See It Running.

Unlock Trapped GPU Capacity

At scale, AI compute capacity is a software problem as much as a hardware one. Most servers are tuned for peak performance on day one and never revisited.

vert-suite continuously fits GPU and CPU speeds to the actual workload and reclaims around 35% of a server's power budget. That means 33% more GPUs per rack on your existing electrical connections: roughly one extra GPU for every three already deployed, with no new hardware, no model retraining and no application rewrites.

Autonomous AI Agent, 40% Less Power

Claude Code runs an autonomous self-improvement loop, based on Andrej Karpathy's “Autoresearch” pattern, against a local GPU server: an RTX Pro 6000 Blackwell running vLLM and Gemma 4 31B. Without optimisation, the server draws 800 W+ at 19–22 tokens per second.

With vert-suite on, power falls to about 450 W for a 10–14% drop in speed. The wall-socket power meter in the bottom right shows it live: optimisation on, off, then on again. Raw vLLM logs run in the top left.

Portable Proof of Concept on AWS

One command takes a raw, Blackwell-backed AWS EC2 instance and sets up vert-suite from scratch. It installs Helm and the NVIDIA GPU Operator, checks that the GPU is exposed and deploys our core services, then opens k9s to confirm the node is healthy.

The same tool works on virtually any other cloud provider.

Multiple Parallel Coding Tasks on AWS

Multiple parallel coding tasks running LLM inference on an AWS EC2 instance with a single Blackwell GPU. Without optimisation, the GPU stack draws around 375 W. With vert-suite enabled, the draw drops to roughly 300 W, as our eBPF framework tunes hardware power in the background.

Cutting-Edge R&D

Verticular started with £300 of second-hand HP Z820 and a single RTX 3060. Today, the work sits at the cutting edge of the Linux kernel and Kubernetes orchestration for LLMs.

The video shows the GPU-to-CPU pressure on our Blackwell server across different performance settings, overlaid with vert-suite's dynamic optimisation. vert-suite detects the operational phases (three back to back here) on its own, with no labelled data.

Talk at Isambard-AI

Our presentation at Isambard-AI, at an event co-hosted by the Bristol Centre for Supercomputing (BriCS) and NVIDIA, on the growing but often silent GPU energy crisis in data centres.

It explains why cloud resource orchestration is not the same as true optimisation, and how vert-suite's lightweight, eBPF-based architecture tunes GPU and CPU frequencies on the fly to cut energy per token by up to 33% without breaking SLAs.