Articles

Traffic goes up ten percent. The invoice goes up thirty. Ask any platform engineer who's sat through a Finance review in the last two years and you'll get the same tired nod. Public cloud pricing was sold as elastic — pay for what you use, nothing more. In practice, a lot of teams are discovering the meter runs faster than the workload it's supposedly measuring. This piece looks at why that gap keeps widening, what's actually driving it under the hood, and what architects are doing instead of just raising the credit limit.

The Bill That Doesn't Match the Traffic Graph

Here's the uncomfortable part: nobody budgeted for this. Cost models built around AWS, Azure, or GCP list prices assumed steady, predictable unit economics. Then egress fees crept in. Then reserved-instance discounts got restructured. Then a service that used to be "included" started metering API calls per thousand. None of it shows up as one dramatic price hike — it's death by a thousand line items on a Cost Explorer dashboard nobody reads until the CFO does.

For baseline workloads running continuously at 70–80% utilization — the stuff that never sleeps, never bursts, just hums along processing orders or serving a database — egress fees and idle-capacity premiums often stack up fast, prompting a shift toward a managed dedicated private cloud with fixed monthly resource limits instead of floating consumption pricing. Why? Because a 24/7 node doesn't care about elasticity. It needs the same sixteen cores at 3 a.m. as it does at 3 p.m. Paying a premium for flexibility you never use is just paying a premium.

Sound familiar? If your traffic graph is a flat line and your bill graph is a staircase, you've found the mismatch this article is about.

What's Actually Happening on the Market

The industry didn't wake up one day and decide public cloud was too expensive. It's been a slow accumulation of receipts, and a handful of very public exits made the conversation impossible to ignore.

The exits everyone quotes

• 37signals (the company behind Basecamp and HEY) pulled its infrastructure off AWS in 2022–2023 and published the math: roughly $3.2 million saved over five years by buying its own hardware and colocating it instead.
• Dropbox famously did this back in 2016, moving the bulk of its storage off S3 onto custom-built infrastructure — a move that still gets cited at every FinOps conference eight years later.
• Ahrefs, the SEO analytics platform, runs almost entirely on owned bare metal for a similar reason: their workload is predictable, storage-heavy, and doesn't benefit from cloud elasticity at all.

None of these were small startups panicking over a bill. They were mature engineering organizations doing the arithmetic and concluding that "cloud-native" and "cloud-hosted" are not the same commitment.

What's being tested right now

• VMware Cloud Foundation (VCF) deployments are seeing renewed interest following Broadcom's 2024 licensing overhaul, which pushed a lot of enterprises to re-evaluate whether staying on hyperscaler VMs was actually cheaper than running VCF on owned or colocated hardware.
• Zero-egress object storage — Cloudflare R2, Backblaze B2, Wasabi — is being piloted specifically to escape the data-transfer line item, which for media-heavy or analytics-heavy companies can dwarf compute costs entirely.
• Kubecost and OpenCost, the CNCF sandbox tooling, are showing up in more Kubernetes stacks not because teams love dashboards, but because nobody could answer "which namespace is burning the budget" without one.
• Spot and preemptible instance orchestration is being tested more aggressively for stateless workloads, while stateful, always-on services get carved out and moved to fixed-cost infrastructure entirely.

At KubeCon and re:Invent alike, the sessions filling rooms these days aren't about scaling up — they're about scaling down intelligently. FinOps Foundation membership has grown steadily since 2021, which tells its own story: cost governance became a discipline because the old assumptions stopped holding.

Why the Math Breaks Down

There's a structural reason this happens, and it's worth naming plainly instead of hand-waving at "cloud is expensive."

• Egress is the trap door. Compute is competitive. Storage is competitive. Moving data out of a hyperscaler is where margins hide, and it's the one cost engineers routinely underestimate at design time.
• Reserved capacity assumes forecasting accuracy nobody has. Commit to a 3-year reserved instance and guess wrong on growth, and you're either overpaying for idle capacity or scrambling on-demand at list price.
• "Pay for what you use" penalizes steady-state workloads. Elasticity pricing is optimized for spiky, unpredictable traffic. A workload that's boringly consistent — think a core banking ledger or an ERP backend — gets none of the benefit and all of the premium baked into the pricing model.
• Multi-service sprawl compounds silently. A managed database here, a message queue there, a serverless function chain in between — each is individually cheap and collectively brutal once volume scales.
• Support and compliance tiers add fixed overhead. Enterprise support contracts, security add-ons, and compliance tooling (SOC 2 evidence collection, for instance) get priced as percentages of spend, so the bill grows even when nothing architecturally changed.

Add it up and you get exactly the pattern engineers keep flagging: infrastructure spend curves upward on a workload that's flat. Not because anyone did anything wrong. Because the pricing model was never built for that shape of demand in the first place.

How Architects Are Responding

This isn't a "leave the cloud entirely" story — that would be its own kind of oversimplification. It's a segmentation story. Teams are increasingly splitting workloads by behavior, not by habit.

The split that's becoming standard

• Bursty, unpredictable, customer-facing traffic — front-end services, seasonal spikes, marketing campaign landing pages — stays on public cloud, where elasticity is genuinely earning its premium.
• Steady-state, always-on, predictable workloads — core databases, ERP systems, internal platforms, batch processing that runs on a fixed schedule — moves to dedicated or colocated infrastructure with a flat monthly cost.
• Regulated or latency-sensitive data — financial transactions, healthcare records, anything with residency requirements — gets placed wherever compliance actually demands it, which increasingly means infrastructure the organization controls directly rather than a shared multi-tenant region.

This is roughly the thinking behind offerings like DXC's Private Cloud+ Dedicated, which packages single-tenant infrastructure — hosted in a customer's own data center, a DXC facility, or a colocation site — with fixed resource limits and full lifecycle management, including patching, backup, and Day 2 operations, so the "steady-state" half of the workload split doesn't require building an internal ops team from scratch just to get out from under variable pricing. DXC reports customers seeing up to 30% lower costs within twelve months of that kind of migration, and it's not because the hardware is magic — it's because fixed capacity, priced once, stops compounding the way metered consumption does.

Questions worth asking before the next renewal

1. Which workloads have run at consistent utilization for six months or more?
2. What percentage of the monthly bill is data transfer, not compute or storage?
3. Would a 3-year fixed-cost commitment beat the current on-demand or reserved-instance blend?
4. Does the compliance or data-residency requirement actually mandate hyperscaler infrastructure, or is that just where it happened to land historically?
5. Is the engineering team spending more time optimizing cost allocation tags than shipping features?

If the answer to that last one is yes, that's usually the tell. Cost governance shouldn't be a full-time job layered on top of the actual work.

Where This Settles

Nobody's predicting a mass exodus back to on-prem racks — that ship, mostly, has sailed, and rightly so for workloads that genuinely need elasticity. But the blanket assumption that "cloud is always cheaper" has taken enough public hits — Broadcom's VMware pricing shake-up, the steady drumbeat of exit case studies, the FinOps movement itself — that it's no longer treated as gospel in architecture reviews. It's treated as one option among several, evaluated the way any infrastructure decision should be: against the actual shape of the workload, not against a marketing slide from 2015.

Flat traffic deserves flat pricing. That's not a controversial statement. It's just taken the industry a decade of surprise invoices to start acting on it.