AI Use Cases: GPUs and LLM APIs Compared
What a workload costs depends on how you run it. Calling a hosted model bills per token; renting GPUs bills per hour. Each page below covers what the workload demands, which models or GPUs suit it, and what providers currently charge.
No infrastructure, scales to zero, access to closed models. Suits variable traffic and most application workloads. 9 use cases.
Training, fine-tuning, open-weight models at steady high utilisation, or data that must stay on hardware you control. 8 use cases.
LLM API use cases
All LLM API prices →Models are chosen from live prices by a quality floor and any hard requirement, then ranked by the estimated cost of a typical request. Figures are per 1,000 requests at the cheapest provider.
Best LLMs for Coding Agents
Long, mostly cached repository context with high reasoning demands
Best LLMs for Reasoning and Research Agents
Multi-step planning, tool use and long outputs on hard problems
Best LLMs for Customer Support Chatbots
Short conversational turns where latency and cost per conversation matter
Best LLMs for RAG Question Answering
Answering from retrieved passages with grounded, cited responses
Best LLMs for Document Summarization
Large inputs, short outputs — input rate and context window decide the bill
Best LLMs for Structured Data Extraction
Turning emails, invoices and forms into JSON at volume
Best LLMs for Classification and Content Moderation
Tiny outputs at very high volume — the cheapest capable model wins
Best LLMs for Vision and Document Understanding
Reading screenshots, scans, charts and photos alongside text
Best LLMs for Translation and Localization
Output roughly equal to input — both token rates count
GPU use cases
All GPU prices →GPUs are ordered by suitability for the workload, not by price. Figures are the lowest current on-demand rate per GPU-hour across providers.
Best GPUs for LLM Training
Multi-GPU nodes with HBM and high-bandwidth interconnects
Best GPUs for AI Inference
Throughput per dollar, with enough VRAM to hold the weights
Best GPUs for Fine-Tuning
LoRA and QLoRA on one GPU, full fine-tunes on a node
Best GPUs for AI Image Generation
Diffusion models fit in 24 GB; batch size buys throughput
Best GPUs for Scientific Computing
FP64 throughput, ECC memory, and node-level interconnect
Best GPUs for Video Rendering and Transcoding
Hardware encoders, VRAM for scene data, and stream density
Best GPUs for RAG Pipelines
An embedding model and a generator, sized separately
Best Budget GPUs for AI
The lowest cost per unit of work, not the lowest hourly rate
Frequently Asked Questions
Should I call an LLM API or rent GPUs and run the model myself?
An API bills per token and needs no infrastructure, so it suits variable or modest traffic and access to closed models. Renting GPUs bills per hour whether or not they are busy, so it pays off when utilisation is steady and high, when you need an open-weight model the APIs do not host, or when data must stay on hardware you control.
How are the API cost estimates calculated?
Each API use case states a typical request: input tokens, how many of them are served from a prompt cache, and output tokens. The estimate applies every provider’s current standard rates to that request and takes the cheapest provider for each model. The request shape is shown on each page so you can scale it to your own traffic.
How are models chosen for each API use case?
Models are selected from live pricing rather than a fixed list. Each use case sets a quality floor as a percentile of the Artificial Analysis Intelligence Index across all priced models, plus any hard requirement such as image input or a minimum context window. Models that clear it are ranked by estimated cost.
How are GPUs ordered on the GPU use-case pages?
By suitability for the workload — memory capacity, bandwidth, interconnect and precision support — not by price. The comparison table shows each GPU’s specs beside its lowest and median on-demand rate per GPU-hour, its lowest spot rate, and for multi-GPU workloads the cheapest node of the typical size. Spot and reserved rates are never used as the headline price.