Best value cloud GPUs right now

Start with memory. If the model, its KV cache and the runtime do not fit, nothing else matters. Use the per-GB list filtered to the size you need; the LLM VRAM calculator turns a model into a number.

Then ask what runs out first. Generating tokens for one user at a time is limited by memory bandwidth, so the per-TB/s list is the one to read for interactive serving. Training and batched serving are limited by arithmetic, so read the per-TFLOPS list, and the FP8 column if your software runs in FP8. The training versus inference guide and HBM versus GDDR explain why.

When the lists disagree, the job decides. A consumer card can top the per-GB and per-TFLOPS lists while lacking what a production job may need, such as NVLink between cards or a data-center memory system. The RTX PRO 6000 guide compares a professional and a consumer card built on the same chip.

Specs are per family. The spec table holds one row per GPU family, so the per-TFLOPS and per-TB/s figures divide the family's cheapest price by the family's published figure. Where a family has slower variants, such as the PCIe version of a GPU sold mainly as SXM, the cheapest offer may be the slower one, and the figure flatters it. The per-GB figure avoids this: it uses the price of the exact memory size. Check the GPU specs chart and the rent page before you book.

Dense figures only. Vendors headline throughput with structured sparsity, which is twice the dense number and applies only to pruned models. This page divides by the dense figure.

Which prices count. The price is the cheapest current on-demand offer per GPU-hour: in stock, from secure (non peer-to-peer) providers, seen in the last 15 minutes. MIG slices, fractional plans and laptop parts are left out, because a slice is not the whole GPU the specs describe. Storage, network transfer and tax are extra.

Specs come from GPUPerHour's spec table, served by the public /api/gpu-specs endpoint, each row checked against the vendor's datasheet. Prices come from live provider listings when the page renders. Price per GB is the cheapest price at each memory size divided by that size, taking the best size in the family. Price per TFLOPS is the family's cheapest price divided by its dense FP16 or FP8 figure. Price per TB/s divides by memory bandwidth in TB/s.

The figures are free to reuse under CC BY 4.0; please cite GPUPerHour and link to this page, with the observation time.