bandwidth, the startup aims to extract performance increases that defy conventional engineering wisdom.
In May, Kog thrust itself into the global spotlight, capturing the front page of Hacker News with a technical preview that demonstrated blistering single-request decoding speeds of 3,000 tokens per second (TPS) on standard hardware. Utilizing its proprietary Kog Inference Engine (KIE) and an open-sourced, purpose-built small language model dubbed Laneformer 2B, the company proved that extreme low-latency inference is achievable on the exact enterprise infrastructure already sitting in datacenters worldwide.
Yet, Kog’s journey is only beginning. Founded by veteran entrepreneur Gaël Delalleau, the 11-person startup operates at the intersection of solid-state physics and offensive cybersecurity—a unique mindset that allows its engineers to reverse-engineer hardware down to the assembly and binary levels. As the company targets scaling its breakthrough speeds to massive, mainstream large language models (LLMs) ahead of a planned Series A funding round, it stands as a standard-bearer for Europe’s growing push toward technological sovereignty. Backed by institutional giants like Bpifrance and French Tech 2030, Kog is positioning itself to answer a critical question: Can clever software render specialized hardware obsolete, or at least push the boundaries of conventional silicon further than anyone thought possible?
Detailed Chronology: From Hacker News Sensation to European Deep-Tech Contender
The narrative of Kog is one of rapid momentum punctuated by deep technical excavation. The timeline of its ascent reveals a calculated strategy designed to disrupt the status quo of AI infrastructure.
The May Breakthrough and Public Debut
The catalyst for Kog’s public prominence arrived in May, a month already marked by historic shifts in AI markets, including the blockbuster IPO debut of hardware rival Cerebras. Eschewing custom silicon entirely, Kog released a technical preview that immediately captivated the developer community, rocketing to the front page of Hacker News.
The preview demonstrated that extremely fast, single-request decoding is entirely feasible on standard datacenter GPUs already owned by major enterprises. Using standard configurations of the AMD MI300X and Nvidia H200, the startup showcased its ability to process data at unprecedented speeds. While some early observers expressed disappointment that these optimizations did not extend to consumer-grade laptop GPUs, enterprise architects and AI engineers immediately recognized the paradigm-shifting potential.
Early Market Validation and Business Inquiries
The ripple effects of the Hacker News debut were immediate. Rather than facing polite indifference, Kog was met with an immediate deluge of commercial interest. According to CEO Gaël Delalleau, the preview generated over 200 tangible business leads from companies desperate to escape the latency bottlenecks plaguing modern AI deployments.
Early feedback pointed directly to software engineering workflows as the prime candidate for immediate deployment. Developers utilizing advanced coding assistants—such as veteran users of Anthropic’s Claude Code—are intimately familiar with the frustrating delays of waiting hours for complex logic loops to compile and execute. Recognizing that time equates to money, industry giants like Anthropic have begun charging price multiples for premium tiers like "Fast Mode." Kog realized it could target this exact frustration, alongside emerging application developers building prompt-to-game and prompt-to-app ecosystems where marginal speed improvements directly correlate with exponential revenue gains.
Shifting Focus to Larger Models
As Kog engaged with its expanding roster of prospective enterprise customers, the startup encountered a sobering market reality: the ecosystem was not yet mature enough for widespread adoption. Specifically, prospective clients expressed a strong reluctance to engage in the complex, resource-intensive fine-tuning of small models.
Responding swiftly to this feedback, Kog pivoted its internal roadmap. Since the initial launch, the startup has dedicated its engineering bandwidth entirely to accelerating the performance of larger, more complex foundational models to meet real-world enterprise demand. This evolution culminated in the open-sourcing of Laneformer 2B, a 2-billion-parameter model designed to prove the viability of Kog’s foundational approach, while the team races toward implementing its optimization methodologies onto heavier, industry-standard LLMs.
Supporting Context & Metrics: The Physics of Hardware and Low-Level Hacking
To understand how Kog achieves its extraordinary performance claims, one must examine the unconventional background of its founder and the rigorous methodology that shapes its engineering culture.
The Intersection of Physics and Offensive Cybersecurity
Gaël Delalleau is not a traditional academic researcher. Though he studied solid-state physics at France’s prestigious École Polytechnique—instilling in him a fundamental respect for the immutable laws of physics governing silicon behavior—his professional formative years were spent in the trenches of offensive cybersecurity. As a veteran white-hat hacker and a four-time finalist at DEFCON’s legendary CTF (Capture the Flag) tournament, Delalleau cultivated a mindset centered on reverse engineering.
"There’s this mindset of understanding the laws of physics, and the laws of the GPU in order to make the most of them," Delalleau explains. Drawing on his cybersecurity background, he adds that low-level hacking taught him "to reverse-engineer things at a very low level — down to assembly language and binary code — to understand how it works, and to try to use it to achieve a goal for which it wasn’t necessarily designed."
This philosophy directly informs the Kog Inference Engine (KIE). While traditional machine learning engineers interact with GPUs through high-level abstractions, frameworks, and standard APIs, Kog’s team digs deep into the bare-metal architecture. They treat memory bandwidth—which modern GPUs possess in abundance—as a resource to be aggressively unlocked by bypassing inefficiencies in standard software stacks.
The Metrics of Speed: Breaking Down the Numbers
The primary metric driving Kog’s value proposition is its demonstration of 3,000 tokens per second (TPS) per request during its initial technical preview. To put this figure into perspective, standard commercial inference engines running comparable small models often hover well below this threshold, constrained by memory latency and inefficient scheduling.
However, industry skeptics were quick to point out a crucial caveat: the 3,000 TPS milestone was achieved using Laneformer 2B, a purpose-built, highly compact model featuring roughly 2 billion parameters. Scaling this level of performance to mainstream large language models containing tens or hundreds of billions of parameters represents an immense technical leap.
The Competitive Landscape: Software-Agnostic Optimization
Kog is not operating in an isolated vacuum. The broader French deep-tech ecosystem has emerged as a hotbed for inference innovation. For instance, fellow French startup ZML recently released free, hardware-agnostic software designed to speed up inference across a diverse array of AI chips while bypassing Nvidia’s ubiquitous CUDA framework.
Despite such regional peers, Delalleau draws a distinction between Kog and other market players. He likens Kog’s granular, deeply technical approach more closely to Stanford University’s renowned Hazy Research lab, emphasizing an unyielding focus on deep-level GPU acceleration rather than broad, surface-level hardware abstraction.
Official Statements and Industry Insights
The discourse surrounding Kog’s technology highlights a profound philosophical divide in the modern AI hardware market: Is custom silicon mandatory for future progress, or can clever, low-level software extract infinite utility from existing infrastructure?
Defending the GPU Paradigm
Amid intense market enthusiasm for proprietary ASICs and specialized inference chips, Delalleau remains bullish on the enduring relevance of conventional graphics processing units.
"GPUs have a bright future," Delalleau asserts, directly challenging the growing narrative that graphics cards are inherently ill-suited for efficient token decoding. In his view, the persistent belief that standard GPUs are bottlenecked for LLM inference is a fundamental misconception. Modern hardware architectures feature sprawling memory bandwidth reserves that simply require intelligent, low-level software orchestration to unleash.
The Bottleneck of Manual Engineering
This extreme, bare-metal approach, however, comes with inherent operational friction. Because Kog’s engineers must reverse-engineer hardware architectures down to assembly and binary code, the process is intensely manual and time-consuming.
"For every new GPU, we’ll dedicate several weeks or even months, to really dig into the details and conduct GPU engineering research on that hardware," Delalleau notes.
With a tight-knit team of just 11 people, this meticulous dedication imposes a strict operational ceiling on the sheer number of distinct chip architectures Kog can support simultaneously in the near term. To overcome this scaling limitation over the long run, the startup plans to integrate its proprietary methodologies into agent-based automation pipelines, allowing the system to autonomously map and optimize support for new chips and evolving model architectures.
Future Outlook: The Road to Series A and European Sovereignty
As the artificial intelligence sector looks toward the horizon, Kog stands at a pivotal juncture of commercial validation and technical execution.
The September Milestone and Series A Horizons
The immediate future for Kog hinges on a definitive proof-of-concept. To transition from a viral Hacker News sensation to a heavily backed enterprise staple, the startup must demonstrate that its software optimization techniques can successfully scale to major, industry-standard large language models while maintaining order-of-magnitude speed improvements.
According to Delalleau, this crucial validation phase is rapidly approaching. The company anticipates implementing its optimization engine on its first major enterprise-grade model at a 10x speed multiplier by September. Successfully clearing this technical hurdle will provide the concrete customer traction required to anchor its upcoming Series A funding round.
Sovereignty Tailwinds and Ecosystem Support
Beyond commercial metrics, Kog is a primary beneficiary of Europe’s broader strategic push for technological independence. As the European Union and national governments aggressively foster local capabilities in AI pre-training, infrastructure, and deployment, Kog enjoys powerful institutional tailwinds.
The startup has already secured vital backing from cloud infrastructure pioneer Scaleway, alongside substantial financial and programmatic support from France’s public investment bank Bpifrance and the prestigious French Tech 2030 initiative. This robust domestic backing ensures that even as Kog competes on the global stage against well-funded Silicon Valley behemoths, it remains deeply anchored in the burgeoning sovereign European AI corridor.
Conclusion
Kog’s bet is audacious: that the limits of AI inference are defined not by the physical constraints of silicon, but by the limitations of the software layers written on top of them. By fusing solid-state physics with the mindset of a low-level security hacker, the French startup is rewriting the rules of GPU utilization. If Kog successfully proves its 10x acceleration metrics on heavy LLMs this autumn, it will not only secure its Series A funding—it will fundamentally reshape how the global enterprise computes the future of artificial intelligence.
