PrivateLink & VPC peering for inference
The fastest way to weaken your data-boundary claims is to route inference traffic over the public internet. AWS PrivateLink for LLM inference eliminates that gap — your application connects to the inference gateway through an interface VPC endpoint, traffic stays on the AWS backbone, and there is no public egress path to audit or worry about. It's the networking layer of private inference, and it's simpler than most teams expect.
Solutions Engineer
Why public internet routing weakens your security posture
When your application sends prompts to an inference API over the public internet, the traffic traverses infrastructure you don't control — ISPs, transit networks, and the provider's load balancers. You can encrypt it, and you should. But the data has left your network boundary before the first token is generated. For regulated workloads, that's a compliance finding. For security-conscious teams, it's an unnecessary attack surface. Private connectivity — PrivateLink, VPC peering, or Direct Connect — removes that surface. Prompts travel over the cloud provider's backbone, inside your VPC, and the only endpoints that can reach the inference gateway are the ones you explicitly authorize.
PrivateLink vs VPC peering vs Direct Connect
Three private connectivity options, each for a different topology. Choosing the right one depends on whether the inference gateway lives in your VPC, in a separate VPC, or across your on-prem boundary:
- AWS PrivateLink. Creates an interface VPC endpoint in your VPC that maps to the inference service. Traffic stays on the AWS backbone and never hits the internet. Private DNS means your application connects to
inference.your-vpc.internaland the endpoint handles the routing. Best when the inference gateway is in a separate VPC or managed account. - VPC peering. Connects two VPCs directly over AWS networking. Useful when you manage both the application VPC and the inference VPC yourself. Simpler than PrivateLink for same-account deployments but requires non-overlapping CIDR ranges and route table management.
- AWS Direct Connect. Extends your on-prem network into AWS over a dedicated physical connection. Use this when the GPUs are in AWS but your application is in your data center — common in hybrid on-prem/VPC deployments.
Setting up an interface VPC endpoint
PrivateLink connects your application to the inference gateway through an elastic network interface in your subnet. The gateway's security group controls which VPC endpoints can reach it. Your application sees a private DNS name, not a public IP. Here's the Terraform for the endpoint side:
resource "aws_vpc_endpoint" "hyperinfer" {
vpc_id = aws_vpc.main.id
service_name = "com.hyperinfer.inference"
vpc_endpoint_type = "Interface"
subnet_ids = aws_subnet.private[*].id
security_group_ids = [aws_security_group.inference.id]
private_dns_enabled = true
}What private DNS gives you
With private_dns_enabled = true, AWS Route 53 Resolver maps the public DNS name of the inference service to the private IP of your VPC endpoint. Your application code connects to the same hostname whether it's running in your VPC or not — but inside the VPC, the traffic resolves locally and never leaves the AWS network. No code changes needed. No separate DNS configuration. You get private routing without your application knowing it happened.
Data residency and the compliance angle
Private connectivity reinforces data residency claims in ways that public internet routing cannot. When traffic stays on the AWS backbone within a single region — Frankfurt to Frankfurt, Singapore to Singapore — you have documentation from the cloud provider that the data never transited a public network. That's concrete evidence for auditors reviewing your GDPR compliance posture. Pair it with single-tenant GPUs in the same region and zero data retention, and you've eliminated the three most common data-residency audit findings: cross-region traffic routing, prompts stored in multi-tenant logging pipelines, and data at rest in a provider's environment. The network layer is only one piece, but it's the piece most teams overlook. See HyperInfer's compliance documentation for the full control mapping.
The alternative you shouldn't use
IP allowlists. They're better than nothing, but they're a perimeter control on a public endpoint — a locked door on a house with windows. A misconfigured load balancer, a rotated IP range, or a traffic-routing change can expose your endpoint without warning. PrivateLink makes the endpoint unreachable from the public internet by construction, not by configuration. For a full VPC deployment walkthrough, the networking step is the step that turns a demo into a production architecture.
Related reading
- GuideDeploy DeepSeek-V4 in your own VPCA step-by-step architecture for running frontier open-weight models on dedicated GPUs inside your cloud account — so prompts never leave your security boundary and nothing is ever logged.
- GuideVPC vs on-prem vs air-gapped inferenceVPC, on-prem, private cloud or air-gapped — break down the four private LLM deployment models across data boundaries, operational burden, and compliance readiness.
- ComplianceGDPR & data residency for LLM inferenceGDPR compliance for LLM inference isn't just about a DPA. It's about controller-processor boundaries, cross-border transfer safeguards and region-pinned data residency — all working together.
- GuideAir-gapped LLM inference architectureHow to run frontier open-weight models with zero external connectivity — local registries, physical-media model updates, and the cryptographic controls for classified workloads.
Run this privately, in your own environment
A solutions engineer will scope a zero-retention deployment for your models and volume.