Colocation vs Cloud: The Real TCO Comparison

For steady-state, predictable workloads, colocation typically costs 40-60% less than public cloud over a two-to-three-year horizon once egress fees, support-tier charges, and reserved-instance premiums are priced into the comparison. Cloud remains cheaper for bursty, short-lived, or unpredictable workloads, where colocation’s fixed capex has no chance to amortize. The right answer depends entirely on utilization, not on which model has the lower headline rate.

Key takeaways

• Breakeven lands between 6 and 18 months. Buying hardware and colocating it typically pays back against renting equivalent cloud VMs in that window, faster for steady, high-utilization workloads.
• The $30K/month threshold. Teams spending above roughly $30,000 monthly on stable, predictable cloud workloads are the clearest colocation-migration candidates, with 40-60% savings documented over 2-3 years.
• Egress is not a rounding error. AWS charges $0.09/GB, Azure $0.087/GB, and GCP $0.12/GB for the first tier of internet egress — a 50 TB/month workload adds roughly $4,400-4,500/month on AWS alone.
• Reserved pricing narrows but doesn’t close the gap. 1-3 year Reserved Instances or Savings Plans cut cloud costs 37-72%, but documented case studies still show colocation running a fraction of the equivalent on-demand cloud cost for constant, high-density loads.
• GPU economics favor ownership at high utilization. An amortized, colocated 8x H100 server lands near $1.60-1.70/GPU-hour — below both the ~$3.72/GPU-hour on-demand median and the ~$2.35/GPU-hour one-year cloud contract rate, per our GPU price tracker.
• Hidden costs cut both ways. Cloud hides egress, IOPS, and support tiers; colocation hides hardware refresh cycles, spare parts, and the engineering headcount needed to replace managed services.

Benchmark current wholesale and retail colocation rates in our colocation price index and see the full facility catalog in our data center directory.

Why the headline price is the wrong comparison

Comparing a cloud instance’s hourly rate to a server’s monthly colocation invoice is the single most common TCO error. The two prices are not measuring the same thing:

• The cloud rate bundles compute, a share of the provider’s capex, power, cooling, networking, and a management layer — priced for elasticity, available by the second, cancellable anytime.
• The colocation rate covers only rack space, power, cooling, and physical security. The server itself, its refresh cycle, the OS and application stack, and most of the operational labor are the buyer’s responsibility.

A fair TCO model normalizes both sides to total dollars per unit of compute per month, over the same time horizon, and includes every recurring line item on each side — not just the number printed on the rate card.

What belongs in a cloud TCO model

Cost component	Typical range (2026)	Notes
On-demand compute	Instance-family dependent	Baseline; rarely what steady-state workloads actually pay
Reserved Instances / Savings Plans (1-3 yr)	37-72% below on-demand	Requires committing to instance family/region for the term
Internet egress (AWS)	$0.09/GB (first 10 TB), down to $0.05/GB above 150 TB	First 100 GB/month free
Internet egress (Azure)	$0.087/GB below 50 TB	Converges with AWS above 50 TB
Internet egress (GCP Premium Tier)	$0.12/GB (first 1 TB), $0.08/GB above 10 TB	Standard Tier flat at ~$0.085/GB
Cross-AZ / inter-region transfer	$0.01-0.02/GB each direction	Easy to miss in multi-AZ architectures
Storage IOPS, snapshots, support tiers	Workload-dependent	Enterprise support commonly 3-10% of monthly spend

Source: EgressCost.com pricing comparison, nOps EC2 pricing breakdown.

Egress is the line item modelers skip most often. At $0.09/GB, a workload that streams 50 TB out of AWS in a month adds roughly $4,400-4,500 to the bill before any compute is counted — and GPU inference and training pipelines routinely move multiples of that, per Spheron’s 2026 GPU bandwidth analysis.

What belongs in a colocation TCO model

Cost component	Typical range (2026)	Notes
Hardware capex	Server-class dependent	Depreciate over the refresh cycle, not paid monthly
Colocation fee (retail rack)	$500-1,500/month	Full 42U rack, 5-10 kW; see our colocation pricing guide
Colocation fee (wholesale, $/kW)	$130-215+/kW/month, market-dependent	Primary US markets vs. secondary markets
Hardware refresh cycle	4-6 years	Budget a full hardware replacement, not just repairs
Remote hands / labor	$100-250/hour beyond included allowance	Or in-house ops headcount
Cross-connects, IP transit	$50-350/month per connection; $0.10-1.00/Mbps	No egress fee for outbound bandwidth itself

Colocation has no equivalent to cloud egress pricing — bandwidth is contracted separately (IP transit) rather than billed per gigabyte on outbound traffic, which is exactly why bandwidth-heavy, steady workloads see the largest swing in favor of colocation.

Worked example: a steady, high-density workload

The clearest real-world data point comes from a documented SEO/analytics firm comparison cited by Network World: running the same high-density compute capacity cost $17,557/month on AWS versus $1,550/month colocated — an 11x gap for a single equivalent server. Scaled to 850 servers, the same comparison ran $39.5 million over 30 months colocated versus $447.7 million over 30 months on equivalent AWS EC2 instances.

Scenario	Monthly cost (per equivalent server)	30-month total
AWS EC2 on-demand	$17,557	~$527K
Colocated owned hardware	$1,550	~$46.5K

This is an extreme case — high-density, constant-load, hardware-specific compute is where cloud pricing is least competitive. Reserved Instances or Savings Plans would narrow the gap by 37-72%, per nOps, but even at the top of that discount range the on-demand comparison shown above would still land several multiples above colocation. The lesson generalizes: the more predictable and hardware-specific the workload, the wider the colocation advantage.

For typical enterprise workloads — application servers, databases, steady-state AI inference — industry TCO audits converge on a more moderate but still material 40-60% cost reduction over 2-3 years when moving from public cloud to colocation, once egress and hidden fees are modeled, per Databank and 3EX Hosting.

The GPU case: buy-and-colocate vs. rent

AI compute makes the buy-vs-rent decision sharper because the hardware capex is large, well-documented, and easy to amortize against a known power draw.

An 8-GPU HGX H100 server costs $250,000-320,000 fully configured, per Mercatus, and draws up to 10.2 kW at full load per NVIDIA’s DGX SuperPod electrical specification. Amortizing a $280,000 mid-range configuration over 3 years and colocating it at a representative $170/kW/month wholesale rate:

Line item	Monthly cost	Per-GPU-hour (8 GPUs, 730 hrs)
Hardware amortization ($280K / 36 months)	$7,778	$1.33
Power + space (10.2 kW @ $170/kW)	$1,734	$0.30
Total owned + colocated	$9,512	~$1.63

Compare that to current GPU price tracker benchmarks: H100 on-demand cloud rental runs a median of $3.72/GPU-hour (range $1.49-11.06), and the best published one-year contract rate sits near $2.35/GPU-hour. Owned-and-colocated hardware undercuts both — but only if utilization holds near 100% for the full 3-year term. A GPU sitting idle 40% of the time effectively pays $2.72/GPU-hour for used capacity, erasing most of the advantage. See our H100 rental price guide for current provider-by-provider rates.

When cloud still wins

• Anything shorter than the breakeven window. If the workload’s committed life is under 6-18 months, hardware capex will not amortize before the project ends.
• Genuinely bursty or seasonal load. Retail traffic spikes, batch jobs, and R&D experimentation benefit from paying only for consumed compute.
• Workloads needing managed services. Serverless functions, managed databases, and global CDN/edge delivery would require significant engineering investment to replicate on owned infrastructure — factor that opportunity cost into any TCO model.
• Early-stage products. Before product-market fit, the flexibility to resize or kill infrastructure outweighs a cost delta that may never materialize at scale.

Capex vs. opex: the accounting dimension

TCO comparisons often stop at the cash number and miss how each model hits the income statement. Cloud spend is opex — it flows through the P&L in the period incurred, with no balance-sheet asset and no depreciation schedule to manage. Colocation splits into two pieces: the colocation fee itself is opex (rent for space, power, and cooling), while the hardware is capex, depreciated over its useful life (commonly 3-5 years for servers) and carried on the balance sheet until disposal or refresh.

For a finance team, this matters in three ways. First, capex requires upfront cash or financing — the $250,000-320,000 GPU server example above is paid (or financed) at purchase, not spread evenly like a cloud bill, which changes near-term cash flow even when the 3-year TCO favors colocation. Second, depreciation creates a tax shield that pure opex does not, which shifts the after-tax comparison somewhat in colocation’s favor for capital-rich buyers. Third, colocation hardware is a disposable or resellable asset at end of life, while cloud spend has zero residual value — a difference TCO models built purely on monthly cash cost tend to ignore. None of this changes which model is cheaper on a pre-tax cash basis, but it does change which model a specific balance sheet can absorb.

What to do

1. Pull 90 days of actual cloud billing, not list prices — including egress, cross-AZ transfer, storage, and support-tier line items, which routinely add 15-30% beyond compute.
2. Segment workloads by utilization pattern. Steady-state, hardware-specific, or bandwidth-heavy workloads are colocation candidates; bursty and experimental workloads should stay in the cloud.
3. Model 3-year TCO on both sides, including hardware refresh cycles and remote-hands labor for colocation, and realistic Reserved Instance/Savings Plan discounting for cloud — not on-demand list price on either side.
4. Benchmark colocation rates against the current market using our colocation price index and market statistics before assuming a quoted rate is competitive.
5. Get a colocation quote to compare directly against your current cloud invoice through our quote service, and review facility options in the relevant sector guide.