ServicesAbout UsContact Us
Sell

Own vs. API

Own your compute.

Compare what you pay for tokens today with what it costs to run open models on servers you own.

Run the numbersSource hardware
What you use today
billion

Input and output together. Your API dashboard shows this.

per million

Blended across input and output.

The servers

GPU generation

Example from public reports. Used costs less than new.

tokens/sec

Depends on the model and serving stack.

% of the time
kW
Running it
per kW a month

Colocation with power included.

a month
years
% of price

What it sells for when you move to the next generation.

The estimate

Owning costs about 3.4× less.

API$120K a month
Owning$35.3K a month
Servers needed
2 (36.8B tokens a month)
Hardware up front
$600K
Pays back in
6 months
Per million tokens
$1.77 owned vs. $6.00
Saved over 3 years
$3M
What owning costs a month
Hardware, net of resale
$11.7K
Power and space
$3.7K
People and support
$20K

A rough estimate from your inputs, not a quote. Server prices are examples. Open-weight models aren’t a match for every frontier model, so test yours on your own work first.

Revolut reported cutting its AI coding costs about 10× after moving to self-hosted models on 56 NVIDIA B200 GPUs. The Register, November 2026

When owning pays

Three signs it's time.

  1. Your volume is steady and heavy.

    Servers cost the same whether they're busy or not. Owning pays when your token use is high every day, not just in spikes.

  2. Your data has to stay in-house.

    On your own hardware, prompts, code, and customer data never leave servers you control.

  3. An open model does the job.

    Open-weight models now handle a lot of everyday work, like coding help, search, and support. Test one on your own tasks first.

What it takes

From API bill to your own racks.

  1. 01

    Pick a model.

    Test open-weight models on your real work.

  2. 02

    Size the cluster.

    Your token volume sets how many servers.

  3. 03

    Find power and space.

    Your data center or a colocation site.

  4. 04

    Source the servers.

    New, or used from data center refreshes.

  5. 05

    Sell when you upgrade.

    Resale helps pay for the next generation.

Ready to price hardware

Tell us what you need.

Model, how many, and where it's going. New and used, and you inspect before you pay.

Source hardware

Questions people ask

Is self-hosting an LLM cheaper than paying for an API?

At high, steady volume it often is, because you pay for the hardware once instead of for every token. At low or spiky volume an API usually costs less. The calculator above shows where your usage falls.

How much does it cost to run an LLM on your own servers?

Four things add up: the servers, power and space to run them, the people who keep them running, and what the hardware is worth when you upgrade. Spread over the years you keep it, that's your monthly cost.

How many GPUs do I need to self-host a model?

It depends on the model's size and how many tokens you serve. A mid-size open-weight model fits on one 8-GPU server, and you add servers as your volume grows. The calculator sizes the cluster from your token volume.

What is an on-premise or private LLM?

A language model that runs on hardware you own or lease, in your data center or a colocation facility, instead of through a provider's API. Your data stays on your servers.

Should I buy new or used GPU servers for inference?

Both work. Used servers from data center refreshes cost less. New servers make sense for the latest generation. Inspect used units before you pay.

What happens to the hardware when I upgrade?

Sell it. Data center GPUs keep resale value, and selling the old generation helps pay for the next one. You can sell it on Replenysh too.