Top AI Product

Expanse, launching out of Y Combinator’s current batch, attacks an embarrassing number nobody wants to put on a slide: datacenters run at roughly 30–40% effective GPU utilisation. The rest is paid for and idle. Expanse’s job is to make it useful.

## Three things it actually does

The product wraps three capabilities. **Resource prediction** right-sizes job submissions before they reach the scheduler — most researchers over-request, because under-requesting kills their job and over-requesting just wastes someone else’s capacity. **Optimisation suggestions** surface code and config changes researchers can apply themselves rather than hand-holding through migrations. **Failure prediction** catches jobs that will fail before they burn hours of GPU time. The pitch deliberately stays out of the workflow: deploy Expanse, change nothing in your scripts.

## Why the team matters

Four engineers from HPC and GPU training at the largest quant funds and national supercomputing centres. The technical lead, Ismaeel, built the first multimodal HPC resource predictor as research at EPCC — Edinburgh’s Parallel Computing Centre — and reportedly beat every published baseline. That’s the kind of domain background that decides whether a “GPU optimisation” startup is real or marketing.

## Why it matters

Frontier training is GPU-bound and frontier inference is becoming GPU-bound, and the curve of “buy more H200s” eventually breaks. Squeezing actual work out of capacity that’s already in the building is the unglamorous next step — and one of the cleanest places left for a startup to add value without trying to out-train OpenAI.