Run AI on your own servers
Which open models run on which GPU server, and what owning one costs.
Which server runs which model
| Model, by size | RTX PRO 6000, 96 GB per GPU | H200, 141 GB per GPU | B200, 180 GB per GPU | B300, 288 GB per GPU |
|---|---|---|---|---|
| Qwen3.8-27B, 27B parameters | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| DeepSeek V4-Flash, 284B parameters | 2 GPUs | 4 GPUs | 4 GPUs | 4 GPUs |
| GLM-5.3, 743B parameters | No recipe | 1 server | 1 server | 1 server |
| DeepSeek V4-Pro, 1.6T parameters | No recipe | 1 server | 1 server | 1 server |
| Kimi K3, 2.8T parameters | No recipe | 2 servers | 2 servers | 1 server |
8-GPU servers. Lime marks the GPUs one copy of the model uses. Leading open models as of October 2026, from each model’s serving recipe (Qwen, DeepSeek, GLM, Kimi).
What one server costs to own
H200 server, upfront
$383K
- Then $5K a month
- About $16K a month over 3 years
B200 server, upfront
$520K
- Then $7K a month
- About $21K a month over 3 years
B300 server, upfront
$596K
- Then $7K a month
- About $24K a month over 3 years
Ballpark, October 2026. The server, shipping, install, colocation setup and deposit, and our $25K setup, then colocation at $280 per kW, our management, and parts, bandwidth and insurance. Before taxes and financing.
When owning beats an API
Steady, heavy use
Many people or systems on the models all day.
Data that can’t leave
Contracts, patient records, source code, client files.
A fixed bill
The same every month, however many tokens you use.
Your model, your version
Fine-tune it and keep it as long as you need.
Plan your private AI servers.
- A call with our CTO about what you need
- Ballpark costs and timelines for your budget
- A straight answer if it’s not a fit
Prefer email? hello@amcompute.com
Questions about running AI yourself
- Can we run ChatGPT on our own servers?
- Not ChatGPT itself. Open-weight models such as Qwen, DeepSeek, GLM and Kimi run on your own hardware.
- Is running our own AI cheaper than an API?
- Only with steady, heavy use. One H200 server costs about $16K a month over three years, busy or idle. Owning pays off when many people or systems use the models all day, or when the data can’t leave your hardware.
- Can we start with one server?
- Yes. One 8-GPU H200 server runs GLM-5.3 or DeepSeek V4-Pro, and a 27B model like Qwen3.8-27B needs a single GPU. Our setup fee is a flat $25K for 1 to 7 servers, not a percentage of what you buy.
- Which server runs the largest models?
- Kimi K3, at 2.8 trillion parameters, fits on one 8-GPU B300 server. On H200 or B200 it takes two servers.
- Can we run local AI on a workstation?
- For one person, yes. A 27B model such as Qwen3.8-27B, compressed to 4-bit, runs on a single RTX 5090. A team, or the larger models, needs a server in a data center.
- Do we need a data center?
- Yes for these servers. Each draws 10-16 kW, more than an office can supply. We host them in a U.S. colocation data center, connected to your offices over a private link or VPN.
- What software runs the models?
- Open-source inference servers such as vLLM and SGLang load the model and serve it to your apps over an API like OpenAI’s. Your team or software vendor picks the models. We keep the servers running.