Expanse, an AI infrastructure startup, has emerged from stealth with a £4 million seed round led by Crane Venture Partners, with participation from PXN Ventures and angel investors including former DeepMind researchers and leaders from AI infrastructure teams.
Before any AI workload runs, engineers have to estimate how many GPUs and CPUs it will need, along with how much memory and how long it will take. Those estimates decide how billions of dollars of infrastructure are allocated every day. Expanse's software predicts what a workload requires before it runs and recommends the best GPU, CPU, memory and runtime configuration before any resources are committed. When jobs are sized correctly in advance, organisations get more work done on the infrastructure they already own. They see fewer failed jobs and higher utilisation, and they need to buy less extra compute. Monitoring and observability platforms explain what happened after a workload has finished, but Expanse works before the job starts. Its models are installed directly in a customer's cloud or on-premises infrastructure and analyse workloads without any code or telemetry leaving that environment.
GPUs are now one of the world's most constrained resources, yet much of that capacity goes unused. Industry estimates suggest that approximately 30% of cloud spend is lost through over-allocation. Microsoft Research has reported GPU utilisation of around 50% across many of its internal deep learning workloads. New GPU capacity can take months to procure, and power, cooling and planning constraints make data centres increasingly hard to expand. For many organisations, getting more out of existing infrastructure is faster, cheaper and more practical than adding hardware.
Four University of Edinburgh graduate engineers founded the startup: Ismaeel Bashir, Nikodem Bieniek, Eren Mendi and Yafet Melake. All four previously built and operated large-scale compute infrastructure in quantitative finance and at national supercomputing facilities. While at the Edinburgh Parallel Computing Centre (EPCC), Bashir developed a multimodal HPC resource prediction system that set a new benchmark for predicting resource requirements across HPC workloads. In one production deployment, Expanse identified nearly £6 million of idle compute capacity in a single month.
The funding will go towards expanding the engineering team and speeding up product development. It will also help bring the platform to more organisations across AI infrastructure, quantitative finance, life sciences, research and high performance computing.
Today, every AI workload begins with an educated guess. Engineers shouldn't have to predict exactly how much compute their code will need before they press run. The machine should carry the uncertainty, not the person.
The future of AI won't be defined only by who builds the biggest clusters, but by who uses them most intelligently. Expanse gives organisations something they don't have today: certainty. By predicting exactly what every workload needs before it runs, customers recover existing capacity, reduce waste and extend the life of some of the world's most valuable infrastructure. We believe that's a new and important category in AI infrastructure.
The widespread adoption of AI by businesses and everyday users has led to an unprecedented focus on the infrastructure required to service this technology and many of the challenges that engineers face. As Expanse has identified, one of those challenges is optimisation and helping businesses to make better use of their compute resource, which can lower running costs and increase scalability. PXN is proud to be backing Ismaeel and the team, who are tackling this head on.
Everyone, from hyperscalers to start-ups, struggles to make efficient use of their most valuable resource: GPU hours. Mapping a constant stream of workloads across shared infrastructure with competing priorities and unpredictable computational demands remains one of the hardest problems in AI infrastructure.
