Private LLM for companies: self-hosted, or a private API with the same guarantees
A company wants a private LLM so that its prompts are not used for training, are not kept, are processed where it decides and are seen only by providers it has approved. Standard Compute gives you those four guarantees on one API, for frontier and open-source models, without running GPUs yourself.
Create a company workspaceTalk to usSee pricing
Self-hosting an open model is the other route. It gives you the same control, plus the bill for hardware, an inference stack and the people to run it. This page sets the two side by side so IT, security and finance can pick on facts.
What makes an LLM private for a company
The four questions a security review asks, and how Standard Compute answers each of them.
| Question | What private means | Standard Compute |
|---|---|---|
| Is our data used for training? | What private means: Never, by anyone in the path | Standard Compute: Standard Compute never trains on your prompts or outputs. With a region rule or zero retention on, every request goes only to providers that commit not to train on it |
| Is our data kept? | What private means: No copy after the answer | Standard Compute: Turn on zero data retention for the company, a team or a key. We process prompts in memory and never write them to storage or logs |
| Where is it processed? | What private means: Where you decide | Standard Compute: A Europe rule serves every request, including retries and fallbacks, only from approved European inference endpoints |
| Who can see it? | What private means: Only providers you approve | Standard Compute: Provider allow-lists per company, team or key. New providers are never added without an admin's approval |
Self-hosted LLM or a private API: what each one takes
Self-hosting puts an open-weight model on servers you control. Nothing leaves your network, and you own capacity planning, GPU costs, upgrades, failover and security patching. It suits a company with a platform team and steady, high volume on one or two models.
A private API keeps the guarantees in policy and contract instead of hardware. On Standard Compute you get Claude, GPT and Gemini next to GLM, DeepSeek and MiniMax on one key, and each step is routed to the cheapest model that can do it and the cheapest provider fast enough to serve it.
| Self-hosted open model | Standard Compute | |
|---|---|---|
| Models | Self-hosted open model: The open-weight models you deploy | Standard Compute: Frontier models (Claude, GPT, Gemini) and open models (GLM, DeepSeek, MiniMax and more) on one API |
| Start-up time | Self-hosted open model: Weeks: hardware, inference server, monitoring | Standard Compute: Minutes: create a company workspace and a key |
| Cost | Self-hosted open model: GPUs paid whether used or not, plus staff | Standard Compute: Flat monthly plan per person or app, from $19 per key |
| Failover | Self-hosted open model: Yours to build | Standard Compute: Automatic failover to another approved provider of the same model, inside your region and provider rules |
| Training and retention | Self-hosted open model: Under your control | Standard Compute: No training by Standard Compute; zero data retention per company, team or key |
| Processing location | Self-hosted open model: Your data centre | Standard Compute: Approved European inference endpoints under a Europe rule; EU gateway in Finland; account data in Frankfurt |
| Contract | Self-hosted open model: None needed for your own servers | Standard Compute: A DPA you sign in the dashboard: no training, no retention of your content, 72-hour breach notice |
Open models served in Europe
Run leading open models, including GLM, DeepSeek and MiniMax, alongside Claude, GPT and Gemini through one API. GLM, DeepSeek and MiniMax are also served by European providers for Europe-restricted workspaces.
That covers the usual reason for self-hosting an open model in Europe: the model weights are open, and the processing location and retention are fixed by your rules.
Privacy controls and what they cost
Region rules and zero data retention are free. The rest can be bought one at a time or together in Enterprise.
| Control | Standard Compute | Per month |
|---|---|---|
| Europe rule | Standard Compute: Every request, including retries and fallbacks, is served only by approved European inference endpoints. If none can serve it, the request is refused before the prompt is sent | Per month: Free |
| Zero data retention | Standard Compute: Requests go only to inference endpoints that commit to keeping no copy of your data. Teams and keys inherit it and can't switch it off | Per month: Free |
| No training | Standard Compute: Standard Compute never trains on your prompts or outputs. With a region rule or zero retention on, every request goes only to providers that commit not to train on it | Per month: Included |
| Provider allow-lists | Standard Compute: Choose exactly which inference providers may handle your data, per company, team or key. Fallbacks can never route around the list | Per month: $199 |
| SSO | Standard Compute: Microsoft Entra ID or any SAML identity provider for the dashboard | Per month: $399 |
| Signed DPA | Standard Compute: No training, no retention of your content, breach notice within 72 hours | Per month: $299 |
| Enterprise | Standard Compute: Every feature above in one plan, self-serve | Per month: $1,499 |
Prices in USD, from /pricing. Model usage is billed on flat monthly plans per person or app.
Set up a private LLM API in four steps
1. Create a company workspace and turn on zero data retention for the whole company.
2. Set a Europe rule if your data must be processed in Europe.
3. Choose the providers that may serve your data with a provider allow-list.
4. Give every person and app its own key with a monthly budget, and point the OpenAI SDK, Claude Code or any OpenAI-compatible tool at the new base URL.
from openai import OpenAI
client = OpenAI(
base_url="https://eu.stdcmpt.com/v1",
api_key="SC_KEY_WITH_EUROPE_RULE",
)
reply = client.chat.completions.create(
model="anthropic/claude-sonnet-5.5",
messages=[{"role": "user", "content": "Summarise this contract"}],
)Questions from IT, security and finance
- Do we need to self-host to have a private LLM?
- No. Self-hosting is one way to control training, retention, location and who sees your data. Standard Compute gives you the same four controls by policy and contract: no training, zero data retention, a Europe rule and provider allow-lists.
- Can we use Claude and GPT privately, or only open models?
- Both. Claude, GPT, Gemini and open models such as GLM, DeepSeek and MiniMax run on the same key, under the same retention, region and provider rules.
- Does Standard Compute store our prompts?
- No. We process your prompts in memory to serve each request and never write them to storage or logs. Request records hold metadata only (model, provider, tokens, cost and status).
- Where is our data processed?
- With a Europe rule, every request is served only by approved European inference endpoints. Our EU gateway terminates TLS and runs in Finland, and your account data is stored in Frankfurt.
- What does a private LLM API cost?
- Region rules and zero data retention are free. Model usage is a flat monthly plan per person or app, from $19 to $2,499 per key. Enterprise, with every governance feature, is $1,499 a month.
Related pages
- Zero data retention LLM API for companies
- LLM APIs that don't train on your data
- Self-hosted LLM gateway vs a managed gateway
- Open-source models for business
- AI data residency for companies
- Standard Compute for companies