Learn the stack. Numbers included.
Every figure on this page is live, dated, and linked to the page it was read from — the same readings as the rest of the directory, not a number typed into a guide once and left to rot.
Top GPU clouds · October 2026
Ranked by cheapest live H100-SXM-80GB rateTop inference providers · October 2026
Ranked by cheapest blended $/1M tokensNew here? Pick a path
Ordered routes through the directory · each ends somewhere you can act
Start here if "the AI stack" is a phrase you have nodded at without a picture behind it.
- 1 · The ten layers
- 2 · Browse by layer
The part most reading on this subject skips, and the part that decides the architecture.
Once you know the shape and the money, this is the decision itself.
For when the model works and now has to reach your systems.
- 1 · MCP servers
- 2 · Agent skills
Reading the numbers
The terms our tables use, defined once
Per single GPU, per hour — an 8×H100 node at $18.40/hr is $2.30 here. Always normalized per GPU, so two sellers quoting different node sizes can be compared at all.
3 parts input to 1 part output — a typical chat mix. Heavy generation? Weight output yourself in the cost calculator. Input and output are never averaged into one rate anywhere else on this site.
Dearest ÷ cheapest for identical hardware or weights. Today: 3.8× on the widest GPU. It measures a market, never a product’s own range.
No trend until a second capture exists. Where you see it, we have met that seller once — it is honesty about how long we have been looking, not a missing number.