We are not seeing a mass “back to on-prem.” Cloud spend keeps compounding. What we are seeing is targeted repatriation of specific workloads where the economics are lopsided mostly high, steady utilization and storage/egress-heavy use cases. The rational end-state is hybrid.
The signal (not the noise)
- Cloud adoption is still growing at double-digit rates globally. That’s not a market in retreat.
- Repatriation is real but selective. Companies pull parts of workloads back (e.g., storage tiers, steady compute) while keeping elastic or managed services in cloud.
- Hybrid (cloud + on-prem) is the operating model that keeps winning.
When on-prem legitimately beats cloud
If you can say “yes” to most of these, on-prem probably wins on cost:
- High, steady utilization (think 24/7 services) you can keep hot for years.
- Storage-heavy + high egress patterns (media, analytics, data platforms pushing lots of bytes out).
- Predictable growth and the ability to buy, rack, power, and automate efficiently.
- AI at scale with tightly scheduled clusters you keep busy (training).
Why: You amortize hardware and pay relatively flat power/space/ops. If you keep the metal busy, your unit cost falls below cloud’s per-hour pricing.
When cloud still wins (even on cost)
- Seasonal demand where paying for idle iron would kill you.
- Managed services that replace headcount and operational risk.
- Speed and optionality new regions, services, and partnerships you can’t replicate easily on-prem.
The cost reality (no BS)
Using typical, conservative numbers:
- Cloud on-demand pricing vs on-prem: on-prem becomes cheaper only when average utilization gets to about ~50%+.
- If you’ve already negotiated 1-year or 3-year commitments in cloud, the break-even shifts up: ~80%+ utilization before on-prem pulls ahead.
These are directional, not absolute. Your actual break-even depends on your prices, power, headcount, amortization period, and workload shape. But the pattern is stable: on-prem only wins if you keep it hot.
A simple, ruthless decision framework
- Segment your estate by utilization, storage/egress ratio, compliance, and ops overhead.
- Optimize cloud first (rightsizing, commitments, storage class, architecture). If the unit economics are still ugly, proceed.
- Pilot one repatriation candidate (steady, high-egress service) into colo (colocation data center) with full-loaded costs and SLOs.
- Standardize: document the playbook and repeat where it clearly beats cloud. Ignore everything else.
Bottom line
The smart move isn’t ideology, it’s hybrid pragmatism. Keep elastic and high-velocity innovation in cloud. Pull back the steady, storage-intensive, and predictably busy workloads where the math is unambiguous.