On-premise vs cloud: total cost of ownership
On this page
- Why most TCO comparisons are wrong
- What on-premise actually costs
- What cloud actually costs
- A worked example: 20 VMs and 2 databases, 5 years
- When cloud is NOT cheaper
- FAQ
- Is the cloud actually cheaper than on-premise?
- What on-premise costs do people forget in TCO comparisons?
- What cloud costs surprise people after migrating?
- Should we keep anything on-premise?
Every vendor in this debate has a TCO calculator that proves their side wins. Here is the comparison with both thumbs off the scale: what on-prem really costs, what cloud really costs, and a worked example you can adapt to your own estate.
Why most TCO comparisons are wrong
They compare a server’s purchase price to a cloud instance’s monthly rate and stop. Both sides hide their real costs in other budgets:
- On-prem hides costs in salaries, facilities, and future capex. The server was $12,000 once; the power, the licensing, the refresh in year five, and the fraction of an engineer who babysits it never appear on the same spreadsheet.
- Cloud hides costs in meters and growth. The instance is $70/month; the egress, the backups, the managed-service premium, and the resources nobody deleted appear on next month’s bill.
An honest comparison prices both sides fully, over the same 5-year window, at realistic utilization. That last clause matters most: on-prem hardware is bought for peak plus growth and typically runs at 15 to 30% utilization, while cloud lets you pay closer to actual usage, if you have the discipline to.
What on-premise actually costs
For each category, ask “what would we spend over 5 years,” not “what did the box cost”:
| Category | What’s in it | Typical share of true TCO |
|---|---|---|
| Hardware | Servers, storage, network gear, plus refresh every 4 to 6 years and spares | 20 to 30% |
| Software licensing | Virtualization (VMware’s repricing sharply raised this line), OS, database, backup tooling | 10 to 20% |
| Facilities | Colocation fees or power, cooling, UPS, physical security for a server room | 10 to 15% |
| Connectivity | Business circuits, redundant links, firewalls | 5 to 10% |
| Disaster recovery | A second site or DR service; if you don’t pay this, price the risk instead | 5 to 15% |
| People time | Patching, hardware swaps, capacity planning, backup verification, upgrade projects | 25 to 40% |
Two on-prem specifics deserve emphasis. Overprovisioning is structural: you buy for the busiest projected hour of year four, so most capacity idles for most of its life. And people time is the buried lede: even half an engineer’s time on infrastructure care is $40,000 to 70,000/year that a TCO comparison must count, because after migration that time goes back into product work (or the role’s next hire doesn’t happen).
What cloud actually costs
The cloud side of the ledger, honestly stated:
- The run rate, which you should estimate properly before migrating, and which drops 25 to 40% below on-demand list with rightsizing and commitments.
- The meters people forget: egress at roughly $0.09/GB to the internet, NAT and cross-zone traffic, per-request charges, backup and snapshot storage growing unattended.
- The managed premium. RDS costs more than the same Postgres on a VM; you’re buying back DBA hours. That’s usually a good trade and it belongs on the ledger as a cost, not hidden as zero.
- Migration itself: one-time engineering, parallel-running both environments during cutover, and team training. Budget 10 to 30% of first-year run cost for a typical estate.
- Growth drift. Creating resources is frictionless, so ungoverned accounts drift up 5 to 10% per quarter. Governance (tagging, budgets, monthly review per our bill shock guide) is a standing cost of doing cloud well.
A worked example: 20 VMs and 2 databases, 5 years
A typical SMB estate: 20 production and internal VMs, 2 database servers, ~30 TB of storage, currently on 3 aging hypervisor hosts and a SAN in a colo rack.
On-premise, annualized over 5 years:
| Item | $/year |
|---|---|
| Hardware refresh (3 hosts + SAN + network, ~$90k every 5 yrs) | $18,000 |
| Virtualization + OS + backup licensing | $15,000 |
| Colo rack, power, connectivity | $18,000 |
| DR (minimal warm standby at a second site) | $10,000 |
| People time (0.5 FTE sysadmin, loaded) | $50,000 |
| Total | ≈ $111,000/yr |
Cloud equivalent, annualized:
| Item | $/year |
|---|---|
| 20 rightsized VMs + 2 managed databases (with 1-yr commitments) | $58,000 |
| Storage (30 TB, tiered) + backups | $9,000 |
| Egress + networking (moderate traffic) | $5,000 |
| Remaining ops time (0.15 FTE: cloud doesn’t run itself) | $15,000 |
| Governance and tooling | $3,000 |
| Run rate | ≈ $90,000/yr |
| One-time migration (~$60k amortized over 5 yrs) | $12,000 |
| Total | ≈ $102,000/yr |
Verdict: roughly 8% cheaper in raw dollars, and the honest summary is that the dollar gap is modest while the structural gains are not: DR went from “minimal” to tested and real, capacity became elastic, the hardware-refresh capex spike disappeared, and 0.35 FTE of engineering time moved back to product work. If that time has any product value, the effective gap widens well past 20%. This is the typical shape: cloud wins on utilization, people, and risk more than on the sticker price.
Run the same numbers with sloppy cloud practice (no rightsizing, on-demand rates, unmanaged growth) and cloud costs $130,000+/year and loses. Discipline is not optional; it’s where the savings live.
When cloud is NOT cheaper
Saying this out loud is the point of an honest comparison:
- High-utilization, stable compute on owned hardware. A workload that genuinely keeps servers busy 24/7 (render farms, some HPC, mature high-traffic monoliths with flat load) removes cloud’s utilization advantage. Owned hardware run hot is very cheap per unit of work.
- Egress-heavy workloads. Serving hundreds of TB monthly to the internet at $0.09/GB dwarfs any compute savings. Video platforms and backup services either stay on-prem, use a CDN aggressively, or pick providers with structurally cheap egress like OCI.
- Very large, stable storage. Petabytes that mostly sit still can cost less on owned disks than on cloud storage billed monthly, forever.
- Sunk staffing you’d keep regardless. If the team maintaining the servers isn’t redeployable or reducible, the people line doesn’t actually move, and the TCO case weakens accordingly. Count what changes, not what’s theoretically assignable.
- Punitive cloud licensing. Some vendors price their software per cloud core in ways that make the cloud math ugly. Check your licenses before you check the instance prices.
None of these mean “never migrate”; they mean those specific workloads may belong in the on-prem column of a hybrid design, chosen deliberately per the migration strategy rather than by inertia.
FAQ
Is the cloud actually cheaper than on-premise?
For most SMB and mid-market estates, yes, by 20 to 40% over 5 years, but only when the comparison is honest on both sides and the cloud side is run with discipline (rightsizing, commitments, off-hours shutdown). For stable, high-utilization workloads on hardware you already own, with admin staffing you’d keep anyway, on-prem frequently wins on pure dollars. The cloud’s advantage is largest where utilization is low and people time is the hidden cost.
What on-premise costs do people forget in TCO comparisons?
The big five: hardware refresh (servers and storage age out every 4 to 6 years), power and cooling, software licensing (virtualization especially, after recent price increases), a real disaster-recovery capability, and staff time spent on patching, hardware, and capacity planning rather than product work. People time is usually the largest and the least visible, because it sits in a salary budget, not an infrastructure one.
What cloud costs surprise people after migrating?
Egress fees (data out to the internet at roughly $0.09/GB), the premium on managed services versus raw VMs, cross-zone and NAT traffic charges, backup storage that nobody lifecycle-manages, and quiet growth: it’s so easy to create resources that spend drifts up 5 to 10% a quarter without governance. A migration also carries a real one-time cost, typically 10 to 30% of first-year run cost.
Should we keep anything on-premise?
Sometimes. Common keeps: latency-critical systems tied to physical operations (manufacturing, lab equipment), workloads with data-residency constraints your provider can’t satisfy, very large stable storage with heavy egress, and licensed software priced punitively in cloud. A hybrid posture with those exceptions, and everything else in cloud, is a normal end state, not a failure.
Want this comparison run on your actual estate instead of a representative one? Talk to a Webisoft cloud engineer. We’ll build the 5-year TCO model from your inventory and your bills, and if the answer is “stay on-prem” or “move half,” that’s the answer you’ll get.
Frequently asked questions
Is the cloud actually cheaper than on-premise?
For most SMB and mid-market estates, yes, by 20 to 40% over 5 years, but only when the comparison is honest on both sides and the cloud side is run with discipline (rightsizing, commitments, off-hours shutdown). For stable, high-utilization workloads on hardware you already own, with admin staffing you'd keep anyway, on-prem frequently wins on pure dollars. The cloud's advantage is largest where utilization is low and people time is the hidden cost.
What on-premise costs do people forget in TCO comparisons?
The big five: hardware refresh (servers and storage age out every 4 to 6 years), power and cooling, software licensing (virtualization especially, after recent price increases), a real disaster-recovery capability, and staff time spent on patching, hardware, and capacity planning rather than product work. People time is usually the largest and the least visible, because it sits in a salary budget, not an infrastructure one.
What cloud costs surprise people after migrating?
Egress fees (data out to the internet at roughly $0.09/GB), the premium on managed services versus raw VMs, cross-zone and NAT traffic charges, backup storage that nobody lifecycle-manages, and quiet growth: it's so easy to create resources that spend drifts up 5 to 10% a quarter without governance. A migration also carries a real one-time cost, typically 10 to 30% of first-year run cost.
Should we keep anything on-premise?
Sometimes. Common keeps: latency-critical systems tied to physical operations (manufacturing, lab equipment), workloads with data-residency constraints your provider can't satisfy, very large stable storage with heavy egress, and licensed software priced punitively in cloud. A hybrid posture with those exceptions, and everything else in cloud, is a normal end state, not a failure.