Plan your hybrid cloud colocation: what moves, what stays, how it connects
Seven quick questions. The planner sorts your workloads between colocation and the cloud, sizes the private links, names the metro and dates the move. No account and no prices.
Which clouds do you use now?
Pick every one that applies.
No account and no prices. Your plan shows as soon as you finish.
Already know what you need? Send us your setup and you will have an answer by email within 24 hours.
Hybrid cloud colocation means running your steady systems on hardware you own inside a rented data center, while the rest stays in AWS, Azure or Google Cloud, and a dedicated network link carries traffic between them without touching the open internet. You keep the cloud for what it does well, move the always-on systems to space where a monthly fee for room and power covers them, and decide one workload at a time.
What hybrid colocation looks like day to day
In hybrid colocation, the colocation provider supplies the building, power, cooling and security, and you supply the servers, storage and network gear that fill your cabinets or cage. From that cage, a cable runs to a cloud provider's edge router in the same building or one nearby, and your traffic reaches your cloud region on the provider's own network.
Three pieces make it work:
- The space. Cabinets or a cage, sized in racks and kilowatts.
- The link. A private connection to each cloud, and two of them for anything that matters.
- The split. A clear rule for which systems sit on each side, so nothing ends up straddling the link by accident.
It is a narrower idea than hybrid cloud in general, which can mean a cloud account plus a server closet. Here, your own hardware sits in a professional data center close to the cloud's front door.
of organizations in Flexera's 2026 survey run hybrid cloud, up three points on the year before.
Source: Flexera, 2026 State of the Cloud Report, March 2026.
of cloud spend is wasted by respondents' own estimate, the first rise in five years, which Flexera ties to fast growth in cloud AI workloads.
Source: Flexera, press release, March 18, 2026.
Which workloads move in a hybrid IT colocation plan, and which stay put
Most hybrid IT colocation plans come down to one question asked of each workload: does it run at a steady size all month, or does it rise and fall? Steady, busy systems suit owned hardware, because a rack costs the same at 3 a.m. as at noon. Spiky or short-lived systems suit the cloud, because you rent the capacity only while you use it.
Four things can override that first sort:
- Data gravity. Big datasets pull the systems that read them toward wherever they sit. Move a database without its main apps and the link ends up in the middle of every query.
- Latency. Anything that talks to your users, or to cloud services hundreds of times a second, needs to stay close to them.
- Licensing. Software sold per server or per core often fits owned hardware better, and some older systems were never built for the cloud.
- Compliance. Regulated data can live on either side, but inside your own cage the physical controls are yours to show an auditor.
Here is where the planner starts each workload before your own answers adjust it:
- Always-on app serversColocationSame size around the clock, so owned hardware stays busy.
- DatabasesColocationWhen their main apps move too; otherwise either side, next to the apps.
- ERP and legacy systemsColocationSteady, often licensed per server, rarely rebuilt for the cloud.
- AI trainingColocationGPUs run near full for weeks at a time.
- Storage and backupColocationWhen traffic is heavy or a DR copy is in play; otherwise either side.
- Virtual desktopsEither sideFollows where your people and their files are.
- DR copyOpposite sideAlways the other side from production.
- Analytics and data warehouseCloudManaged services stretch for big queries, unless the raw data has moved.
- AI inferenceCloudDemand follows users through the day.
- Web front endsCloudAutoscaling and content delivery do the most for them.
- Dev and testCloudStarts and stops all day; pay by the hour.
- Bursty batch jobsCloudRented capacity covers peaks, and the bill stops with the job.
Hybrid cloud vs on premise vs colocation, side by side
Weighing hybrid cloud vs on premise is really a question of who owns each layer, and who pays for capacity that sits idle. On premise means your own building and your own power. All in the cloud means you rent everything by usage. Hybrid with colocation sits between: servers you own, in a rented building someone else runs, linked privately to the cloud.
| Hybrid with colocation | All in the cloud | On premise | |
|---|---|---|---|
| Servers and storage | Yours | The cloud provider's | Yours |
| Building, power, cooling | Run by the colocation provider | Run by the cloud provider | Run by you |
| How you pay | Monthly for space and power, plus hardware, plus cloud usage for what stays | By the hour, the gigabyte and the request | The building, its power and staff, plus hardware |
| Steady, always-on systems | Fit owned hardware well | Billed every hour they run | Fit well if the building has room |
| Sudden growth | Cloud side in minutes, colocation side in weeks | Minutes | Months, often longer |
| Physical control for audits | Yours inside the cage | Shared with the provider | Fully yours |
| Path to the cloud | Private link, often in the same building | Already there | Private link carried by a network provider, usually farther |
If you landed here comparing cloud vs on premise vs hybrid, the table is the short answer: hybrid with colocation keeps the hardware control of on premise without running a building, and keeps the cloud for the parts that change size.
Colocation hybrid cloud links: Direct Connect, ExpressRoute, Interconnect and FastConnect
A colocation hybrid cloud is only as dependable as the private link between your cage and each cloud, so it pays to know what each provider sells before you pick a building. Every provider has its own name for the same idea: a physical port in a data center that hands your traffic to its network without crossing the public internet. When the provider's router sits in your building, the link is a short cable between cages, called a cross-connect. When it does not, a network provider carries the traffic to the nearest location.
| Cloud | Private connection | Sizes the provider offers |
|---|---|---|
| AWS | Direct Connect | Dedicated ports at 1, 10, 100 or 400 Gbps. Hosted connections through a partner from 50 Mbps to 25 Gbps. |
| Azure | ExpressRoute | Circuits from 50 Mbps to 10 Gbps through a provider. ExpressRoute Direct ports at 10, 100 or 400 Gbps. |
| Google Cloud | Cloud Interconnect | Dedicated Interconnect links at 10, 100 or 400 Gbps, with 400 at some facilities. Partner Interconnect attachments from 50 Mbps to 50 Gbps. |
| Oracle Cloud | FastConnect | 1, 10, 100 or 400 Gbps per cross-connect when colocated with Oracle or through a third-party provider; 1, 10 or 100 Gbps through Oracle partners. |
Sources: AWS Direct Connect User Guide, dedicated and hosted connections; Microsoft Learn, ExpressRoute overview; Google Cloud, Cloud Interconnect overview and colocation facility locations; Oracle, FastConnect overview. Checked November 2026.
Two paths, every time
One connection is one point of failure. Each provider's own guidance points the same way:
- AWS rates two connections at more than one location for a 99.9% SLA, and separate connections on separate devices in more than one location for 99.99%.
- Azure builds two connections into every ExpressRoute circuit and recommends two circuits at two peering locations for the most resilient setup.
- Google needs at least four Dedicated Interconnect connections for its 99.99% setup, spread across two edge zones, and across two metros when you use two regions.
Sources: AWS, Direct Connect resiliency toolkit; Microsoft Learn, ExpressRoute overview; Google Cloud, 99.99% availability for Dedicated Interconnect.
Sizing the port
Start from the outbound data line on your cloud bill, then leave room for backups and the first bulk copy. Converted to a steady rate (terabytes times 8 bits, over the seconds in a month), 50 TB a month is about 152 Mbps and 200 TB a month is about 608 Mbps around the clock. Averages hide the nightly peaks, so the planner rounds up to the next standard port.
Pick the metro that matches your cloud region
Your main cloud region decides the metro for your racks, because the shortest path between the two has the least delay and the fewest hands on the fiber. The pairs below come from each provider's own region and connection location lists; we name metros, not buildings.
| Cloud | Region | Metro to look in | Why |
|---|---|---|---|
| AWS | us-east-1 | Northern Virginia | N. Virginia region; Direct Connect locations in Ashburn and Reston |
| AWS | us-east-2 | Columbus | Ohio region; Direct Connect location in Columbus |
| AWS | us-west-2 | Portland and Hillsboro | Oregon region; Direct Connect location in Portland |
| AWS | us-west-1 | Silicon Valley | N. California region; locations in San Jose, Santa Clara and Milpitas |
| Azure | East US, East US 2 | Northern Virginia | Both regions are in Virginia |
| Azure | Central US | Chicago | Region in Iowa; Chicago is the nearest large peering location |
| Azure | South Central US | Dallas | Region in Texas; Dallas is a peering location |
| Azure | West US 2 | Seattle | Region in Washington; Seattle is a peering location |
| us-east4 | Northern Virginia | Google lists Ashburn as its low-latency metro | |
| us-east5 | Columbus | Google lists Columbus as its low-latency metro | |
| us-central1 | Council Bluffs | Google lists Council Bluffs as its low-latency metro | |
| us-south1 | Dallas | Google lists Dallas as its low-latency metro | |
| us-west1, us-west2 | Portland; Los Angeles | Google lists each as the low-latency metro | |
| Oracle | us-ashburn-1, us-chicago-1 | Northern Virginia; Chicago | Named for Ashburn and Chicago |
| Oracle | us-phoenix-1, us-sanjose-1 | Phoenix; Silicon Valley | Named for Phoenix and San Jose |
Sources: AWS, Regions and Direct Connect locations; Microsoft Learn, Azure regions list and ExpressRoute locations; Google Cloud, colocation facility locations; Oracle, Regions. Checked November 2026.
Using more than one cloud
Microsoft notes that an ExpressRoute peering location does not have to match your Azure region: one location reaches every Azure region in the same geopolitical region. Chicago and Dallas appear on the AWS, Azure and Google connection location lists, which makes either one a workable meeting point for a multi-cloud cage. If your clouds point to different metros, start where most of the traffic goes and test latency to the other.
vacancy in Northern Virginia in the first half of 2026, against a record low of 1.4% across North America's primary markets. In the busiest metro, start the search well ahead of your date.
Source: CBRE, North America Data Center Trends H1 2026: Northern Virginia release, September 8, 2026; primary market release.
Hybrid cloud hosting vs colocation: who owns and runs the hardware
Hybrid cloud hosting usually means a provider rents you servers it owns and manages, linked to your cloud, while in colocation the servers are yours and the provider runs only the building around them.
Hosting suits teams without hardware staff, or systems with a short life where buying servers makes little sense. Colocation suits teams that already run their own hardware, want to choose it, or plan to keep it for its full working life. Some setups use both: a few hosted servers for a side system and a cage for the core.
The line to watch is who fixes things. With hosting, a failed disk is the provider's ticket. In colocation it is yours, usually handled through the building's remote hands service at an hourly or monthly rate.
AI workloads in a hybrid cloud and colocation setup
AI shows why teams pair hybrid cloud and colocation so clearly: training keeps GPUs busy for weeks at a time, while serving answers to users rises and falls with demand. One split puts GPU racks you own in colocation for training and keeps inference, the apps that call it and most storage in the cloud. The reverse works too, when serving is steady and training happens now and then.
- The building matters as much as the metro. GPU racks draw several times the power of a typical server rack, and many need liquid cooling, so confirm what each building can support per rack today.
- Size the link for the data, not the users. Training sets and model checkpoints are large. The link between sides carries those, so it needs more headroom than user traffic alone suggests.
- Keep one copy of the truth. Decide which side holds the master copy of each dataset, and sync the other on a schedule.
Moving to a data center hybrid cloud in stages
A data center hybrid cloud move goes best in stages that each prove the last one worked: connect first, move copies next, and leave the databases for the end. The planner dates each stage back from your go-live.
- Requirements and shortlist. Workload list, racks and power, cloud links and dates agreed inside your team, then a short list of buildings.
- Contract and space. Terms signed and the cage or cabinets handed over.
- Private links. Cross-connects and cloud ports ordered, turned up and tested on both paths before anything depends on them.
- Backup and DR copies. The first data across the link, with no users relying on it yet.
- Steady systems. App servers, storage and legacy systems, one group at a time.
- Databases last. The most careful cutover, once everything they serve is settled.
Timing around a cloud commitment renewal
If you have a committed-spend agreement with AWS, Azure or Google, put its renewal date on the same calendar as the move. Whatever you commit to at renewal still covers systems that have not left yet, so a renewal that lands mid-move should be sized for what will stay and for how long the move overlaps. The planner shows where your renewal falls against each stage.
Who this is for, and who should look elsewhere
We help
- Mid-market and enterprise IT teams with a growing cloud bill and steady workloads (databases, storage, ERP, always-on app servers), usually 5 to 50 racks and sometimes a few megawatts, staying connected to AWS, Azure or Google Cloud.
- Teams with a cloud commitment renewing who want part of the spend moved first.
- Companies with data gravity, latency or compliance reasons to keep certain systems on their own hardware.
- AI teams training on their own GPUs in colocation and serving from the cloud, or the other way around.
This site is not for
- Managed service providers, resellers and provider sales teams. There is a short note for you on our page for providers.
- Students, certification study and job searches. The cloud providers' own training sites cover exam topics in far more depth.
- One or two servers, or a single rack with no cloud link. A retail colocation provider can set that up directly.
Who runs this site
The Metro Colo Advisory team runs Hybrid Cloud Colocation. The firm advises companies buying data center space and cloud links, and it isn't owned by or tied to any provider.
- How we're paidColocation providers and cloud connection services pay us when a company we introduce signs with them, so an introduction costs you nothing. Larger buyers planning a partial move off the cloud can hire us for fixed-fee consulting, and we name the fee before any work starts.
- What we do with your detailsWe read your request and answer it ourselves. No provider sees your name or company until you agree, and we tell you which providers before we contact them.
- Where we workData center markets across the US, with most requests landing near the large cloud regions.
- RepliesEvery request gets an answer by email inside 24 hours. If colocation is the wrong move for you, the answer says so.
Questions buyers ask
Is hybrid cloud colocation cheaper than staying in the cloud?
It can be for systems that run at a steady size all month, and it usually is not for systems that spike and go quiet. We don't put a savings figure on a plan before we have seen the bill and the server counts behind it. The planner tells you what would move and how big the footprint is; the cost comparison comes after that, with real numbers.
What is the difference between colocation and a cloud on-ramp?
Colocation is the space, power and cooling you rent for your own hardware. A cloud on-ramp is the cloud provider's own network equipment sitting in a data center, so a short cable from your cage reaches the provider directly. A building can offer both, and that pairing is what keeps the private link short.
Do I need a separate private connection for each cloud?
Each cloud sells its own: AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect and Oracle FastConnect. Some network providers sell one port in your cage that reaches several clouds as virtual connections, which cuts the number of cables. Either way, plan two paths to every cloud that runs something important.
How long does a hybrid colocation setup take?
The planner works with about 31 to 47 weeks from first requirements to the last cutover, depending on size: contract and space, cross-connects and cloud ports, then the move itself in stages. Your own date may be shorter if space and on-ramps are ready in the building you pick, or longer in a tight market.
Which metro should I pick for AWS us-east-1 or Azure East US?
Northern Virginia. AWS lists Direct Connect locations in Ashburn and Reston, and Microsoft places East US and East US 2 in Virginia. CBRE put vacancy there at 0.2% in the first half of 2026, so start the search early.
Can HIPAA or PCI data live in colocation?
Yes. The framework applies to how you protect the data, wherever it sits. In colocation you run the controls on your own hardware, and the provider covers the building: ask each one for its current audit reports and how its physical controls map to yours.
Do you work with managed service providers, resellers or students?
This site is for companies buying space and cloud links for their own systems. If you sell colocation, cloud or managed services, or you are studying for a certification, we are not the right fit, and the form will say so before anything is sent.
What happens after I send my setup?
We read it, check it against what the planner showed you, and reply by email within 24 hours, sometimes with a few questions. No provider hears about you until you say yes, and we tell you which ones first.