When is on-prem actually cheaper than the cloud?
A question that keeps coming back in ops forums: when does an on-premises server actually become cheaper than AWS, Azure, or even a plain VPS? In one recent thread the most-upvoted reply — a couple hundred votes — was worth pinning:
That’s the answer and the catch in two sentences. Let me take both seriously, because the honest version is more useful than either “cloud is a scam” or “nobody ever got fired for picking AWS.”
The math: when owning (or renting flat) wins
Cloud pricing is a meter. You pay by the hour, by the gigabyte moved, by the request. That’s brilliant when your usage is unpredictable — and punishing when it’s steady, because you’re paying a premium for flexibility you’re no longer using.
The clearest example is egress, the fee to move your own data out. On AWS the headline rate has sat around $0.09/GB since 2018. A workload pushing 50 TB a month runs roughly $53,000 a year in transfer fees alone — before storage, before compute. (More on finding that line on your own bill: the meter you can’t see.) A flat-rate dedicated server from a provider like Hetzner or OVH rents for a fixed monthly price — on the order of tens to a couple hundred dollars — with bandwidth included. That box costs the same whether you serve 1 TB or 50.
So owned or flat-rate hardware tends to win when three things are true at once:
- Steady — it runs most of the day, most of the year (not a spike you scale up for once a quarter).
- Meter-heavy — its bill scales with your success: egress, always-on compute, per-request fees.
- Long-lived — it’ll be around for more than a year, so the setup cost amortizes.
Hit all three and the savings aren’t a rounding error — they’re typically the kind of 40–70% cut that shows up on the invoice every single month, and grows as you grow.
When the cloud genuinely wins (and I’ll say so)
The same honesty cuts the other way. Keep it on the meter when:
- Load is spiky or seasonal. If traffic triples for one week a year and idles the rest, the meter is your friend — moving to fixed hardware means paying for the peak all year long.
- You need to scale to zero. A workload that’s busy some days and dark others is exactly what per-second billing is good at.
- The product is brand new. Before you know your real usage shape, flexibility is worth more than the last 30% of savings.
- The workload is tiny. If a $5 VPS covers it, that’s already cheaper than an hour of anyone’s time spent optimizing. Leave it alone.
If your setup looks like that list, moving it off the cloud won’t pay off — and I’ll tell you that rather than sell you a migration.
That box costs the same whether you serve 1 TB or 50.
The real objection isn’t the hardware — it’s “who runs it at 2am”
Notice what the top-voted answer conceded on cost, then charged back on: you move to the cloud to manage less stuff. That’s the true objection, and a reply in the same thread put a number on it — you’ll pay staff time to rack it, patch it, and carry the pager, and you give up the redundancy a single colo or closet can’t match.
That’s all real. It’s also the single reason most small businesses stay on a bill they already know is too high. But look closely at the objection: it’s about operations, not location. It’s “who runs it,” not “where does it live."
"On-prem vs cloud” is the wrong frame
The real axis was never on-prem versus cloud. It’s metered versus flat — and it has four corners, not two:
| You run it | Someone else runs it | |
|---|---|---|
| Metered | (rare) | Hyperscaler cloud — the expensive default |
| Flat-rate | DIY server / colo | Managed flat-rate hardware |
Most people argue about the two diagonal corners — expensive-but-hands-off cloud versus cheap-but-you-run-it DIY — as if those were the only options. The corner that actually solves the SMB’s problem is the fourth one: flat-rate hardware that someone else operates. You get the cheaper, predictable bill without hiring a sysadmin or babysitting a box. That’s what “leaving the cloud” really means for a small business — you’re leaving the meter, not the datacenter.
The decision, in three questions a CFO can answer
- Is the workload steady and long-lived — running most of the day, most of the year? If no, keep it on the meter.
- Does its bill scale with success — egress, per-request, always-on compute? If no, the savings are small; don’t bother.
- If yes to both, the only question left is who operates it — and that’s a priced, solvable decision, not a reason to keep overpaying every month.
Want to run those three questions across your own stack? I put them on a one-page Leave the meter worksheet — the test, how to spot the meter on your bill, and a fill-in table to score each workload move-or-keep. Free, no strings.
If your cloud bill is riding a meter for a workload that never actually changes, that gap is exactly what I cost out — flat-rate hardware, run for you, versus what you’re paying now. Every message comes straight to me; I read and reply to each one myself, usually within a business day. Send me your setup and I’ll send back a one-page teardown showing which lines would move, what they’d cost flat, and — if it wouldn’t pay off — I’ll tell you that too.