Capability 02 — Deployment
Three ways to run your AI. All private.
Every deployment mode we offer keeps your data out of consumer chatbots. No training on your prompts. No leaking to public models. The difference is where the hardware lives and who manages it.
Deployment modes
The three modes.
On-premises
Open-weight models deployed on GPU servers inside your building. Data never leaves. You own the hardware, we configure and deploy.
- Models — DeepSeek, Llama, Mistral, Qwen, GLM
- Hardware — NVIDIA H100/H200/B200, AMD MI300X
- Serving — vLLM, SGLang, TensorRT-LLM
- Privacy — Complete. Air-gapped option available
Private GPU cloud
Open-weight models running on dedicated GPU instances in a private cloud. The GPUs are yours alone — no sharing with other customers.
- Models — Same open-weight catalog
- Privacy — Dedicated instances, US-based infrastructure
- Cost — Monthly operational. No hardware purchase
Via API
Access the most capable models in the world through API endpoints we configure and secure. Or serve your own open-weight model through hosted endpoints.
- Models — Claude, GPT, Gemini, Grok
- Privacy — Contractual: under the enterprise agreements we configure, providers do not train on your data
- Cost — Pay per token. Lowest barrier to entry
Privacy
Privacy across all modes.
| Concern | On-prem | Hosted Dedicated | Via API |
|---|---|---|---|
| Data used to train public models? | No | No | No (enterprise terms) |
| Data leaves your jurisdiction? | No | No (US-based infrastructure) | No (enterprise terms) |
| You control the model weights? | Yes | Yes | No (frontier) / Yes (open-weight served) |
| Audit trail available? | Yes | Yes | Yes |
Decision guide
How to choose.
| On-prem | Hosted Dedicated | Via API | |
|---|---|---|---|
| Upfront cost | High (hardware purchase) | None | None |
| Ongoing cost | Low (power, maintenance) | Medium (monthly) | Variable (per token) |
| Control | Maximum | High | Low–Medium |
| Intelligence ceiling | Open-weight | Open-weight | Frontier (Claude, GPT, Gemini) |
| Best if | Data must stay on-site, long-term high volume | You want control without CAPEX, predictable spend | You need max intelligence, variable or low volume |
Many of our clients start with one mode and evolve. A bank might prototype via API, move to hosted dedicated for production, and eventually bring it on-prem. We handle all three transitions.
Contact
Not sure which deployment fits?
We'll walk through your requirements — data sensitivity, budget, use case, timeline — and recommend the right architecture.
Find your deployment model →