Expanse Raises $5.3M to Recover Wasted GPU Capacity

The most expensive GPU in an AI cluster may be the one another engineer cannot use because a cautious estimate is holding it. Resource requests are made before the workload produces evidence, so the person submitting a job absorbs the pain of asking for too little while the wider cluster absorbs the cost of asking for too much.

Expanse has raised a $5.3M Seed round to put a prediction layer in front of that decision. Crane Venture Partners led the financing, with PXN Ventures and additional investors supporting software designed to predict GPU, CPU, memory, runtime, completion probability, and failure risk before a workload enters the queue.

The financing gives the young company more room to hire engineers, accelerate product development, and deploy across AI labs, quantitative finance, life sciences, research computing, and high-performance computing. It also puts Expanse in the middle of a difficult commercial test: whether infrastructure teams will trust a new prediction system enough to lower the protective buffer built into their resource requests.

What Expanse Raised

Expanse announced the Seed round on September 16, 2026. The company-issued release names Crane Venture Partners as lead investor, PXN Ventures as a participant, and unnamed angels that include former DeepMind researchers and AI-infrastructure leaders. Founder disclosures identify a wider syndicate that includes Y Combinator, Standard Capital, Scribble Ventures, AIP Seed, General Advance, Ritual Capital, Robinhood Ventures, and Twenty Two Ventures.

The company was founded in October 2025 by four University of Edinburgh students and alumni. Ismaeel Bashir is co-founder and CEO, Nikodem Bieniek is co-founder and CTO, Eren Mendi is co-founder and CPO, and Yafet Melake is co-founder and COO. The team joined Y Combinator's Spring 2026 batch after working across national supercomputing and quantitative-finance environments.

Expanse did not disclose its valuation, financing terms, ownership changes, investor allocations, revenue, pricing, customer count, or total capital raised. Those blanks matter because the $5.3M announcement describes the resources entering the company, not yet the scale of the business leaving it.

Why Compute Requests Become Expensive Guesses

Shared clusters create an asymmetry that good engineers learn quickly. Under-request memory or runtime and a job may fail after hours of work. Over-request resources and the job is more likely to finish, but unused GPUs or memory remain unavailable to colleagues while queues lengthen around them. The locally rational decision becomes a system-wide capacity problem.

This is not only a startup thesis. A 2024 Microsoft Research study examined 400 low-utilization deep-learning jobs and found 706 issues across data operations, model behavior, job configuration, and software dependencies. The researchers reported that 84.99% of the issues could be addressed with a small number of code or script changes, showing how much useful capacity can disappear inside ordinary workload decisions.

The market pressure is also moving beyond the cluster. The International Energy Agency reported that global data-center electricity demand rose 17% in 2025 and consumption from AI-focused data centers rose 50%. Gartner forecasts $1.366T of AI-infrastructure spending in 2026. More hardware is arriving, but power, cooling, interconnection, and capital remain slow enough that recovering existing capacity has become its own infrastructure category.

How Expanse Works Before the Scheduler

Expanse's product has two connected parts. Analyse studies workload code, cluster telemetry, hardware, and the history of similar jobs to estimate resource allocation, runtime, memory use, completion probability, and failure risk before submission. Diagnose examines code, logs, metrics, and the observed result after execution, turning each run into additional evidence for future predictions.

The software works alongside SLURM, Kubernetes, and Nomad rather than replacing the scheduler. It can operate in customer-controlled cloud, on-premises, or hybrid environments. Expanse says its models and management layer deploy inside the customer's infrastructure so code, logs, telemetry, and job data do not leave that environment.

That timing separates Expanse from tools focused mainly on observability after a workload has begun. Monitoring can show that a cluster is underused. Expanse is trying to change the allocation decision before capacity is committed, then connect the prediction with the observed outcome so the model becomes specific to that environment.

The Evidence Expanse Still Has to Turn Into Trust

Expanse says it identified nearly $8M of idle compute capacity in one production deployment. The company also reports that teams reserved 2 to 2.5 times what they used across 122,000 production HPC jobs. In a June 2026 benchmark across two national HPC systems, Expanse says it predicted runtime within 10% and memory within 5% at the median, while frontier language models were materially less accurate.

Those figures are useful signals, but they remain company-reported. Expanse has not published named customers, audited savings, contract values, retention, or a customer-by-customer performance record. The central commercial question is not whether wasted compute exists. Independent research already establishes that. The question is whether one prediction layer can stay calibrated across different schedulers, workload types, hardware generations, research cultures, and security requirements.

The company's self-hosted design addresses part of that sales burden by keeping sensitive code and telemetry inside the customer's environment. It also increases the importance of deployment quality. A system that learns each cluster must integrate cleanly, show its evidence, communicate uncertainty, and earn enough confidence that infrastructure teams act on the recommendation.

What the $5.3M Changes

The round gives Expanse capital to expand engineering and put the product in more environments where the cost of a bad request is visible. AI labs care about training and inference capacity. Quantitative firms care about research throughput. Life-sciences teams may be running simulations and computational workloads on shared systems. National and institutional HPC operators manage long queues where one oversized reservation can delay many users.

Expanse enters a market that already includes schedulers, observability vendors, FinOps products, cloud-optimization software, and compute marketplaces. DevCuration has also tracked businesses such as Liquid Compute, which attacks capacity economics through a regulated market concept rather than workload prediction. Expanse's narrower claim is that better evidence at submission time can recover capacity before customers buy, rent, or trade for more of it.

That claim will now meet the people whose caution created the opportunity. The company must persuade engineers that requesting less hardware does not mean accepting more failed jobs, and persuade infrastructure leaders that the recovered capacity is dependable enough to change budgets and planning. The next phase of Expanse will be measured inside those ordinary submission decisions, one smaller request and one completed workload at a time.