Skip to main content

Zettafleet

Category: AI Infrastructure

An end-to-end platform for training large language models (LLMs) that enables distributed training across multiple datacenters without requiring high-bandwidth interconnects like NVLink. Zettafleet was founded in 2024. The company is led by Tomas Paulik. Based in London, United Kingdom. Total funding raised: $0. Latest round: No external funding disclosed.

Founded
2024
Headquarters
London, United Kingdom
Total funding
$0

Value proposition

Enables training of state-of-the-art LLMs at a fraction of traditional cost by automatically provisioning GPU clusters and utilizing compute across multiple datacenters without requiring NVLink or InfiniBand infrastructure, eliminating manual cluster provisioning and orchestration.

Products and solutions

Data Processing Pipeline (S3 and Hugging Face dataset integration with tokenizer support), Pre-training Platform (experimental), Continued Pre-training Solution, Post-training Tools (Reinforcement Learning, Supervised Fine-Tuning, Direct Preference Optimization), Zettafleet CLI for fleet management, Real-time Training Monitoring Dashboard

Unique value

World's first production-grade platform offering distributed LLM training that can utilize compute across multiple datacenters for a single model training run without dependency on high-bandwidth communication links (NVLink/InfiniBand) between AI accelerators.

Target customer

AI research labs, enterprises building custom LLMs, organizations with proprietary data requiring custom model training, and teams seeking to reduce dependency on Nvidia's hardware ecosystem

Industries served

Artificial Intelligence, Machine Learning Infrastructure, Cloud Computing, Enterprise Software, Deep Learning Research, AI Model Development

Technology advantage

Proprietary distributed training technology that eliminates the need for Nvidia's specialized NVLink infrastructure, enabling organizations to train LLMs using geographically distributed compute resources. The platform automatically handles GPU cluster provisioning, code distribution, and training orchestration while providing real-time monitoring of loss curves and perplexity across all nodes and rounds.

How they differentiate

Only platform offering distributed LLM training across multiple datacenters without requiring high-bandwidth interconnects (NVLink/InfiniBand), enabling training using geographically distributed compute resources and reducing dependency on Nvidia's hardware ecosystem

Main competitors

Together AI, Anyscale, MosaicML (Databricks)

Key partnerships

Cambridge University Machine Learning Systems Lab (CaMLSys) - research origin, University of Cambridge Department of Computer Science and Technology, EPSRC Centre for Doctoral Training in Sensor Technologies and Applications

Major milestones

Company incorporated April 11, 2024 in London, Developed world's first 1.3B parameter LLM trained using distributed approach across multiple datacenters without NVLink dependency, Publicly launched platform on February 23, 2026, Featured in Sifted article as stealth AI startup challenging Nvidia's dominance

Growth metrics

Pre-revenue, early-stage startup; publicly launched platform February 23, 2026 after emerging from stealth mode; team size under 10 people

Market positioning

Early-stage stealth startup emerging from Cambridge University research, positioning as Nvidia-alternative in distributed AI infrastructure; pre-revenue with production platform launched Feb 2026

Geographic focus

United Kingdom (London headquarters), with technology designed for global multi-datacenter deployment across any geographic locations

Patents and IP

No registered patents publicly disclosed as of latest update

About Tomas Paulik

PhD Candidate in Machine Learning at University of Cambridge (2023-present), member of Cambridge Machine Learning Systems Lab, researcher specializing in sensor-based edge computing and federated learning

Official website: