idle Get early access

GPU cost intelligence for LLM inference

Stop overpaying for GPUs.

Know what every token costs. idle splits every GPU-hour you pay for into traffic served and capacity left idle, then ranks the fixes in dollars.

Get free GPU price alerts Find where to run your model
Mon Tue Wed Thu Fri Sat Sun
Serving traffic Paid for, idle: $2,200 this week Example: one 8× H100 endpoint, by hour. Hover for detail.

Today’s GPU prices, side by side.

Rental price and availability for the GPUs inference teams actually run, across GPU clouds and marketplaces, refreshed every 15 minutes. Free, no account.

Open the full radar

GPU Radar

Cheapest $/GPU-hour per provider, on-demand and spot side by side

Loading prices…

On-demand is billed by the hour and stays yours until you stop it; community-cloud listings (RunPod, Akash) count as on-demand and carry a “community” tag. Spot is spare capacity at a discount that the provider can reclaim at any time. Click a row for the range and every region.

How we compare: every price is USD per GPU-hour, a multi-GPU node’s price divided by its GPU count. Storage, egress and taxes are not included. Marketplace prices are per listing and change fast. Each price shows when it was fetched; anything older than 48 hours is marked stale.

Every saving and every risk, ranked in dollars.

idle reads your usage and the live GPU market, then sorts what it finds by what it’s worth each month. The biggest leak is always at the top.

Open the demo workspace

Example workspace

  1. $7,390/mo

    Move doc-summarize off AWS H100 on-demand

    Same 4× H100 on RunPod spot at $1.89/hr vs $4.42/hr. Nightly summaries tolerate preemption.

  2. $6,110/mo

    Right-size support-agent from 8× to 5× H100

    Load test holds p95 at 980 ms on 5 GPUs at current peak. Three GPUs sit idle most of the day.

  3. $5,510/mo

    Move transcribe from 2× H100 to 1× A100 80GB

    Whisper needs under 5 GB and runs at 22% utilization. One A100 covers peak with headroom.

  4. Risk

    support-agent has no failover capacity in US-East

    H100 availability is low at CoreWeave and none at AWS in this region. A node loss means degraded latency.

How it works

From bill to ranked fixes in three steps.

No agents to deploy and no traffic to reroute. You keep running what you run; idle tells you what it costs and where it leaks.

  1. 01

    Connect, read-only

    Point idle at your cloud billing and inference metrics. Nothing is installed in your request path, and nothing can change your infrastructure.

  2. 02

    Every GPU-hour gets a price

    idle matches what you run against live market prices from GPU Radar, then splits each hour into traffic served and capacity left idle.

  3. 03

    Act on a ranked list

    Savings and risks arrive sorted by dollars per month, each with the endpoint it applies to and the effort it takes.

Products

The platform

Four tools, one read-only view of your inference stack. Nothing we run sits in your request path: idle only reads, never routes.

GPU Radar

Free

Live rental prices and availability across providers, so you always see the cheapest GPU that can run your model.

See the demo

Inference FinOps

Design partners, from $299/mo

Cost per token and per endpoint, idle time, and ranked savings, read from your billing and metrics.

See the demo

Load testing

Early access

Push synthetic traffic at your endpoints to find where latency and cost break before real users do.

See the demo

Synthetic monitoring

Waitlist

Probes that tell you when an endpoint gets slow or expensive, before the bill or the pager does.

See the demo

Design partners

Find your idle spend with us, first.

We’re working with a handful of teams serving open models on H100s and A100s. You get a cost review and early pricing; we get your feedback while the product takes shape.

  • 30 minFree GPU cost review with the team before you commit to anything
  • $299/moDesign partner price, locked for 12 months
  • Read-onlyYou send usage and billing; idle never touches your infrastructure

FAQ

Questions teams ask first.

Anything else? Get early access and ask us directly.

Does idle sit in my request path?

No. idle only reads billing and metrics; it never routes, proxies or touches live traffic. If idle went down, your endpoints would not notice.

What does it cost?

GPU Radar, the planning tools and the price alerts are free, no account needed. Inference FinOps is open to design partners from $299/mo (apply below). Load testing is in early access and synthetic monitoring has a waitlist.

Where do the GPU prices come from?

Prices come from public provider price APIs and listings across GPU clouds and marketplaces, each with the time it was fetched. Check the provider before you buy.

What data does the platform need?

Usage and cost data: GPU hours, what you pay for them, and request metrics such as tokens and latency per endpoint. Cost analysis does not need your prompts or model outputs.

How do the price alerts work?

Pick the GPUs you care about and confirm your email. You get a message when their rental price or availability moves. Unsubscribe from any email in one click.