Connecting on-premises systems to a cloud platform takes two decisions, one about the network link and one about how you operate both sides. For steady, high-volume traffic, the best connection is a dedicated private link such as AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect, or Oracle FastConnect, often provisioned through a colocation provider. A site-to-site IPsec VPN is the backup path and the fastest way to start. When the goal is one operating model across both sides, Control Plane runs workloads active-active across cloud regions and on-premises locations and handles Day 2 operations, while platform extensions such as AWS Outposts, Azure Local with Azure Arc, and Google Distributed Cloud run a provider’s services on your hardware.
Disclosure: Control Plane publishes this guide.
TL;DR
The best hybrid cloud connection depends on what you are moving and how much of it moves.
- A dedicated private link (Direct Connect, ExpressRoute, Google Cloud Interconnect, Oracle FastConnect) suits sustained, high-volume, or latency-sensitive traffic.
- A site-to-site IPsec VPN is the quickest to deploy and works as a backup path or for lighter traffic.
- Dedicated links are unencrypted by default. Add MACsec or IPsec where your policy requires encryption on every link.
- Private service endpoints (AWS PrivateLink, Azure Private Link, Google Private Service Connect) let on-premises clients reach a specific cloud service on a private IP, with no public endpoint, over a VPN or interconnect.
- Platform extensions such as AWS Outposts, Azure Local, and Google Distributed Cloud run cloud services on your hardware, so they change the operating model as well as the network.
- A nightly backup, a real-time hybrid app, and a regulated database each need a different connection pattern.
- Control Plane’s Managed Kubernetes joins on-premises servers as Control Plane locations, so workloads run active-active across cloud regions and on premises with Day 2 operations handled. Cloud Wormhole, Control Plane’s private connectivity feature, lets those workloads reach private databases and services through an outbound agent connection with no inbound ports opened. Both complement dedicated interconnects and VPNs.
What are the best hybrid cloud solutions for connecting on-premises systems to a cloud platform?
The right connection depends on data volume, latency sensitivity, and whether you need cloud-consistent operations on both sides.
| Option | How it works | Bandwidth | Setup time | Security | Cost profile | Best fit |
|---|---|---|---|---|---|---|
| Site-to-site IPsec VPN (each cloud’s VPN gateway) | Encrypted IPsec tunnels over the public internet between your gateway and the cloud VPN endpoint | Roughly 1 to 3 Gbps per tunnel (AWS 1.25 Gbps standard or 5 Gbps large-bandwidth; Google 1 to 3 Gbps; Azure up to about 2.3 Gbps depending on SKU); aggregate tunnels with ECMP | Minutes to hours | Encrypted by default (IPsec) | Lowest upfront; hourly gateway fees plus internet-rate egress | Dev and test, branch offices, backup path, light or bursty traffic |
| Dedicated interconnect (AWS Direct Connect dedicated, Azure ExpressRoute Direct, Google Cloud Dedicated Interconnect, Oracle FastConnect) | Physical cross-connect at a colocation facility to the cloud provider’s router, giving a private circuit | 1 Gbps to 400 Gbps per port (AWS and Oracle start at 1 Gbps; ExpressRoute Direct and Google Dedicated Interconnect start at 10 Gbps) | Weeks to months (port order, letter of authorization, cross-connect) | Unencrypted by default; MACsec on 10 Gbps and faster ports at supported locations, plus IPsec overlay | Port fee; inbound free; outbound at reduced rates (AWS, Google), metered or unlimited plans (Azure), or no transfer charge (Oracle) | Sustained high-volume transfers and latency-sensitive production workloads |
| Partner or hosted interconnect through a colocation or network provider | A network provider delivers a virtual circuit to the cloud over its existing physical connections | From 50 Mbps to a provider maximum (Azure 10 Gbps, AWS hosted 25 Gbps, Google Partner Interconnect 50 Gbps) | Days to weeks | Unencrypted by default; IPsec overlay recommended | Smaller commitment than dedicated; partner fees plus transfer charges | Organizations without equipment in a cloud provider’s interconnect facility |
| SD-WAN with cloud on-ramps (Azure Virtual WAN, AWS Cloud WAN, Google Network Connectivity Center) | SD-WAN appliances at branches connect to cloud WAN hubs that manage routing and policy | Depends on the underlay internet or MPLS links | Days to weeks for the SD-WAN overlay; cloud hub setup in hours | Encrypted tunnels between SD-WAN edges and cloud hubs | SD-WAN licensing plus cloud WAN hub and data processing fees | Multi-branch enterprises moving from MPLS to internet with cloud-first routing |
| Private service endpoints (AWS PrivateLink, Azure Private Link, Google Private Service Connect) reached over VPN or interconnect | A private IP address in your VPC or VNet exposes one cloud service or your own service | Limited by the underlying VPN or interconnect | Hours, once the underlying link exists | Service reached on a private IP with no public endpoint; encryption comes from the VPN or from TLS | Endpoint-hour and data processing fees | Exposing one service, such as an API or a managed database, to on-premises consumers |
| Agent-based outbound tunnels (application-level connectors that call out from the private network) | An agent inside your network opens an outbound connection to the cloud service, so the service can reach hosts behind your firewall | Limited by the agent host and tunnel overhead | Minutes to hours | Outbound only; no inbound firewall rules; encrypted in transit | Low infrastructure cost; licensing varies | Cloud workloads that call a few private databases or internal APIs |
| Storage and data gateways (AWS Storage Gateway, Azure File Sync, or similar) | A virtual appliance or agent caches cloud storage locally and synchronizes data | Optimized for storage protocols (NFS, SMB, iSCSI) | Hours to days | Encrypted in transit and at rest by default | Gateway cost plus cloud storage and transfer fees | Backup, archive, file share migration, and storage tiering |
How do you choose a connection for what you are connecting?
The workload, the data volume, and the compliance constraints decide the connection more than the cloud provider does.
| What you are connecting | Recommended connection | Watch out for |
|---|---|---|
| Users and branch offices to cloud apps | SD-WAN with cloud on-ramps, or a site-to-site VPN per branch | Managing a separate VPN per branch gets hard as sites grow; SD-WAN centralizes policy but adds licensing cost |
| On-premises database to cloud applications | Dedicated or partner interconnect routed into the VPC or VNet (Transit Gateway, VNet gateway, or Cloud Router) | Chatty queries multiply round-trip latency; keep the app and database in the same metro or add a cache |
| Nightly bulk data transfer or backup | Dedicated interconnect for large recurring volumes; VPN or storage gateway for smaller volumes | For one-time seeds of tens of terabytes or more, use AWS Snowball, Azure Data Box, or Google Transfer Appliance; use parallel streams, because one TCP flow rarely fills a large link |
| Real-time hybrid apps with tight latency needs | Dedicated interconnect with two connections in separate facilities near the cloud region, and a second metro for 99.99% availability | Every network hop adds latency, so keep chatty, latency-critical components on the same side |
| Regulated data that must stay on premises | Agent-based outbound tunnel or routed interconnect with strict segmentation; consider a platform extension | Encryption in transit must meet your policy; keep private hostnames off public resolvers |
| Legacy VMs being migrated over months | Dedicated interconnect for bulk migration; VPN as interim for light traffic | Overlapping IP ranges block routing; re-IP, or NAT through AWS private NAT gateway, Azure VPN Gateway NAT, or Private Service Connect |
| Kubernetes clusters on both sides | Dedicated interconnect or VPN plus cross-cloud private networking patterns; consider a multi-cluster mesh | Pod CIDRs must be routable or NATed and service CIDRs stay cluster-internal; advertising per-node pod ranges can exceed BGP prefix limits |
| A few internal APIs that a cloud workload must call | Agent-based outbound tunnel or routed private connectivity over VPN | Opening inbound ports for a handful of API calls adds risk that an outbound tunnel avoids |
How do you make the on-premises connection resilient?
A single link to the cloud is a single point of failure, and each cloud provider publishes guidance on how to avoid it.
Two links in different locations
Every provider recommends at least two physical connections, and connections in different facilities for the highest availability. The AWS Direct Connect Resiliency Toolkit is a connection wizard whose Maximum Resiliency model orders separate connections on separate devices in more than one location for a 99.99% SLA. Azure ExpressRoute provisions every circuit with primary and secondary connections run active-active and publishes high-availability guidance, plus separate disaster-recovery guidance for geo-redundant circuits. Google Cloud Interconnect defines 99.9% and 99.99% topologies, and 99.99% needs four connections across two metros.
VPN backup over a dedicated link
A site-to-site VPN in parallel with a dedicated interconnect is standard practice. For the same prefix, AWS and Azure prefer Direct Connect or ExpressRoute over a VPN automatically. On Google Cloud, set a worse advertised route priority on the VPN’s BGP sessions. On premises, give the dedicated link the higher local preference. A backup VPN usually carries far less than a 10 Gbps circuit, so decide which traffic gets priority during failover, and keep both directions on the same path, because asymmetric routing breaks stateful firewalls.
BGP routing and failover
All four major providers use BGP over dedicated interconnects. Advertise the same prefixes over both links. Set a higher local preference on routes learned from the dedicated link to steer traffic leaving your network. To steer return traffic, prepend your AS path on the backup or use provider controls such as AWS local-preference BGP communities or Google Cloud advertised route priority. Never advertise more-specific prefixes over the backup, because longest prefix match overrides all of these.
Default BGP hold times of 60 to 180 seconds (90 on AWS, 180 on Azure, 60 on Google Cloud) delay failure detection. Bidirectional Forwarding Detection (BFD) detects a failed link in under a second on AWS and Azure and in about five seconds on Google Cloud, and end-to-end failover then takes a few more seconds while routes withdraw. Enable BFD on your own router as well.
MTU and route limits
Check MTU on every path. Direct Connect private virtual interfaces support jumbo frames of 9001 bytes and transit virtual interfaces 8500, ExpressRoute uses 1500, and Google Cloud Interconnect allows up to 8896. IPsec tunnels need a smaller MTU of about 1400 with MSS clamping, so failing over from a jumbo-frame link to a VPN drops large packets unless path MTU discovery and clamping work. Each provider also limits how many prefixes a BGP session accepts, for example 100 on an AWS Direct Connect private or transit virtual interface, and exceeding the limit takes the session down, so summarize routes.
Monitoring
Monitor link utilization, BGP session state, packet loss, and latency on every circuit. Amazon CloudWatch metrics for Direct Connect, Azure Monitor for ExpressRoute, and Google Cloud Monitoring metrics for Cloud Interconnect give per-connection telemetry. Alert on BGP flaps, latency changes from your baseline, and utilization above 50% on either link of a redundant pair, because the surviving link must carry the full load.
Do you need to bring the cloud on premises?
A platform extension changes how you operate, so it fits a narrower set of cases than a network link.
| Platform | What runs on premises | Managed by | Best fit |
|---|---|---|---|
| AWS Outposts | EC2, EBS, ECS, EKS, RDS, and S3 on AWS-owned racks (Outposts servers support EC2 and ECS only) | AWS (hardware, software, patches); needs a continuous link to its parent Region | Teams standardized on AWS APIs that need local compute and storage |
| Azure Local (formerly Azure Stack HCI) with Azure Arc | Azure Local VMs enabled by Azure Arc and AKS on your hardware | Shared; Microsoft for the software, customer for the hardware | Windows-heavy and Azure-native teams needing hybrid VMs and containers |
| Google Distributed Cloud (connected and air-gapped) | GKE and VM workloads on Google-managed hardware (connected) or Google-supplied hardware run fully disconnected (air-gapped); a software-only edition runs on your servers | Google in connected; customer or a trusted partner in air-gapped, with Google-supplied hardware and software; customer hardware for software-only | Kubernetes-first teams and regulated or air-gapped environments |
| Oracle Cloud@Customer and Dedicated Region | Full OCI services (Dedicated Region) or a subset such as Exadata Database and compute (Cloud@Customer) | Oracle (hardware, software, operations) | Oracle Database and enterprise apps with strict residency needs |
| VMware Cloud Foundation | vSphere, vSAN, NSX, and Kubernetes on customer hardware, with cloud options through VMware Cloud on AWS, Azure VMware Solution, or Google Cloud VMware Engine | Customer, with Broadcom support | Organizations with deep VMware investments extending to the cloud |
| Nutanix Cloud Platform | Nutanix AHV or ESXi, Kubernetes, and one management plane across on-premises and cloud | Customer, with Nutanix support | Hyperconverged infrastructure teams wanting one management plane |
| Red Hat OpenShift | Kubernetes on bare metal, VMs, or cloud with a consistent developer experience | Customer, with Red Hat support | Kubernetes standardization across data centers and clouds |
Platform extensions make sense when data residency rules keep data inside a facility, when an application needs lower latency to local systems than any WAN link can give, or when local processing of sensor or telemetry data cuts transfer costs. Fully disconnected operation calls for an air-gapped product such as Google Distributed Cloud air-gapped, because Outposts depends on its Region link. You take on hardware lifecycle management, physical security, and capacity planning that a public cloud region would otherwise handle. The 9 best hybrid cloud platforms in 2026 compares these platforms in depth.
What security controls does a hybrid connection need?
A hybrid connection joins two trust zones, so the controls must cover the link and the traffic on it.
Encryption over private links
Dedicated interconnects are isolated from other traffic and unencrypted by default. PCI DSS requires strong cryptography for cardholder data sent over open, public networks, and a private dedicated circuit is generally outside that requirement, but many organizations and assessors still expect encryption on every link. Where your policy requires it, add MACsec (IEEE 802.1AE) at the port or run IPsec over the circuit. MACsec protects only the hop between your router and the provider’s edge, and it needs MACsec-capable optics and a pre-shared key. A standard AWS IPsec tunnel tops out near 1.25 Gbps, so encrypting a 10 Gbps circuit needs several tunnels with ECMP through a Transit Gateway. Application-layer TLS stays the last line of defense either way.
DNS resolution across both sides
Hybrid DNS is a frequent source of misconfiguration. Cloud workloads that resolve on-premises hostnames need conditional forwarding, and on-premises resolvers need forwarding for cloud private DNS zones. Amazon Route 53 Resolver endpoints, Azure DNS Private Resolver inbound and outbound endpoints with forwarding rulesets, and Google Cloud DNS inbound server policies and outbound forwarding zones all support this. Private-endpoint DNS zones, such as the privatelink zones on Azure, need conditional forwarders from on premises.
Firewall and segmentation
Treat the hybrid link as a boundary between zones. Apply stateful firewall rules on both sides, allow only known protocols and ports, and segment on-premises networks so a compromised host cannot reach cloud VPCs unchecked. Cloud network security groups and on-premises firewall policies should express the same intent.
Identity federation
SAML 2.0 or OpenID Connect federation signs users in once across both sides. Workload identity federation (AWS IAM Roles Anywhere, Azure workload identity federation, Google Workload Identity Federation) gives on-premises services short-lived cloud credentials in place of stored keys.
Logging
Log every connection that crosses the hybrid boundary. VPN gateways and interconnects expose BGP session events and per-link metrics. Send these and cloud flow logs (VPC Flow Logs on AWS and Google Cloud, virtual network flow logs on Azure) to a central SIEM, and keep them for the period your compliance framework sets. The hybrid cloud security and compliance risks guide covers the wider risk list and mitigations.
How Control Plane fits
Control Plane is the unbreakable platform for running production workloads. It runs serverless, standard, stateful, and cron workloads natively on AWS, GCP, Azure, and Oracle Cloud Infrastructure (OCI), VM workloads on locations joined through Control Plane’s Managed Kubernetes, and workloads on Kubernetes clusters you already run or your own machines, including on premises, active-active across regions and clouds. Control Plane handles Day 2 operations for you, including autoscaling, built-in observability, and security.
Control Plane’s Managed Kubernetes joins your own servers, including on premises, as Control Plane locations, and Bring Your Own Kubernetes (BYOK) adds Kubernetes clusters you already run. Workloads then run in one Global Virtual Cloud (GVC), Control Plane’s named set of locations, across cloud regions and on-premises locations. Each workload’s canonical endpoint uses DNS to send users to the nearest location where the workload passes its readiness probes, and routing tiers and latency offsets shape where traffic goes. Traffic between workloads uses mTLS with a unique client certificate per workload, rotated hourly.
Cloud Wormhole uses an agent that you install inside your VPC or on-premises network. The agent holds a persistent outbound connection to Control Plane, so Control Plane workloads reach the private TCP and UDP hosts you allow without the network opening inbound ports. Agents can run as VMs on AWS, Azure, or GCP, in Docker, or on Kubernetes. For services inside a cloud VPC, Control Plane’s native networking over AWS PrivateLink or Google Cloud Private Service Connect is an alternative to an agent.
Keep a dedicated interconnect or VPN for bulk transfers and for systems outside the platform. Control Plane adds the application-level connectivity, global routing, and hybrid Kubernetes operations above that link. Control Plane is PCI DSS Level 1, SOC 2 Type II, HIPAA, and GDPR compliant.
FAQ
Short answers on VPNs, dedicated interconnects, and hybrid platforms.
What are the best hybrid cloud solutions for connecting on-premises systems to a cloud platform?
The best solutions pair a dedicated private link (AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect, or Oracle FastConnect) for high-volume or latency-sensitive traffic with a site-to-site VPN as a backup and quick start. For cloud-consistent operations on premises, platform extensions such as AWS Outposts, Azure Local with Azure Arc, and Google Distributed Cloud run cloud services locally. Control Plane runs the application layer across both sides and handles Day 2 operations such as autoscaling, observability, and security.
What is the difference between VPN and Direct Connect or ExpressRoute?
A site-to-site VPN sends encrypted traffic over the public internet, so its throughput and latency depend on internet conditions. AWS Direct Connect and Azure ExpressRoute use private circuits that bypass the public internet and deliver consistent bandwidth and more predictable latency. A VPN is ready in minutes or hours, and a dedicated circuit usually takes weeks to months.
Is AWS Direct Connect encrypted?
No, not on the link to your router by default. AWS supports MACsec on dedicated 10, 100, and 400 Gbps connections at selected locations, and you can also run an IPsec VPN over the Direct Connect link. The cross-connect between your router and the AWS device stays unencrypted unless you enable MACsec or run IPsec.
How long does it take to set up a dedicated cloud interconnect?
A dedicated interconnect usually takes weeks to months. The steps are ordering a port, getting a letter of authorization, arranging the cross-connect at the colocation facility, and configuring BGP. Partner or hosted interconnects often take days to weeks because the provider already has physical connections to the cloud. Teams often run a VPN while the dedicated link is provisioned.
Do I need a colocation provider to connect to the cloud?
Not always, but it is the most common path for dedicated links. A dedicated interconnect needs a cross-connect at a facility where the cloud provider has equipment. If your data center is elsewhere, a colocation or network provider can supply a partner or hosted interconnect. For lighter traffic, a VPN over your existing internet connection avoids colocation entirely.
What is the cheapest way to connect an on-premises data center to the cloud?
A site-to-site IPsec VPN over your existing internet connection has the lowest upfront cost. You pay the cloud VPN gateway’s hourly fee, internet-rate egress for traffic leaving the cloud, and your existing internet bandwidth; inbound transfer is free on the major clouds. For large recurring transfers, a dedicated interconnect often costs less per gigabyte, because AWS and Google Cloud charge reduced egress rates over it, Azure offers metered or unlimited plans, and Oracle FastConnect charges only for the port. Add colocation cross-connect and carrier fees before comparing.
How do I connect on-premises Kubernetes to the cloud?
Start with a dedicated interconnect or VPN for the network layer, and plan CIDRs so pod and service ranges do not overlap or exceed BGP prefix limits. For the orchestration layer, the options are a multi-cluster service mesh, a platform extension such as Google Distributed Cloud or Azure Local with AKS, or Control Plane, which turns on-premises servers into a managed cluster or adds a cluster you already run and treats either as a location alongside cloud regions.
How does Control Plane connect on-premises systems to the cloud?
Control Plane connects in both directions. A Cloud Wormhole agent you install on your network dials out to Control Plane, so Control Plane workloads can call hosts behind your firewall. For compute, on-premises servers join as Control Plane locations, so the same workloads also run next to your data, and a dedicated link or VPN still carries bulk transfers and systems outside the platform.
Can I use more than one cloud provider’s interconnect at the same time?
Yes. Many enterprises keep dedicated interconnects to two or more providers, each ending at its own router in the colocation facility. Google Cross-Cloud Interconnect and the Oracle Interconnect for Microsoft Azure also link clouds directly without your own colocation router. Cross-cloud private networking patterns cover connecting VPCs and VNets across providers.
Run workloads active-active across cloud regions and on-premises locations. Talk to the Control Plane team or sign up free.
Originally published on the Control Plane blog.
Top comments (0)