An AI job can wait weeks for its turn on a cluster, run for hours or days, and then disappear with an out-of-memory error because somebody guessed one number wrong.
By then, the loss is larger than GPU time. Code has to be checked again. The workload returns to the queue. A researcher who thought the experiment was running discovers that the result they were waiting for never arrived.
Ismaeel Bashir knew that ritual from the inside.
During research at the Edinburgh Parallel Computing Centre, he worked on predicting what jobs on national supercomputers would need before they ran. Later, while running machine learning at a major quantitative fund, the same problem remained: request too little hardware and a job can fail; request too much and expensive capacity sits reserved but unused.
That experience became Expanse, founded in 2025 by Ismaeel Bashir alongside Nikodem Bieniek, Eren Mendi and Yafet Melake. The San Francisco- and London-based company has now raised $5.3 million in seed funding, led by Crane Venture Partners with participation from PXN Ventures and angel investors including former DeepMind researchers and AI infrastructure leaders. The money will expand engineering and deployment across AI, quantitative finance, life sciences, research and high-performance computing.
Bashir reduces the thesis to one line:
“The machine should carry the uncertainty, not the person.”
Expanse tries to do that before the scheduler commits the hardware.
Its analyse system can inspect a SLURM script, source file, Kubernetes manifest or Nomad jobspec and recommend CPU, memory, GPU count and walltime. But the more interesting detail sits underneath the recommendation: Expanse returns confidence and supporting evidence, including similar successful or failed executions from the customer’s own cluster.
If the workload is genuinely new and no comparable history exists, the documentation says Expanse does not manufacture a confident answer. It marks the recommendation with low confidence and missing-evidence warnings. That makes the product less like an infrastructure chatbot and more like a prediction system that is expected to show what it actually knows.
Then the job runs.
If it fails, diagnose works in the opposite direction: it examines telemetry, logs, captured source and similar executions to identify the root cause and suggest a fix. The system can produce a patch, but the current workflow is explicitly review-first, the user sees the difference before anything is applied, and Expanse does not silently rewrite the files.
The failed execution then becomes evidence for the next prediction.
That closed loop is what separates Expanse from ordinary observability. A monitoring dashboard can show that a GPU sat idle yesterday. Expanse is trying to stop the wrong amount of hardware from being reserved tomorrow. Its models run alongside existing schedulers rather than replacing them, with support across SLURM, Kubernetes and Nomad. On serving fleets, the same intelligence can operate behind the scenes to right-size models and expose prediction distributions through an API.
The potential waste is not theoretical, although the current figures come from Expanse itself. The company says one cluster revealed nearly $8 million of idle compute capacity in a single month, while analysis across 122,000 production HPC jobs found teams typically reserved two to 2.5 times the resources actually used. Expanse also reports a June benchmark where its models predicted runtime and memory substantially more accurately than frontier LLMs; those results remain company-reported rather than an independent product benchmark.
Privacy is another part of the design. Expanse’s current product architecture can keep workload code, logs and telemetry inside infrastructure controlled by the customer, while the control plane handles identity and registration without receiving workload data.
The founders have even released an open-source wastage scanner for SLURM and Kubernetes that calculates CPU and memory waste locally, giving teams a way to inspect the problem before adopting the broader platform.
The seed round will help Expanse place that intelligence layer on more clusters.
But its larger argument is already visible.
The next AI infrastructure advantage may not come from buying another rack of GPUs. It may come from finally knowing what the GPUs already installed are actually needed for.
