Pomegra
Edinburgh's Expanse secures $5.3 million in seed funding led by Crane Venture Partners to deploy predictive software that calculates AI workload resource needs before a single job runs.
- Expanse raised $5.3M in seed funding led by Crane Venture Partners, with PXN Group, AIP Seed, and DeepMind alumni angels also participating.
- The company's software pre-calculates GPU, CPU, and memory requirements before AI workloads execute, eliminating costly over-allocation at the point of scheduling.
- In one early production deployment, Expanse identified nearly $8M of idle compute capacity within a single month.
Lead
Expanse, a University of Edinburgh spinout founded in October 2025, announced a $5.3 million seed round on September 16, 2026, led by Crane Venture Partners. The raise - also backed by PXN Group, AIP Seed, and a group of angels including former DeepMind researchers and AI infrastructure leaders - funds a product that sits inside a customer's compute environment and tells operators what a workload will need before it runs, not after it has already consumed resources and failed.
What Does Expanse Actually Do?
The company's software analyzes code and hardware telemetry, builds deep learning models of workload behavior over time, and generates per-job predictions of GPU count, CPU allocation, memory requirements, and runtime. Those predictions are fed back into the scheduler before execution begins. The model sharpens with every completed job, meaning accuracy improves as the system accumulates a facility's workload history.
The distinction from conventional observability tooling is deliberate: most monitoring platforms tell engineers what happened after a job finishes - after the wasted capacity has already been booked and charged. Expanse works at the planning stage, before the reservation is made.
Critically, the software installs directly into a customer's cloud or on-premises environment. Code and telemetry do not leave the deployment boundary, which matters to the research institutions, quantitative finance firms, and large-scale AI labs the company is targeting.
Why Does GPU Over-Allocation Persist?
The underlying problem is structural. Before a workload runs, an engineer must specify how many GPUs it will use, how much memory it needs, and how long it will take. Those estimates determine how billions of dollars of infrastructure are allocated daily. Get them wrong in one direction and jobs fail mid-run; get them wrong in the other and GPUs sit idle at full reservation cost.
Industry figures put over-provisioning losses at roughly 30% of cloud spend. Microsoft Research has separately reported average GPU utilization near 50% across internal deep learning workloads. The practical consequence is that organizations are paying for infrastructure that is not running useful computation - a gap that has widened as AI job complexity has grown faster than the tooling built to schedule it.
How Bad Is the Waste in Practice?
The production numbers Expanse has published suggest the problem is not marginal. In one deployment, its software surfaced $7.9 million in idle compute capacity inside a single month. That figure comes from a single customer environment, which implies the aggregate waste across the industry is substantial enough to attract serious capital even at the seed stage.
The founding team has direct operational exposure to this failure mode. Ismaeel Bashir, one of the four co-founders, worked at the Edinburgh Parallel Computing Centre (EPCC) and developed a multimodal HPC resource prediction system that set benchmarks for estimating requirements across HPC workloads. Co-founders Nikodem Bieniek, Eren Mendi, and Yafet Melake bring experience from quantitative finance infrastructure and national supercomputing facilities.
Competitive and Market Context
The AI infrastructure tooling market has attracted considerable attention as enterprises push GPU spending higher and begin demanding better utilization metrics from their providers. Expanse competes at a different layer from observability platforms and cloud cost management tools: its pitch is prediction, not retrospective analysis or billing optimization.
That framing positions the company against the scheduler-adjacent tools built into existing HPC and cloud orchestration stacks, none of which currently perform pre-run workload modeling at the fidelity Expanse claims. Whether the deep learning approach to resource forecasting holds up at enterprise scale across heterogeneous workload types remains the core technical question the seed capital will need to answer.
Crane Venture Partners, the round's lead, focuses on enterprise software infrastructure in Europe. Its participation signals that the firm sees the prediction-before-execution angle as sufficiently differentiated from the monitoring tools already in the market.
Outlook
Expanse enters the market with a narrow, well-defined thesis: eliminate the estimation step entirely by replacing it with a learned model. The $5.3 million gives the four-person team roughly 18 to 24 months to broaden its production deployments, demonstrate accuracy across workload types, and build the sales infrastructure necessary to reach enterprise buyers before better-funded competitors close the differentiation gap. The GPU waste problem is real and quantified. The question is whether a pre-run prediction layer becomes standard infrastructure or gets absorbed into the platforms it currently sits beside.
