LLM Hangar / Tools / Self-hosted LLM cost calculator
Self-hosted LLM cost calculator: GPU hour vs API tokens
Every published "self-hosting vs API" article lands on a different break-even number because each one assumes a different GPU price, a different tokens-per-second figure, and a different number of hours the instance is actually on. This calculator makes those three inputs explicit so you can see what your own numbers say. The reasoning behind it is in Self-hosted LLM cost in 2026: what the guides get wrong.
Your numbers
Monthly figures use 30.4 days. All amounts in USD. Change any field and the results update.
How to read this
Utilization is the whole game. A GPU instance bills for every hour it is on, whether it served one request or a thousand. If your traffic uses a tenth of the instance's capacity, your real cost per token is ten times the headline figure; as one of the better decision guides puts it, "10% utilization makes every number 10x worse" (digitalapplied, May 2026). The hours-per-day field is where that shows up. Business hours only cuts the GPU line by two thirds.
Stop is not terminate. A stopped instance stops the GPU meter but keeps billing its disk. On AWS that is the EBS volume holding the model weights, which for a large model is hundreds of gigabytes. The calculator ignores storage because it is small next to the GPU line, but a forgotten stopped instance is not free, and neither is a prepared stage you keep for fast reboots. LLM Hangar's budget caps and auto-destroy exist because of exactly this.
Published break-even claims range from 2 million tokens a day to 11 billion a month because nobody measured. The inputs you enter above are the reason the articles disagree; the arithmetic is not in dispute. The only numbers LLM Hangar has measured so far are a cold boot of 21 minutes 26 seconds for DeepSeek V4 Flash on 2x H200 and an observed $7 to $14 an hour for that shape on RunPod, both in the DeepSeek V4 Flash guide.
Sometimes cost is not the question. If client data may not leave your account, the API column is not an option and the calculator only tells you what the constraint costs. That is a common reason to self-host and it has nothing to do with break-even.
If the GPU line is what you want to reduce, start with the instance choice: deploying an LLM on AWS without the CLI walks through which shapes fit which models and what they cost by region.
Questions
Is self-hosting an LLM cheaper than using an API?
Only above a token volume that depends on three numbers: the GPU rate you actually pay in your region, the tokens per second you actually get from the model on that GPU, and how many hours a day the instance is on. Below that volume the API is cheaper; above it, self-hosting is. Published guides disagree by three orders of magnitude because they assume different values for those numbers instead of measuring them.
At what volume does self-hosting break even?
Break-even tokens per day equals your monthly self-hosting cost divided by 30.4 and by the blended API price per token. With a single L40S at about $2 an hour on demand and the cheapest API prices on the market, that is well over 100 million tokens a day, which one L40S cannot serve; with an expensive API or a spot instance it can be under 10 million. Enter your own numbers rather than trusting a single published figure.
Do I need Kubernetes to self-host an LLM?
No. One GPU instance running vLLM or SGLang behind a key-authenticated endpoint serves most teams. Kubernetes becomes relevant when you run many models or need autoscaling across nodes, and it is also where most of the DevOps cost in published estimates comes from.
What is the utilization trap in self-hosting?
A GPU instance bills by the hour whether it serves one request or a thousand. If your traffic only uses 10 percent of the instance's capacity, your effective cost per token is ten times the headline figure. Running the endpoint only during working hours, or on a self-destruct timer, is the single biggest lever on the real number.