5 UK startups solving for AI's energy problem

From ballooning training costs to overheating data centers, AI’s growing appetite is straining the infrastructure meant to support it. By some estimates, models like GPT-4 consume enough electricity to power thousands of homes for months, while daily inference workloads now rival the original training in power demand.

While Big Tech scrambles to expand capacity and governments weigh regulatory guardrails, a new wave of UK startups is offering a different answer: make AI less hungry.

These five ventures aren’t just optimizing around the margins - they’re rewriting the rules of compute, storage, and cooling from first principles. Here’s who to watch - from our deep dive into compute & energy (will push out more soon).

Literal Labs

In a world dominated by neural networks, Literal Labs is taking a contrarian bet: ditching biological mimicry for mathematical precision. Their models, based on Tsetlin Machines and propositional logic, offer up to 54× faster inferencing and 52× lower energy use than standard deep learning architectures. Designed for edge environments where power and explainability matter most - like healthcare, defence, or IoT - their compact models are transparent, auditable, and cost-effective.

Graphcore

Few names loom larger in the UK's AI hardware ecosystem than Graphcore. The company’s Intelligent Processing Unit (IPU) is a radical departure from traditional GPUs, designed specifically for graph-based workloads like transformers. By fusing compute and memory into a single architecture, Graphcore eliminates the bottlenecks (and heat spikes) common in legacy designs.

Vaire

Vaire doesn’t just want to cool chips. It wants to redefine physics under the hood. The startup is developing reversible computing hardware that recycles energy typically lost as heat during processing. Instead of one-way logic gates that discard data, Vaire’s chips use reversible logic to retain and reuse inputs - slashing thermal output and power use.

Iceotope

AI may be smart, but it’s hot-headed. Enter Iceotope, a pioneer in liquid immersion cooling that removes 100% of heat generated by high-density compute - with no fans, no water waste, and up to 40% less power consumption per kilowatt of IT equipment.

Deep Green

What if data centers didn’t just consume energy - but gave some back? That’s the radical idea behind Deep Green, which installs modular compute units in public swimming pools, housing blocks, and industrial buildings and repurposes waste heat for free, zero-carbon heating.

AI may be rewriting the rules of cognition, but its real constraint isn’t intelligence - it’s energy. As the world rushes toward ever-larger models and faster inferencing, these UK startups are asking the more interesting question: what if we just did more with less?

From cooling to chips to compute paradigms, they’re proving that the future of AI won’t just be smarter - it’ll be leaner, cleaner, and a whole lot cooler.