Renting a GPU for LLM in 2026: Prices, Cards, and Methods

GPU rentals are evolving rapidly: RTX 5090s for less than $1 per hour sit alongside H100s that have increased in price by about 50% since December. Between marketplaces like RunPod or Vast.ai and dedicated clouds, the right choice depends as much on reliability as on hardware. Here are concrete benchmarks to help you decide, card by card and use case by use case.

GPU Rental: Costs, Configuration, and Reliability to Consider

Several options exist for renting GPUs: marketplaces, GPU clouds, major providers, and services billed by the second. Each presents trade-offs between cost, setup time, and reliability. The risks of downtime must be taken into account, and the cheapest offers do not always guarantee operational security. Among the services to compare are RunPod, Vast.ai, and Lambda.

Choosing the Card Based on Model Size and Memory

The market offers a wide range of cards, from gaming GPUs to NVIDIA data center models, with prices varying significantly across the spectrum. This price diversity might suggest that the cost reflects quality, but the choice should primarily be based on model size and memory needs, often for running open models in 4 bits. For smaller models and coding agents, the 3090 and 5090 provide a good cost/performance ratio. Workstation cards offer interesting capacity but variable value; the Pro 6000 stands out in performance on a single card. For larger models, multi-GPU scalability and fast interconnection needs, the A100, H200, H100, B200, B300, and MI300X are recommended.

API or Rental: Use Cases and Starting Method

It is advisable to opt for rental in cases of batch processing or for precise model tuning, while an API is generally more economical for solo interactive development when the GPU remains idle between calls. It is suggested to start with renting a modest configuration, experiment with concrete workloads, evaluate startup, transfer, and wait times, and then choose between continuing the rental or purchasing based on the intended use. Currently, renting an RTX 5090 costs less than $1 per hour, while H100s are priced about 50% higher than in December.