Kog

What Kog does

Kog (legal name: Kog Labs; DBA: Kog AI) is a Paris-based AI infrastructure startup founded in 2023 by Gaël Delalleau, an École Polytechnique engineer with a background in cybersecurity research and high-performance GPU work. The company builds a hardware-software co-designed inference engine — the Kog Inference Engine (KIE) — purpose-built to eliminate software-imposed speed ceilings on standard datacenter GPUs. KIE integrates a proprietary Laneformer Transformer architecture featuring Delayed Tensor Parallelism (DTP), a single-kernel 'monokernel' runtime, the custom Kog Communication Library (KCCL), and topology-aware GPU memory access; all components are written from scratch in low-level CUDA/HIP and assembly. The system achieves 3,000 output tokens/s per request on 8× AMD MI300X and 2,100 on 8× NVIDIA H200 at batch size 1 in FP16, without speculative decoding, and is exposed as a drop-in replacement for vLLM.

Kog monetizes via API access sold to teams building AI coding agents and agentic workflows, supplemented by a high-touch Design Partner Program for enterprise engagements; no public pricing tiers are disclosed. The company also publishes the Laneformer 2B (2.3B-parameter) coding model on Hugging Face as both a usable checkpoint and a developer-acquisition channel. Distribution is API-first with a live public playground (playground.kog.ai), a technical blog (blog.kog.ai), and featured presence at AMD AI DevDay 2026. The company is backed by Varsity VC and BPI France's Deep Tech Program ($5M seed) and was awarded the French Tech 2030 label by the French government in October 2025; current operations are anchored by an 11-person team.

Name

Kog

Legal name

Kog Labs

Website

https://kog.ai

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

11–50 employees

Short description

Kog is a Paris-based AI infrastructure startup (founded 2023) building a hardware-software co-designed inference engine that delivers 3,000 tokens/s per request on AMD MI300X GPUs for AI coding agents and agentic workflows, via a proprietary monokernel runtime, custom KCCL communication library, and Laneformer architecture.

Ownership category

akta.pro rank

Product category

AI Inference Infrastructure

NAICS

Software Publishers (5132), Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Services (51821), Custom Computer Programming Services (541511)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

Model Deployment, Serving & Inference Platforms (HDAAABAF)

akta.pro secondary industries

AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers) (HDAAAAAI), Model Serving, Inference & Deployment Platforms (APIs, Edge/On-Prem) (HDAEANAC), Model Hosting, Serving & Inference Platforms (HDAAACAB)

Keywords

• Real-time LLM inference
• AI inference engine
• GPU optimization software
• AI coding agents
• Low-latency inference

Headquarters

HQ city

Paris

HQ country

France

HQ region

Europe

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel

Revenue model

1. API Access / Inference-as-a-Service: Kog likely generates revenue by providing API access to its high-speed inference engine. The website prominently features 'Request API Access' CTAs and a public tech preview playground, suggesting a consumption-based or subscription API pricing model targeting developers and enterprises building AI agents. No specific pricing tiers are publicly disclosed.

Pricing tiers

Go-to-market motion4 records

Distribution channels3 records

Marketing channels6 records

Core offering

Kog builds a hardware-software co-designed inference engine (KIE) for ultra-low-latency LLM inference on standard datacenter GPUs (AMD MI300X and NVIDIA H200), achieving 3,000 output tokens per second per request. The platform combines a custom Laneformer model architecture with Delayed Tensor Parallelism (DTP), a monokernel GPU runtime, and a custom KCCL communication library, exposing inference through a vLLM-compatible API and a live playground. The offering targets enterprises building AI coding agents and agentic workflows where token generation speed governs agent iteration throughput.

Product overview

Kog is a Paris-based AI infrastructure startup offering a unified inference platform centered on the Kog Inference Engine (KIE), a hardware-software co-designed system for ultra-low-latency LLM inference. The core product integrates the Laneformer model architecture with Delayed Tensor Parallelism (DTP), the custom Kog Communication Library (KCCL), and a monokernel runtime. The platform includes Kog Labs for research/blog content and a live Playground demo. Users can access the inference engine via API request or through the drop-in vLLM-compatible interface, targeting AI coding agents and agentic workflows that require real-time token generation speeds of 3,000 tokens/s per request.

Differentiator

Problem solved

Functional benefit

Brands

• Kog Inference Engine (KIE): Hardware-software co-design inference engine enabling 3,000 tokens/s per request on AMD MI300X GPUs.
• Laneformer
• Kog CommunicationLibrary (KCCL)
• Kog Playground

Products and services

• Kog Inference Engine (KIE) A hardware-software co-designed inference engine built to run GPUs at their absolute ceiling, delivering 3,000 tokens/s per request via a monokernel runtime, custom KCCL communication library, and Laneformer architecture. Functions as a drop-in vLLM replacement with no code refactoring required, targeting enterprise teams building AI coding agents and agentic workflows.
• Laneformer 2B A 2.3B-parameter instruction-tuned coding model built around Kog's Delayed Tensor Parallelism architecture, designed for AI coding agents and software engineering workflows. Achieves 45.1% on HumanEval+ and 51.6% on MBPP+ while running at 3,000 tokens/s on AMD MI300X.
• Kog Laneformer Architecture A novel Transformer architecture variant in which inter-device communication is delayed by one layer so that compute runs continuously without synchronization pauses, enabling linear scaling across high-end GPUs for low-latency inference.
• Kog Communication Library (KCCL) A custom collective communication layer that replaces standard NCCL/RCCL to unlock linear scaling for tensor parallelism across high-end GPUs, achieving sub-3 microsecond AllReduce latency on AMD MI300X.

Quantifiable outcome

• 3,000 output tokens/s/request on AMD MI300X and 2,100 on NVIDIA H200 at batch size 1, FP16, no speculative decoding
• +6 more outcomes

Segments2 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration2 records

AI capability5 records

Feature8 records

Partnerships

One partnership is on record.

• AMDcoreTechnology or IntegrationAMD featured Kog as an ecosystem partner at AMD AI DevDay 2026 and published Kog's inference benchmark results on AMD's official engineering blog, validating Kog's 3.5x speed improvement on AMD Instinct MI300X GPUs. AMD provides the primary hardware platform for Kog's performance differentiation.

Scale indicators6 records

Recent moves6 records

Expansion highlights6 records

Direct peers

• Modal Labs: Modal provides serverless GPU compute and inference infrastructure for AI applications with a developer-first, Python-native model. Comparable in targeting AI agent builders with low-friction GPU access and inference APIs.
• Replicate: Replicate runs a cloud API for running open-source ML models, including LLM inference, with a focus on simplicity and developer ergonomics. Comparable in API-first inference distribution to developers building AI applications.
• Groq: Groq sells LPU-based ultra-low-latency inference-as-a-service, directly competing on tokens-per-second-per-request as the primary value proposition. Targets developer and enterprise agentic workloads with similar speed-first positioning.
• DeepInfra: DeepInfra provides serverless inference APIs for open-source LLMs with custom inference engine optimizations. Directly comparable API-first inference-as-a-service business model targeting developer adoption.
• Together AI: Together AI runs an inference cloud with proprietary optimizations on multi-GPU stacks, including custom kernels and inference engine work. Comparable as an open-model inference API competitor targeting developer and enterprise agent builders.
• Lepton AI: Lepton AI offers a cloud platform for running AI models with custom inference optimizations on multi-GPU hardware. Comparable as a developer-facing inference platform emphasizing performance and flexible deployment.
• Anyscale: Anyscale (Ray) offers distributed compute and serving infrastructure optimized for AI workloads, including LLM inference. Comparable as a developer-focused platform for scaling and serving large models on multi-GPU clusters.
• Fireworks AI: Fireworks AI operates a model inference and fine-tuning platform optimized for low-latency serving of open-source LLMs. Directly comparable in API-first inference distribution and agent workload targeting.

Broad incumbents

• Hugging Face: Hugging Face operates Inference Endpoints and text-generation-inference (TGI), serving open-source models via API. Comparable as an inference platform serving the same open-model ecosystem, but as a broad incumbent rather than a speed-specialized player.

Emerging players

• Cerebras Systems: Cerebras builds purpose-built inference silicon (CS-3/WSE) and a cloud inference service claiming class-leading tokens/second for LLM serving. Comparable in the speed-first inference narrative, but pursues a custom-silicon route rather than Kog's software-only approach on commodity GPUs.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

• LinkedInlinkedin.com/company/kogai

Revenue estimate

Valuation estimate

Number of profiles

Profiles1 record

Funding overview

Funding rounds1 record

Investors1 record

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Kog