Run open LLMs on GPU servers in your own data center.
Anoman Private AI is designed to bring enterprise GPU servers, the Anoman gateway and guardrails, and curated open models into your data center — for AI workloads you want to keep on your own network. Offered as early access: every deployment is scoped and sized with you.
Configurations
Four sizes, from proof of concept to the largest open models
Reference configurations for enterprise GPU servers.
1–2× NVIDIA RTX PRO 6000
96–192 GB of GPU memory. Models of roughly 8–32B parameters; gpt-oss-120b fits on a single 96 GB card. For proofs of concept and edge sites.
4× NVIDIA RTX PRO 6000
384 GB of GPU memory. 70B models in FP8 for production, 120B mixture-of-experts models, and LoRA fine-tuning. Our default recommendation for on-premise inference.
8× NVIDIA RTX PRO 6000 or 8× NVIDIA H200 NVL
768 GB or 1.13 TB of GPU memory. 235B mixture-of-experts models in FP8 and LoRA fine-tuning of 70B models; the H200 NVL configuration fits 671B models in FP8.
NVIDIA HGX B200 or HGX B300
1.44 TB or 2.3 TB of GPU memory. 400–671B models and full fine-tuning of 70B models. High-density systems, usually placed in a colocation data center.
What's included
What Private AI is designed to include
Anoman gateway and guardrails, on-premise
An OpenAI-compatible API with prompt-injection detection, PII redaction, access control and audit logs, designed to run next to your models.
Curated open models
Open-weight models such as Sahabat-AI and SEA-LION (built for Indonesian and Southeast Asian languages), Qwen and gpt-oss.
Cloud GPUs for development and testing
Development and testing on cloud GPUs in Jakarta can be arranged during scoping, before your own hardware is installed.
Deployment and managed service
Anoman Professional Services installs and hardens the stack, and can run it with you after go-live.
Anoman Flow, optional
Add workflow automation on the same servers once Flow's private edition is available.
FAQ
Questions about Private AI
Which models can Private AI run?
Open-weight models such as Sahabat-AI, SEA-LION, Qwen and gpt-oss. The configurations above list indicative model sizes per tier; we confirm sizing for your models and number of users before any order.
Can we start before buying hardware?
Yes. Development and testing on cloud GPUs in Jakarta can be arranged during scoping, or begin with an AI Readiness Assessment or a proof of concept from Anoman Professional Services.
How is Private AI priced?
Per project. We quote hardware, software and services together after scoping your requirements.
Are there restrictions on who can buy GPU servers?
Advanced NVIDIA GPUs are subject to US export controls, so every hardware order includes end-user screening.
Planning private AI? Talk to us.
Tell us which models you want to run, for how many users, and where the servers will live.