On GPUs you control.
We size the hardware for the models you need, then source it at cost with no markup. Hardware is billed separately from our work.
- One H100 80GB ≈ $25–30k
- A 4×H100 node ≈ $180–240k
- Smaller models far less
What an on-prem deployment includes.
Five things run inside your network when we deploy: the GPU node, the model, an OpenAI-compatible endpoint, access control and an audit log. Here they are in the order we build them.
Scroll to fly inside the perimeter
Five things you own, inside your perimeter.
GPU node, model, endpoint, access control and audit log all run inside your network, on hardware you control.
The model runs on your hardware.
Model selection and GPU sizing come first, then quantization, batching and caching so it serves fast on the hardware you have.
- Llama 4
- Mistral
- Qwen3
- Command A
- DeepSeek V3 / R1
- gpt-oss
- Gemma 3
- Gemma 4
- Qwen3-Coder
- Devstral
- Qwen3.8
- Unlimited-OCR
- DeepSeek-OCR
- GLM-OCR
- BGE
- all-MiniLM
- Whisper
- Voxtral
Same SDK. Change the base URL.
client = OpenAI(
# inside your network
base_url="https://llm.yourcompany.internal/v1",
api_key="your-key",
)Your applications keep their code. The endpoint lives inside your network. Tested-integration matrix on request.
Try it on your data in the 72-hour pilotYou decide who can call which model.
Role-based access control, monitoring and audit logging ship with every production deployment.
| Role | Model | Decision |
|---|---|---|
| support-agents | gpt-oss-20b | allow |
| legal-review | llama-4-scout | allow |
| contractors | llama-4-scout | deny |
Every call is on the record.
Requests are logged inside your network. With Managed Ops we watch it with you: 99.95% uptime SLA, 5-minute incident response, 24/7 across Limassol and Auckland.
| Time | Principal | Model | Route |
|---|---|---|---|
| 09:41:07 | svc-crm-•• | qwen3-32b | internal |
| 09:41:05 | u-ops-•• | gpt-oss-120b | internal |
| 09:41:02 | svc-docs-•• | llama-4-scout | internal |
Pilot, then production, then 24/7 operations.
Each step is priced up front. The pilot readout scopes production, and production hands over to managed operations. We size the hardware and source it at cost, with no markup.
Pilot
$4,900
fixed
- One open-weight model live on your data in 72 hours. The clock starts after access and prerequisites.
- Reference evaluation on your data.
- Readout and production scoping.
Production deployment
from$25,000
one-time, hardware separate
- Model selection and GPU sizing.
- Monitoring, RBAC, audit logging.
- Handover and documentation.
Managed ops
from$1,500
per GPU node / month
- Model and driver updates, monitoring, on-call, SLA tiers.
- 99.95% uptime SLA and 5-minute incident response.
- 24/7 across Limassol and Auckland.
Talk to an engineer.
Tell us the model, the data and the GPUs you have. An engineer from Limassol or Auckland scopes the 72-hour pilot with you.