answers.delivery

How-to · 3 min read

How to tell if self-hosting an LLM beats paying for an API

Your team pays a monthly API bill for a language model, and someone keeps asking whether renting your own GPUs would be cheaper. Blog posts give confident answers in both directions, usually built on someone else's traffic and hardware. You want a number for your own case. How many tokens a month do you need before your own box costs less than the API?

Published
17 Nov 2026
Delivery
No. 12
Topic
Education
Evidence
10 claims quoted from masterainow.com

01
Why it's hard

The two bills behave differently, so comparing them on a single month's figure misleads you. An API charges per token and drops to zero when nobody calls it, while a rented GPU costs the same whether it serves one request or a million. That means utilization decides the answer, and most teams guess it far too high. Throughput is the other trap, because a single chat stream and a batched serving engine on the same card can differ by a lot, and borrowed benchmarks rarely match your setup.

This part describes the usual way. It makes no claims about MasterAI Now, so nothing is quoted here.

02
How MasterAI Now does it

MasterAI Now has a free self-host vs API cost calculator in its AI tools section. You pick an API model and your monthly input and output tokens on one side, then a GPU, the count, your provider's hourly price, your measured throughput, utilization and any extra monthly cost on the other. It shows both monthly bills, the break-even volume where they are equal, and how much of the deployment's capacity your traffic would use. It refuses to guess the rental price and throughput, so the result is built on your numbers.

Enter your own rental price and your own measured throughput to find the volume where the two cross.Quoted from masterainow.com
A card busy 20% of the month costs five times more per token than one kept busy, because the bill is identical and only the denominator moved.Quoted from masterainow.com
Monthly output tokens where both bills are equal.Quoted from masterainow.com

03
Step by step

  1. 1Open the self-host vs API cost calculator from the AI tools menu.
  2. 2Choose the API model you use now and enter your monthly input and output tokens in millions.
  3. 3Pick the GPU and how many cards you would rent, then type the hourly price from your provider.
  4. 4Enter the sustained output tokens per second you measured on your own deployment.
  5. 5Set a realistic utilization and add extra monthly costs like engineering, storage and egress.
  6. 6Read the break-even volume and the capacity used, then compare them with your real traffic.

04
Tips

Measure throughput before you trust any result. Run your own model on the card you plan to rent, at your real prompt lengths and the concurrency you expect, and use that figure.

Run your actual model on your actual card at your actual prompt lengths, and note the aggregate output tokens per second at the concurrency you expect.Quoted from masterainow.com

Be strict about utilization if your traffic has a daily peak. A box sized for lunchtime load sits mostly idle at night, and the page itself warns that this is where self-hosting usually loses. Batch jobs that run flat out are the opposite case.

This is why self-hosting rarely wins for traffic with a daily peak.Quoted from masterainow.com

Remember what the math leaves out. Ops time, quality gaps between the API model and the open model you would run, and cold starts are not in the formula, so treat a narrow win for self-hosting as a tie. Ops time in particular is easy to forget, since someone has to patch, monitor and restart the serving stack every month.

Quality differences between the API model and the open model you would host.Quoted from masterainow.com

Check the date on the API prices before you present the result. Model prices change, so confirm the current rate on the provider's own pricing page if the gap between the two bills is small. A cheaper model tier on the API side can also move the break-even point a long way.

Prices verified February 1, 2025Quoted from masterainow.com

05
Questions

Does the calculator send my numbers anywhere?

No. It runs entirely in your browser, and what you type stays on the page.

Runs entirely in your browserQuoted from masterainow.com

Which GPUs can I pick?

Only cards with a verified VRAM figure, from the RTX 4090 up to the H200 SXM.

Only cards with a verified VRAM figure are listed.Quoted from masterainow.com

Is there a default GPU rental price?

No. You enter the hourly price from your own provider, because the tool does not assume one.

Enter the price from your providerQuoted from masterainow.com

More from MasterAI Now

All 2 answers