Opportunities › GPU and HPC resource utilization optimization › Proposed app
Rightsize
Predict GPU, memory and walltime needs before a SLURM or Kubernetes job is submitted
- Score
- 53
- confidence 53
- Evidence
- 11
- items on the page
- Rivals
- 3
- named in sources
- First version
- 60 d
- estimate
- Type
- Desktop app
- PWA: runs in the browser, installs on Mac and Windows
- Market since
- 2003
- Mature market · AI estimate
ScreensConcept
Desktop app: HPC operators analyze job telemetry, review failure diagnostics, and monitor cluster utilization through detailed dashboards. Sample screens on a computer, in the proposed brand. Their numbers are sample data.
Job estimate
Safe to shrink41 GB
About this appAI proposal
Rightsize reads a job's source code, its submission script and live node telemetry, then predicts what the job will actually need: GPU memory, RAM, CPU cores and walltime. Each estimate comes with a confidence range and a warning if the job is likely to fail at the size requested.
It is built for operators and research or ML teams on SLURM or Kubernetes clusters. The material reports users over-requesting by 2-3x because a failed job can cost days of work. Rightsize gives users a number they can trust enough to request less.
Unlike per-user averages from sacct, it looks at the code and the new job type, not only history. Unlike a general LLM agent, it sees your hardware and topology. Predictions are refreshed as jobs finish on your cluster.
- Tagline
- Request what the job needs. Not what you fear it needs.
- Voice
- PreciseCalmPractical
- Why this name
- Two plain words joined: right and size. It says what the product does, is easy to say, and does not echo any scheduler or vendor name.
In the first version
Members say which features they would use. It tells PWAS what people want; it changes no score.
Pre-submit prediction
Point it at a submission script and get VRAM, memory, CPU and walltime estimates with confidence intervals before the job enters the queue.
No member yet
Failure risk warnings
Flags jobs likely to run out of memory or time at the requested size and shows the smallest request that keeps risk low.
No member yet
Code-aware analysis
Reads training and job code, so a changed batch size or a new model type moves the estimate without waiting for history.
No member yet
Live utilisation dashboard
Shows requested versus used resources per user, group, queue and node, with the waste in hours and in cost.
No member yet
Brand identityAI proposal
Check the name against trademark registers before using it.
Palette
#12343BDeep Teal · primary#F4F1EAPaper · surface#1F2A2EInk · text#E0902BAmber · accent#4F8A6BMoss · data, safe#B5482FBrick · data, risk
Typography
Aa
Rightsize
ABCDEFGHIJKLMNOPQRSTUVWXYZ 0123456789
PricingAI proposal
The material gives no competitor prices, only that pricing is per cluster. These tiers and the dollar figure are a starting point to test in pilots, not a market benchmark.
Pilot
Contact us
- One cluster, paid pilot
- Dashboard and predictions in read-only mode
- Setup help from the team
Team cluster
$1500 a month
- One cluster up to 50 nodes
- Pre-submit predictions and failure warnings
- Utilisation dashboard
- Failure diagnosis
Facility
Contact us
- Large or multi-cluster sites
- Self-hosted deployment
- Scheduler integration support
- Priority support
Would you pay Team cluster · $1500 a month?
Your answer tells PWAS which prices people accept. It changes no score.
No answers yet.
What it could be worthEstimate
What a software business making this much sells for today, from real deals. Pick a monthly revenue.
If it makes
Worth about
$194,000
Sale range $132,000 to $263,000
$120,000 a year in revenue
From 87 software businesses listed and sold on Empire Flippers in the last year, revenue vetted by the broker: they sell for about 19× their monthly revenue, most between 13× and 26×.
Measured on profit, the middle deal is 32× monthly net profit: about 2.7 years of profit.
An estimate, not an offer or financial advice. A real price depends on profit, growth, churn, how much of the work is the owner's, and the buyer.
How it would growAI proposal
How the product would reach new customers through its own use, and how the first ones would be found. A plan to test, not a result.
- Built-in loop
- No loop in the product. It runs inside a cluster operator's own environment. Users see it at submission, but their colleagues at other sites do not, so it spreads through sales, talks and published results.
- Do they become customers?
- Cluster users who see it cannot buy it; the operator decides. To test: whether research users at one site push their operator to start a pilot after seeing their own savings.
- First customers
- University and regional HPC centres running SLURM with GPU partitions, and small AI labs with 50-200 GPUs, where the systems lead can start a pilot.
- Where to reach them
- HPC community forums: Share a free sacct waste report script and invite operators to compare it with predictions.
- Direct outreach to sysadmins: Offer a read-only pilot on one partition, with a report of waste and failure risk after two weeks.
- Conference talks: Present anonymised accuracy results at HPC user-group meetings.
- First milestone
- Three paid pilots on SLURM clusters, each showing at least 20% less requested GPU time with no rise in failures.
Why people already pay for this
- VerifiedHacker News post received 103 points and 27 comments
- Public claimDatacenters run at 30-40% effective utilisation due to over-requesting
- Public claimUsers over-request resources by 2-3x to avoid job failure
- Public claim59% of compute was wasted on a national-scale HPC cluster over one month
- AI inference
Tech stack
Every proposal builds on the same PWA base, installable from the browser on any phone or computer. Dashed boxes are what this app adds.
- AppNext.js and TypeScript
- PWAManifest and service worker
- UITailwind CSS
- Data and authSupabase Postgres
- PaymentsStripe
- EmailResend
- HostingVercel
- This app · Node agentLightweight collector for GPU and host metricsTelemetry must be gathered on cluster nodes, which the shared stack does not do.
- This app · Scheduler integrationSLURM submit plugin and Kubernetes admission hookPredictions must appear at submission time inside the user's normal workflow.
RisksAI inference
- TechnicalAccurate prediction from code and telemetry across many job types is hard; a first version trained on little data may be no better than sacct averages.
- CommercialOperators may be slow to buy, and users may not trust lower requests after a past failure. Pilots must show failures do not rise.
- DataClusters hold sensitive code and data; many sites will require self-hosting or refuse to share job source with a vendor.
- CommercialThe material says an existing company, Expanse, already claims an 8x lead over LLMs and runs paid pilots. A small team must differentiate.
- PlatformEach site customises SLURM or Kubernetes differently, so integration work per customer could be larger than planned.
Everything marked AI proposal or AI inference was written by a model from the evidence on the opportunity page. The proposed name must be checked against trademark registers before it is used.