Công cụ tính chi phí GPU và AI.
Nine calculators for the decisions that come before you rent anything: which GPU is big enough, which is cheapest, and whether you should rent one at all. Each one uses the same live price data as the rest of the site, and every result is a shareable URL.
VRAM fit calculator
Question it answers: will this model fit on this GPU? Pick a model, a precision (FP16, INT8, INT4…) and a context length. The calculator estimates the memory needed for the weights plus the KV cache and lists every rentable GPU with enough headroom, with the cheapest current price for each.
Use it when you're choosing hardware for an open-weights model — for example, to see that a 70B model at INT4 fits on a single 48 GB card but needs two 24 GB cards.
Self-host vs API breakeven
Question it answers: at my traffic, is it cheaper to rent a GPU and serve the model myself, or pay a hosted API per token? It plots both monthly costs against request volume and marks the crossover.
Use it when you're about to commit to a deployment. At low volume, APIs almost always win because you only pay for tokens you use; at steady high volume a rented GPU that stays busy usually wins.
API monthly bill estimator
Question it answers: what will a month of my workload cost on each hosted API? Enter your input and output tokens and get the bill for every provider that serves the model, cheapest first.
Use it when you already know you want an API and need to pick the provider — output tokens often cost 3–5× input tokens, so the ranking changes with your prompt/response mix.
GPU side-by-side
Question it answers: how do these GPUs actually differ? Compare up to four cards on VRAM, FP16/FP8 throughput, power draw and the current cheapest hourly price.
Use it when two cards look similar on paper — an L40S and an A100, say — and you want to see what you gain per dollar.
Nên thuê GPU nào?
Question it answers: I know what I want to do — which GPU should I rent for it right now? Pick a workload (run, fine-tune or train a model, or generate images), a model and what matters most, and get the three best options with provider, price and monthly cost, plus a setup checklist.
Use it when you're starting a project and don't want to cross-reference VRAM tables and price lists by hand.
Ước tính chi phí huấn luyện
Question it answers: what does it cost to train a model of N parameters on D tokens? It applies the standard 6 × parameters × tokens compute rule to each GPU's spec-sheet throughput at a utilization you choose, and prices the GPU-hours at today's cheapest listing.
Use it when you're scoping a pre-training or continued-pretraining run and need a budget and a timeline.
Tính chi phí fine-tune
Question it answers: how much memory, time and money does a full, LoRA or QLoRA fine-tune of a specific open-weights model take on each GPU?
Use it when you have a dataset and a base model and need to know whether one card is enough — QLoRA often fits a 70B model on a single 80 GB GPU.
Tính chi phí inference
Question it answers: what does it cost per million tokens to serve this model myself, on each GPU, and how does that compare with the cheapest hosted API?
Use it when you're deciding how to deploy a model and want throughput and unit cost side by side.
Tính thuê hay mua
Question it answers: should I buy a GPU instead of renting one? It compares renting with owning (depreciation plus electricity) and shows the hours per month at which buying pays off.
Use it when your GPU bill has become a fixed monthly cost and you wonder whether the card would pay for itself.
Limits worth knowing
These are planning tools. Memory estimates can be off by a few GB depending on your inference framework, and throughput depends heavily on batch size and software. The methodology explains the formulas.