Where models run
Hyperscaler AI, self-hosted open-weight models or in-region GPU clouds, behind one gateway.
One gateway in front of every model
A hybrid reference design for regulated workloads: sensitive work stays in your account, other work uses the cloud's model service.
Point at a component to light its links; select it for what it does here and the practice that makes it right on AWS.
Three places a model can run
Chosen per workload, not per company. Many customers use more than one, behind one gateway.
| Option | Fits when | What NGT engineers |
|---|---|---|
| Hyperscaler AI: Amazon Bedrock, Azure OpenAI, Vertex AI | The fastest start; the region meets your data rules | Private endpoints, identity, guardrails, cost tags, pipeline |
| Open-weight models on GPU Kubernetes | Sensitive data must stay in your account or data centre | GPU node groups, serving (vLLM, TGI, Triton), scaling, operations |
| In-region GPU cloud: FPT AI Factory, Token Factory (FPT Smart Cloud) | Data must stay in Vietnam or Japan | Architecture, integration, gateway routing, quotas, operations |
No customer data offshore
For customers whose data may not leave their environment. The offshore team never sees it.
- 01Build on mock data
- Agent core built on mock or synthetic data
- Public models in the build zone
- 02Cross the boundary
- Source and package review
- Licence, SBOM and vulnerability checks
- 03Connect in your zone
- You approve and import the package
- Adapters and data connectors added on your side
- 04Run with real data
- Self-hosted open-weight model in the secure zone
- Your data stays there
One agent, two environments
The same agent code in two zones. Only the models and the data differ.
| Build zone | Customer zone | |
|---|---|---|
| Purpose | Develop and test | Production |
| Models | Public models | Self-hosted open-weight LLM |
| Data | Mock and synthetic | Real customer data |
| Location | FPT delivery environment | The customer's secure environment |
Platforms we build on
Hyperscaler AI, open-source model tooling and FPT group platforms. NGT designs, integrates and runs on them.
Hyperscaler AI services
Open models and serving
vLLMTGINVIDIA TritonAgent frameworks and protocols
DifyMCPA2AFPT group platforms
FPT AI Factory, Token Factory and FPT AI Agents are run by FPT Smart Cloud; CodeVista by FPT Software's AI Center.