Features
See what every GPU-hour costs, and what it served.
From public GPU prices to the idle time inside your own fleet. Start free with the radar and planners; bring your usage data when you want your own numbers.
GPU Radar
What a GPU-hour costs right now, across providers, with the time every price was fetched.
-
On-demand and spot, side by side
One row per provider with separate columns, normalized to $/GPU-hour. No price means a dash, never a guess.
-
Region and price filters
Filter by US, EU or Asia and by pricing type, sort by price, and expand a row for the min–max range and every region.
-
Freshness on every number
Each row says when it was updated and is marked stale after 48 hours.
-
Price index and provider pages
Every GPU on one page with its 30-day trend, and each provider’s prices against the cheapest offer anywhere.
-
Price history
Cheapest vs median on-demand price over time for each GPU, from H100 and H200 to B300 and RTX cards.
-
Price alerts
Pick the GPUs you care about and get an email when their price or availability moves.
-
Open price data
Download the price index as CSV or JSON, or embed a live price card for any GPU on your own site.
-
The GPU Brief
A short newsletter with what moved in GPU prices, archived on the site, with an RSS feed.
Planning tools
Size and price a deployment before you commit, using live radar prices.
-
Where to run
Pick a model or a custom size and precision; get the GPUs it needs per type and the cheapest offer for each, billed in whole instances.
-
Tokens vs GPUs
Compare renting or buying GPUs against paying an API per token, with the breakeven volume on a log-log chart.
-
Token cost by GPU
What 1M output tokens of your model cost on each GPU type at your utilization, next to a hosted API price.
-
Capacity planner
Peak and average requests per second in; replicas per GPU type, the monthly bill and how much of it idles out.
-
Idle cost
GPU, count, utilization and $/GPU-hour in; monthly and yearly idle spend out.
-
GPU and model pages
Head-to-head GPU comparisons and a page per open model with the GPUs it needs and its cost per token.
-
Shareable links
Every planner keeps its inputs in the URL, so a result can be sent or bookmarked.
Inference FinOps
Your own endpoints, priced: what every token costs and how much of your GPU time sits idle.
-
Served vs idle GPU-hours
Daily spend split into the hours that served traffic and the hours that waited, per endpoint.
-
Cost per 1M tokens
Per endpoint and per day, so you can compare models, GPUs and providers on unit cost.
-
Ranked opportunities
Savings and risks ranked in dollars, so the biggest leak is always first.
-
Spend by provider
Where the money goes across clouds and marketplaces.
-
Fleet utilization by hour
A week-by-hour heatmap that shows when your GPUs are busy and when they are not.
-
CSV in and out
Send daily usage and hourly utilization from a file or the ingest API; download endpoints, opportunities and daily spend as CSV.
Load testing
Find the concurrency where latency and cost break, before real users do.
-
Latency by concurrency
p95 latency at each step of a run, with the point where it breaks your SLO.
-
Throughput by concurrency
Tokens per second as load grows, to size replicas against real traffic.
-
Run history
Every run with its steps, so you can compare before and after a change.
Synthetic monitoring
Probes that tell you when an endpoint gets slow or fails, before the pager or the bill does.
-
HTTP probes
Checks against your public endpoints on the interval you choose; private addresses are refused.
-
p95 latency and uptime
Latency for the last 24 hours and uptime for the last 90 days, per endpoint.
-
Incidents
A log of when an endpoint went slow or down.
Read-only by design
- Read-onlyidle never sits in your request path and never changes your infrastructure.
- No promptsCost analysis needs usage and billing, never model inputs or outputs.
- No cloud credentialsYou send data by CSV or API. Revocable keys, stored only as hashes.