How Much Does It Cost to Self-Host an LLM for 200 Employees?
How much does it cost to self-host an LLM for 200 employees? A worked example with GPU rental, hosted API comparison and the arithmetic so you can check it.
How much does it cost to self-host an LLM for 200 employees? The answer depends on how much your people use it, and the more they use it, the better self-hosting looks. This article gives a worked example with every calculation shown, so you can replace our assumptions with your own.
All numbers below are examples, not quotes. Hosted prices are list prices as of November 2026. For your own figures, use the LLM cost calculator.
The cost components
Self-hosting has these cost lines:
| Cost | Notes |
|---|---|
| GPU servers | Rented by the hour or bought. Usually the largest line. |
| Management platform | Loadcore pricing is not published, so it is not included here. |
| Storage and network | Model files are 18 GB to hundreds of GB. Usually small next to the GPUs. |
| People | Someone has to own the service. Often a fraction of an existing IT role. |
The rest of this article focuses on the GPU line, because that is the one you can calculate.
GPU cost: a rented example
Take a rented H200 at $2.49 per hour, the example used on our homepage. A month has about 730 hours:
- 1 server running 24/7: 2.49 x 730 = $1,817.70 per month
- 2 servers running 24/7: 2 x 1,817.70 = $3,635.40 per month
An H200 has 141 GB of memory, enough for gpt-oss 120B with room to spare for long contexts and parallel requests. If you buy hardware instead, you trade a monthly rental for a capital cost and your own power and hosting.
The hosted comparison
List prices per 1M tokens as of November 2026:
| Model | Input | Output |
|---|---|---|
| GPT-6.1 Sol | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
We need an assumption for usage. Here are two scenarios per employee per month:
- Light: 1M input tokens and 0.25M output tokens
- Heavy: 5M input tokens and 1.25M output tokens
For 200 employees, the totals are:
| Scenario | Tokens per month | GPT-6.1 Sol or Claude Sonnet 5.5 | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Light | 200M in, 50M out | 200 x 2 + 50 x 10 = $900 | 200 x 4 + 50 x 20 = $1,800 | 200 x 10 + 50 x 50 = $4,500 |
| Heavy | 1,000M in, 250M out | 1,000 x 2 + 250 x 10 = $4,500 | 1,000 x 4 + 250 x 20 = $9,000 | 1,000 x 10 + 250 x 50 = $22,500 |
What the numbers say
Set them next to the rented server:
| Scenario | Hosted API | 1 rented H200 | 2 rented H200 |
|---|---|---|---|
| Light | $900 to $4,500 | $1,817.70 | $3,635.40 |
| Heavy | $4,500 to $22,500 | $1,817.70 | $3,635.40 |
Once a team uses AI every day, a fixed-cost server wins, and the gap widens with every new user and integration. In the heavy scenario one H200 costs less than half of the cheapest hosted option, and around a twelfth of the frontier option. Even in the light scenario, one server is already on par with Claude Opus 5.5 and far below GPT-6 Astra.
Hosted bills also grow with adoption, which is exactly what you want from an AI rollout. A server lets you encourage usage instead of rationing it.
How many people one server serves depends on the model, the quantization, context length and how many people are active at the same time. Start with one server, watch usage in the Console, and add a second when you need more capacity.
What fits on the hardware
The model decides the hardware. From our model library:
- Gemma 4 31B: 18.3 GB at Q4_K_M, fits one 48 GB GPU comfortably
- Qwen3.8 27B: 17.6 GB at Q4
- gpt-oss 120B: 63.4 GB in MXFP4, fits one 80 GB GPU
- GLM-5.3 and DeepSeek V4.1 Flash: 200 to 500 GB, multi-GPU
Smaller models on cheaper GPUs change the math considerably. A model that fits a 48 GB card does not need an H200-class budget.
Cost is not the only reason
Many companies self-host for reasons that are not in the table:
- Data control. Prompts and completions stay on your servers. See GDPR-compliant AI and on-premise LLM.
- Predictable spend. A server costs the same whether 20 or 200 people use it.
- Per-department access. Loadcore grants models to departments and gives each person a personal key with usage statistics.
Plan for growth
For a business-critical assistant, run the same model on a second server. That doubles capacity and gives you a spare, and at $3,635.40 a month it still costs less than the heavy hosted scenario on any model above. That is why the table includes a two-server column.
A quick way to estimate your own number
- Count your active users, not your headcount.
- Estimate tokens per active user per month, from a pilot if you can.
- Multiply by hosted prices (current ones) to get the hosted baseline.
- Choose a model, find the GPU that fits it, multiply the hourly price by 730.
- Add a second server for capacity and a spare.
- Compare, and add the qualitative factors.
Read Can Open-Source AI Replace ChatGPT in Your Company? for how to decide, and compare options on the self-hosted ChatGPT alternative and self-hosted Claude alternative pages.