Context.dev released GPU Terminal, an open-source GPU price tracker

A team renting an H100 by the hour can now compare 62 providers on one page before they sign a monthly contract. GPU Terminal, released by the YC-backed Context.dev, is a free dashboard that pulls source-backed hourly rates for H100, H200, B200, A100, RTX 5090, RTX 4090, L40S and RTX A6000 cards and shows the price change for each provider over the past week. The site went live on Product Hunt with 106 upvotes.

The marquee on gpu.context.dev states the current low for each card: H100 from $0.12, H200 from $2.00, B200 $3.50, A100 from $0.48, RTX 5090 $0.35, RTX 4090 from $0.26, L40S $0.47 and RTX A6000 from $0.20. The source for each price is one click away.

What the dashboard shows for one card

Pick a GPU family and the page lists every provider with a current observation for it. For H100 the list carries 62 providers. Each row shows the lowest price seen, whether it moved in the past week, and when the row was last observed. OctaSpace sits at $0.12 per GPU-hour at the bottom of the list and Replicate at $5.49 at the top, with Hugging Face at $4.50 and Modal at $3.95 near it. A one-week sparkline tracks each provider's own line.

Filters narrow the list by VRAM, region, interruption risk, billing model, deployment type, local storage, provider platform (hyperscaler, specialist, decentralised marketplace), minimum spend and required GPU count. A toggle hides providers that only list a "starting at" price. The current filter set exports to CSV, and a share link reproduces the view somebody else loaded.

Where the data comes from

The methodology page states that the project reads prices directly from each provider's own page. The GitHub repository, context-dot-dev/gpu-terminal, is Apache 2.0 and four days old, so the open-source history is still thin. Every observation carries a source URL, a timestamp and the collection method, and the site records a provider as "unspecified" rather than filling in a figure it did not read, so a user can see directly where a filter does not apply.

A team moving an inference workload between providers can now point at the GPU Terminal row rather than a vendor's own marketing page, and the row names when the figure was last seen. A team on a long-term contract can read the sparkline for the specific provider they signed with and the next three cheapest, and bring that into a renewal conversation. The site is read-only. For teams that want a copy in their own tooling, the repository is on GitHub.

Source

GPU Terminal by Context.dev, launch on Product Hunt, context-dot-dev/gpu-terminal on GitHub.