June 1, 2026
GPU gossip: waste, wins, and whining
Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
Startup says huge amounts of pricey chip power are being wasted — and commenters are already fighting over who should cash in
TLDR: Expanse says companies and research labs waste massive amounts of costly computing power by playing it safe, and its software can spot that before jobs run. Commenters liked the idea but quickly argued over whether this is real waste, disaster backup space, or savings that should already be reaching customers.
A new startup called Expanse showed up on Hacker News with a juicy claim: big computing centers are wasting staggering amounts of expensive chip time because researchers and companies ask for way more power than they really need. Their pitch is simple in plain English: read the job before it runs, guess what it actually needs, warn if it’s about to crash, and tell people how to fix sloppy code. The founders say one large research cluster wasted the equivalent of $8.5 million in a single month. Naturally, the comments immediately turned into a mix of applause, side-eye, and “okay but who gets the savings?”
The strongest reaction was classic internet pragmatism. One commenter instantly asked why data centers don’t already pass these savings on to customers with better pricing plans, basically turning the launch into a mini trial of the cloud industry. Another brought up an old-school business reality check: sometimes “wasted” capacity isn’t waste at all, because companies keep spare room for disasters and failovers. In other words, Expanse says “look at all this idle money,” while skeptics reply, “that empty space may be there on purpose.”
Then came the funnier energy. One HPC user admitted many job submissions at their institution are “egregiously bad,” and joked they’d support every script being sent through an AI helper — for everyone else, of course. That little wink was the thread’s comic gold: everybody loves optimization, but nobody wants to be the one getting optimized. Even the more random drive-by comments — a traffic-shaper plug, a perf-book networking request — added to the messy launch-day vibe. The real mood? People think the idea is clever, but the comment section wants receipts, money, and maybe a little less confidence from everyone involved.
Key Points
- •Expanse says it improves effective capacity in HPC and GPU clusters by predicting workload resource needs, failures, and optimization opportunities before scheduling.
- •The article states that datacenters often achieve only 30% to 40% effective utilization because users over-request resources to avoid job failures.
- •In a month-long measurement of one national-scale HPC cluster, Expanse reports that 122,000 jobs showed 59% compute waste, equivalent to about $8.5 million at on-demand cloud rates for similar hardware.
- •The company links its product to prior EPCC research on a multimodal HPC resource predictor that reportedly outperformed other baselines by 34% and prompted frontier LLMs by about 8x on the same task.
- •Expanse integrates with SLURM and Kubernetes, collects live telemetry from sources such as DCGM, CUPTI, and cgroups, and offers submit-time prediction, live observability, and failure diagnosis.