In October 2024, DHH posted a number that should have ended the colocation vs cloud debate: "Our cloud-exit savings will now top ten million over five years." 37signals had been paying AWS $3.2 million a year. They bought $700K worth of hardware, moved to colocation, and dropped to $1.3 million. The hardware paid for itself inside the first year.

Six years before that, Dropbox filed its S-1 and quietly disclosed that building its own infrastructure — a project called Magic Pocket — saved $74.6 million over two years compared to staying on AWS S3.

These numbers are public. They've been public for years. And yet every article on "colocation vs cloud hosting" in 2026 still ends with some version of "it depends on your needs."

It does depend. But not on your needs — on two numbers you can calculate in about ten minutes: your monthly cloud spend and the percentage of your workloads that are stable. I've helped clients run this math for six years now, and the answer is almost never ambiguous once you actually do it.

So let's do it. Every dollar amount in this article is from April 2026 published pricing — not "approximate" ranges from a blog post written in 2019.

What You're Actually Comparing (It's Not Features)

Every colocation vs cloud article I've read this year frames the comparison the same way: control vs flexibility. Customization vs scalability. Security vs convenience. These are real differences, but they're the wrong lens for making a decision.

The actual comparison is between two cost structures:

Cloud is pure OpEx. You pay monthly, forever, and the meter never stops. AWS charges you $30.37/month for a t3.medium (2 vCPU, 4GB RAM) — but that's before egress, before cross-AZ transfers, before NAT gateway fees, before snapshot storage, and before the reserved instance you bought for a workload you deprecated in March.

Colocation is CapEx up front plus lower OpEx. You buy hardware ($3,000–8,000 for a decent server), rent rack space ($75–150/month for 1U), and pay for power, connectivity, and the occasional remote hands intervention. The monthly cost is lower, but you're on the hook for hardware refreshes every 4–5 years, and every cross-connect and support call has a price tag.

The question isn't "which gives me more control." The question is: over 36 months, which one costs less — and by how much?

When you frame it that way, the answer stops depending on philosophy and starts depending on arithmetic.

The Cloud Bill Autopsy

Cloud pricing looks simple on the pricing page. It is not simple on the invoice.

I've reviewed cloud bills for about two dozen clients at this point, and I've never seen one where the actual monthly cost was less than 1.3x the sticker price. The worst was 2.4x. The culprits are always the same.

Egress fees. AWS charges $0.09 per GB for data leaving their network. Run a mid-traffic SaaS app pushing 500GB of egress per month and that's $45/month just to send data to your own users. Scale to 5TB and it's $450/month — for bandwidth that costs your colo provider a fraction of that.

Cross-AZ transfer. AWS charges $0.01/GB in each direction for traffic between availability zones. Best practice says deploy across multiple AZs for redundancy. AWS charges you for following their own best practice. A microservices architecture chattering across three AZs can rack up hundreds of dollars monthly in transfer fees that never appear in any pricing calculator.

NAT Gateway. $0.045 per GB processed, on top of the egress charge. It's a tax on a tax. 1TB through a NAT gateway costs $45 in processing fees before the $90 egress charge. You're paying $135 to move data that costs the provider almost nothing.

Snapshot and storage creep. EBS snapshots accumulate quietly. That "temporary" snapshot from January is still there in April, and you're paying for every GB. I've seen clients with $200–400/month in forgotten snapshot costs.

One Hacker News user put it bluntly: "Got an unexpected $85K bill from Google Cloud... submitted support tickets weeks ago and yet to receive a response." That's an extreme case. But the pattern — bill shock from costs you didn't model — is not extreme at all. It's the norm.

The honest multiplier: take your cloud sticker price and multiply by 1.3–2.5x depending on your egress volume and architectural complexity. That's your real monthly cost. Any TCO comparison that uses the base VM price is comparing fiction.

The Colocation Sticker Shock

If you read the last section and thought "great, colocation is obviously cheaper" — slow down. Colo has its own hidden cost layer, and it's just as dishonest about the real number.

A 1U colocation slot advertises at $49–150/month. Hivelocity lists $49/month for 1U. Looks cheap. Here's what's not in that number:

As Brightlio documents in their pricing breakdown: "Remote hands can cost $50–150 per 15-minute intervention." Four incidents a year — and four is optimistic for aging hardware — adds $600–2,400 annually that never showed up in the quote.

The honest multiplier: take the advertised rack price and expect 2–3x once you add a single cross-connect and amortize hardware over 5 years. A $150/month 1U slot realistically costs $300–400/month ($150 rack + $100–200 cross-connect + $50–80 hardware amortization). A quarter-rack at $500/month is really $1,000–1,300/month.

Colo isn't magic. It's cheaper than cloud at scale, but only if you account for everything — and most colo quotes are designed to make you not account for everything.

The 36-Month Math: Three Scenarios

I built these scenarios using verified April 2026 pricing. Every number comes from published rates. Cloud costs use a 1.5x multiplier on base spend (conservative middle of the 1.3–2.5x range). Colo costs include all the hidden items from the table above.

Look at the $2,000/month tier. Colo saves $63,500 over three years — nearly $1,800/month. At the $5,000 tier, you're looking at $169,400 in savings, enough to hire a full-time DevOps engineer and still come out ahead.

But the $500/month tier tells a different story. Yes, colo is cheaper on paper — $268/month savings. But that savings assumes you have someone to manage the hardware. Someone to respond when the server goes down at 3 AM on a Saturday. Someone to plan the hardware refresh at month 48. I had a client at this tier who spent his first colo weekend driving 40 minutes to the data center because a firmware update bricked his RAID controller. He saved $268 that month and burned a Saturday. By month four he was back on a cloud VPS, and when I asked him if colo was worth it, he laughed. If that someone is you — the founder, the solo dev, the person who also does sales and support — the savings aren't savings. They're a pay cut.

As one Hacker News commenter observed: "People say cloud is cheaper, but that's based on horrible assumptions such as..." — and then listed a dozen costs people omit. The same criticism applies in reverse. Colo evangelists who only quote rack price are doing the same math, just in the other direction.

The crossover point: somewhere around $1,500–2,000/month in real cloud spend, colocation starts winning decisively — but only if more than 70% of your workloads are stable (not bursty, not seasonal, not experimental). Below $1,500/month or with highly variable workloads, cloud's flexibility premium is worth paying.

The 2026 Market Reality Nobody Mentions

Here's where the repatriation narrative hits a wall.

North American colocation vacancy hit 1.4% at the end of 2025 — a record low. A healthy commercial real estate market runs 5–8% vacancy. At 1.4%, you're not shopping for colocation. You're competing for it.

Average wholesale pricing reached $195.94 per kW per month for 250–500 kW deployments in H2 2025. That's a national average. In Northern Virginia — the largest data center market in the world — availability for high-density deployments is effectively zero. Silicon Valley is similar. Space exists in Atlanta, Dallas, Phoenix, but even secondary markets are tightening fast.

What this means for your 36-month plan:

• Lead times are longer. Getting a quarter-rack used to take 2–3 weeks. In 2026, plan for 4–8 weeks in Tier 1 markets — assuming space exists at all.
• Contracts are longer. Providers know demand exceeds supply. Expect 24–36 month minimums where 12 months used to be standard.
• Power is the real bottleneck. Even where rack space is technically available, power capacity may not be. AI and GPU workloads have consumed massive power allocations in major markets.
• Price escalation clauses are aggressive. Read the contract carefully. "Power rates subject to annual adjustment" is an open-ended cost increase baked into your 36-month commitment.

Remember that Barclays CIO Survey — 86% plan to move workloads back to private environments? That's 86% of CIOs competing for 1.4% vacancy. The math works. The execution is a different problem entirely.

If you're evaluating colocation in 2026, start the space search months before you order hardware. And if your target market is Northern Virginia, have a backup plan. The 36-month savings don't mean much if month one starts with "we can put you on a waitlist."

Verdict: The Decision Framework

I've spent this entire article arguing that "it depends" is a cop-out. So here's where I put my own framework on the table.

Do not colocate if:

• Your team is under 3 people with no dedicated ops person. The operational burden will consume more value than you save on hosting.
• You need presence in 15+ geographic regions. Colo gives you one or two locations. Cloud gives you the world.
• You're spending under $500/month on cloud. The savings are real but thin, and one bad remote hands incident wipes out a quarter's worth.

One thing I want to be direct about: I can't give you specific colo provider recommendations. Colocation pricing is negotiated, not listed. The numbers in this article are industry averages and published rates. Your actual quote will depend on power draw, contract length, market, and negotiation. What I can give you — and what the rest of the SERP won't — is the framework to evaluate that quote honestly, with all the hidden costs visible.

The Dropbox and 37signals stories are real. The savings are real. But those companies had teams of infrastructure engineers and budgets measured in millions. The question isn't whether colocation saves money at scale — it does. The question is whether your scale justifies the cost of running your own hardware, and whether 2026's market will let you execute on the plan at all.

Run the 36-month math with your actual numbers. If colo saves you $1,500+/month, take it seriously. If the savings are under $500/month, cloud's operational simplicity is worth the premium. The worst outcome isn't choosing wrong — it's not choosing at all, letting your cloud bill compound 30–50% above sticker price, month after month, because you read an article that told you "it depends."

One caveat worth stating explicitly: this entire framework assumes general-purpose compute — web apps, databases, APIs, file storage. If your workload is GPU-heavy (ML training, rendering, inference), the economics are completely different. GPU colo pricing starts at $4,000–8,000+/rack/month in 2026, availability is near zero in Tier 1 markets, and cloud GPU instances have their own pricing insanity. That comparison deserves its own article, not a paragraph bolted onto this one.

It depends on two numbers. Go calculate them.

FAQ

At what monthly spend does colocation start saving money over cloud?

The crossover is around $1,500–2,000/month in actual cloud spend (after hidden costs, not sticker price). Below that, colo's fixed costs — cross-connects at $100–300/month, remote hands at $150–250/hour when things break, hardware amortization — eat too far into the savings. Above $2,000/month with mostly stable workloads, the gap opens fast: $1,700+/month in savings at the $2K tier over 36 months.

Can I colocate a single server, or do I need a full rack?

Single-server 1U colocation exists. Hivelocity starts at $49/month, and the broader market runs $75–150/month for 1U. But the economics aren't great — one cross-connect at $200/month pushes your effective cost to $249 — over 5x the rack price alone before you touch remote hands or bandwidth. Single-server colo makes sense for specific compliance or latency requirements, not for cost optimization.

How long does colocation setup take from scratch?

Budget 3–8 weeks. Hardware procurement runs 1–3 weeks depending on config. Colo provisioning takes 1–2 weeks once the contract is signed. But in a 1.4% vacancy market, finding available space in your target geography could add weeks or months to the front end. Start the space search before you order hardware.

What happens to my servers if the data center goes bankrupt?

Your hardware is your property — they can't legally seize it. That's the good news. The bad news is everything else. Power gets cut. Network goes dark. And your "reasonable access" clause means nothing when the building has a padlock on the door and a receiver sorting through creditor claims. I watched a client lose 11 hours of uptime during a provider acquisition in 2022 — not a bankruptcy, just a change of ownership — because the handover team deprioritized single-rack tenants. Three things to lock down in your contract before you sign: an equipment removal clause with 30-day notice (not 90), a right to enter the facility within 48 hours of any ownership event, and a prohibition on prepaying more than 3 months. And keep current backups offsite — not at another facility from the same parent company.

Is hybrid cloud + colocation worth the complexity?

At the $2,000+/month level, yes — but keep it boring. Stable workloads (databases, application servers, file storage) go to colo. Bursty workloads (batch processing, seasonal spikes, dev/staging) stay in cloud. Connect them with a site-to-site VPN or a direct cross-connect to a cloud on-ramp. Don't attempt dynamic orchestration between the two unless you have a dedicated platform team. Sophisticated hybrid setups erase the cost savings for most organizations under 10 engineers.

Does colocation make compliance (HIPAA/PCI) easier?

It makes compliance more controllable, not easier. You own the hardware, control the encryption keys, and decide who physically accesses the machines — that simplifies data residency and physical access requirements. But it shifts the compliance burden onto you. In cloud, the provider handles facility certifications. In colo, you need to verify your facility's SOC 2/SSAE 18 status and potentially bring in your own compliance auditor. The tradeoff: more control, more responsibility.