The same H100 costs 69% more at one cloud than another.
AI Compute Arbitrage prices your workload across AWS, Google Cloud, Azure, CoreWeave, Lambda, RunPod and Vast.ai in one shot — then tells you which one to actually buy, and what you save by switching.
Coverage
Hyperscalers, GPU clouds and marketplaces — priced the same way
AWS
Deepest region coverage and the widest managed-service catalog. You pay a premium for both.
48 offers · from $1.006/hr per GPU
CoreWeave
GPU-native cloud with InfiniBand fabrics. Fast multi-node scheduling at well under hyperscaler rates.
42 offers · from $0.734/hr per GPU
Google Cloud
Strong networking and TPU alternatives. Committed-use discounts reward predictable workloads.
48 offers · from $0.916/hr per GPU
Lambda Labs
Simple flat pricing and preinstalled ML images. Capacity for the newest GPUs sells out quickly.
28 offers · from $0.605/hr per GPU
Microsoft Azure
Enterprise agreements and hybrid identity. Typically the priciest per GPU-hour of the big three.
48 offers · from $1.149/hr per GPU
RunPod
Per-second billing and instant pod start. Community cloud tiers undercut everyone but AWS spot.
48 offers · from $0.400/hr per GPU
Vast.ai
Open marketplace of third-party hosts. Cheapest per GPU-hour, most variable in reliability.
48 offers · from $0.278/hr per GPU
Savings calculator
Price your next run before you launch it
No account needed. Change the workload and every provider reprices — including how many nodes you would actually have to rent.
You save
69.1%
$20,366.64 versus Microsoft Azure
Cheapest total
$9,095.76
8 × H100 SXM 80GB for 720 h
Baseline total
$29,462.40
On Microsoft Azure
Annualized saving
$247,794.12
If this workload ran every month
Vast.ai
US East · 1 × 8-GPU instance
- Rate
- $12.63/hr
- Per GPU-hour
- $1.579/hr
- Total
- $9,095.76
- Saving
- $20,366.64 (69.1%)
Lowest total bill for the run
RunPod
US East · 1 × 8-GPU instance
- Rate
- $14.71/hr
- Per GPU-hour
- $1.839/hr
- Total
- $10,592.64
- Saving
- $18,869.76 (64.0%)
Capacity you can get today
Vast.ai
US East · 1 × 8-GPU instance
- Rate
- $12.63/hr
- Per GPU-hour
- $1.579/hr
- Total
- $9,095.76
- Saving
- $20,366.64 (69.1%)
Price efficiency weighed against reliability
Total cost by provider
Best offer per provider for 8 × H100 SXM 80GB over 720 hours
Cheapest option for this workload
45 offers priced. Sign up to save this as a project and track it over time.
Create a free accountFeatures
Everything you need to stop guessing where to run a job
Comparison engine
Enter GPU, count, hours and region. Every matching offer is priced — including how many nodes you would actually have to rent when a shape overshoots.
Savings math, not vibes
Savings are measured against the priciest hyperscaler quote for the same workload, so the number survives contact with your finance team.
Seven providers, one table
AWS, Google Cloud, Azure, CoreWeave, Lambda Labs, RunPod and Vast.ai — hyperscalers, neoclouds and marketplaces side by side.
Cheapest, fastest, best value
Three rankings, because the cheapest node is not always the one you want. Value blends price efficiency with reliability and provisioning speed.
Dashboard analytics
Spend trend, GPU-hours, utilization, provider mix and realized savings per project — all derived from your own workloads.
Price history
Daily price points per provider so you can see who is drifting up and who is discounting before you sign anything.
Watchlists
Star the offers you are tracking. Sold-out capacity at a good price is worth watching, not forgetting.
A real REST API
/api/providers, /api/gpus, /api/pricing, /api/projects, /api/users. Same engine the UI uses — script it into CI or your scheduler.
Pricing
The tool that saves you money should not cost money
The whole platform runs on free tiers — free hosting, free database, free auth. There is no paid plan to upsell you to.
Free
$0forever
Everything in this repository. Self-host it and you are done.
- Unlimited price comparisons
- All seven providers
- Up to 5 saved projects
- 60 days of price history
- Full REST API
- Automated workload scheduling
- Team workspaces
Team
Most popular$0while in preview
For groups that want shared projects and longer history.
- Everything in Free
- Unlimited projects
- 12 months of price history
- Shared watchlists
- Spot-price tracking
- CSV export
- Automated workload scheduling
Self-hosted
$0MIT licensed
Clone it, point it at your own database, deploy it anywhere free.
- Full source
- Bring your own pricing feed
- SQLite or Postgres
- Deploy on Vercel or Cloudflare free tiers
- No telemetry
FAQ
Questions people actually ask
Where do the prices come from?
The bundled catalog is an illustrative reference set — public list prices, jittered per listing, with a synthetic daily history so the trend charts have something to show. It exists so the app is fully usable without a single paid API key. Swapping in a live feed means implementing one interface (PricingSource); no UI or API code changes.
How is 'savings' calculated?
We price your exact workload — GPU model, count, hours, region — on every provider that offers it. The baseline is the most expensive hyperscaler quote for that same workload, because that is usually the bill teams are trying to escape. Savings is the gap between that baseline and the option you pick.
What if a provider only rents 8-GPU nodes and I need 3?
Then you get charged for 8. The engine computes how many nodes you would actually have to rent, prices those, and flags the idle GPUs in the results. A per-GPU rate that ignores instance shapes makes cheap providers look cheaper than they are.
What does the value score mean?
It blends price efficiency (throughput per dollar, normalized across candidates), the provider's reliability score, and how quickly it provisions capacity. The cheapest marketplace listing and a hyperscaler with a real SLA are not the same purchase, and one number should not pretend they are.
Is spot pricing included?
Yes, for the providers that offer interruptible capacity. Switch the pricing mode and the comparison reprices against spot rates, hiding providers that have no spot tier. Remember spot capacity can be reclaimed mid-run — the savings are real but so is the risk.
Does it cost anything to run?
No. Next.js on Vercel or Cloudflare's free tier, SQLite locally or Supabase's free Postgres, Better Auth for sign-in, and no paid APIs anywhere in the stack. The one-command setup creates the database, seeds it and starts the dev server.
Can it schedule my jobs automatically?
Not yet. The data model and the pricing engine were built with it in mind — projects already carry a selected offer and a repricing routine — but automated scheduling, Kubernetes integration and spot reservation are on the roadmap, not in the MVP.
Can I use this for real purchasing decisions?
Use it to shortlist. Always confirm the live rate, region availability and egress terms with the provider before committing spend — GPU pricing moves faster than any catalog.
Contact
Tell us what you are running
Missing a provider? Want a live pricing feed wired in? Found a number that looks wrong? Send it over — the pricing model is meant to be argued with.