Barclays’ CIO survey put a number on the mood: 83% of enterprise CIOs plan to repatriate at least some workloads. That figure launched a hundred “the cloud is over” headlines. Here is the number the headlines skip — only about 8% of companies move entire workloads back, and Flexera’s State of the Cloud data suggests roughly a fifth of migrated workloads have returned on-prem in some form. Repatriation is real. It is also selective, and the interesting question is not whether to leave the cloud — it’s which workloads clear the bar.
This brief gives you the actual math: the two workload profiles that consistently win on-prem, the thresholds that decide it, and where AWS, Google Cloud, Broadcom/VMware, and Red Hat sit as the traffic starts to move.
What changed
Three forces converged between 2024 and 2026. First, the FinOps easy wins ran out — after rightsizing, reserved capacity, and spot, steady-state cloud bills stopped shrinking, and finance started asking why a workload with flat utilization needs elastic pricing. Second, GPU economics got loud: with on-demand GPU pricing running roughly $2–8 per hour as of mid-2026, an owned inference cluster starts beating rented capacity once sustained utilization crosses the 60–70% range. Third, the exit tax fell. Under pressure from the EU Data Act, both Google Cloud and AWS now waive egress fees for customers moving data off their platforms — which removed the single largest one-time cost of leaving.
The takeaway for the meeting: the barriers to selective repatriation dropped at the same time the savings case matured. That’s why the survey numbers spiked — not because the cloud stopped working.
The numbers, minus the hype
Most repatriation coverage is either hyperscaler damage control or hosting-vendor cheerleading. Here is what the data actually supports, with the hype version alongside it.
| The headline version | What the data says | |
|---|---|---|
| CIO intent | “86% of CIOs are leaving the cloud” | Barclays found 83% plan to repatriate some workloads — intent, not exodus |
| Full exits | “Enterprises are abandoning AWS” | Only ~8% of companies move entire workloads; the rest rebalance |
| Volume moved | “The great cloud reversal” | Flexera data points to roughly one-fifth of migrated workloads coming back in some form |
| Savings | “Cut your bill 80% overnight” | Steady-state workloads average ~32% savings over a 3–5 year horizon — real, not miraculous |
| GPU math | “Never rent a GPU again” | Owned clusters win above ~60–70% sustained utilization at $2–8/hr on-demand rates; below that, keep renting |
Read the table as a diagnostic. If your estate is mostly variable, the honest savings number for you is near zero. If 40% of your compute runs flat around the clock, the 32% figure is worth a business case.
The workloads that actually leave
Two profiles dominate real repatriation projects, and they share one trait: predictability.
Profile 1: steady-state compute and storage
Databases with stable IOPS, internal line-of-business applications, batch processing, and archive-heavy storage. The pattern that qualifies: utilization that varies less than roughly ±20% month over month, a demand forecast you trust for three years, and a 3–5 year amortization window. These workloads pay cloud’s elasticity premium and never use the elasticity. Moved to owned or hosted infrastructure, they are the source of that ~32% average saving — sometimes more when storage egress was a recurring line item, a pattern we quantified in our egress fee benchmark.
Profile 2: sustained-utilization GPU inference
Training bursts belong in the cloud. Inference that hums along at high utilization does not. At $2–8 per GPU-hour on demand, a cluster you keep above 60–70% utilization typically pays for itself well inside its useful life — as of mid-2026 hardware pricing, commonly within two years. The break-even is sensitive to power costs and your ability to actually keep the cluster busy, so run the model before you buy; our on-prem GPU cluster cost guide walks through the full TCO stack, including the parts vendors leave out.
The workloads that stay
Be equally honest about what doesn’t move, because failed repatriations are expensive twice.
- Spiky and seasonal demand. Retail peaks, tax season, campaign traffic — elasticity is what you’re paying for, and it’s worth it.
- Anything welded to higher-order managed services. If the workload leans on BigQuery, DynamoDB, or a managed streaming stack, the replatforming labor usually eats the infrastructure savings. Count the engineering hours before you count the server savings.
- Globally distributed, latency-sensitive front ends. Replicating a hyperscaler’s edge footprint on your own is not a savings project.
- Anything owned by a team with no ops bench. Repatriation converts an invoice into an operational responsibility. If you don’t have the staff, you don’t have the option.
The verdict a VP can repeat: repatriate the base load, rent the spike.
How AWS and Google Cloud are playing defense
AWS’s response has been pragmatic rather than panicked. It waives egress for customers moving off the platform — conditional, and coordinated through support rather than automatic — and it is leaning on the two levers that keep steady-state workloads in place: deeper commitment discounts and hybrid hardware. Outposts is effectively AWS’s “repatriate without leaving” play, and its European Sovereign Cloud targets the compliance-driven slice of repatriation demand. AWS’s real strength is gravity — the breadth of managed services makes full exits genuinely painful. Its weakness is the flip side: on raw, predictable compute, list pricing loses to owned hardware, and AWS knows it.
Google Cloud moved first on exit fees, announcing free egress for departing customers back in January 2024 — but read the terms: the waiver applies when you close the account and move everything within a 60-day window, which is stricter than the press release implied. Google’s retention story is stickiness of the productive kind: BigQuery and Vertex AI are hard to replicate on-prem, and GKE Enterprise gives it a credible hybrid answer for Kubernetes estates. Where Google is thinner is the on-prem hardware story — there is no Outposts equivalent with real traction, so its repatriation-prone customers tend to leave rather than hybridize.
The landing zone problem: Broadcom/VMware vs Red Hat
Here’s the twist most repatriation math ignores: the default landing zone got expensive at exactly the wrong moment. Broadcom’s overhaul of VMware — subscription-only licensing, the product catalog consolidated down to a handful of bundles around VMware Cloud Foundation — has produced widely reported renewal increases in the 2–5x range for estates pushed from bare vSphere into the full VCF stack, with some customers reporting far worse. The April 2025 announcement of a 72-core license minimum was reversed back to 16 cores after customer pushback, but the episode did lasting damage to renewal psychology. If you repatriate onto VMware without negotiating hard, you can hand your cloud savings straight to Broadcom.
That said, dismiss VCF at your peril. It remains the most complete private-cloud stack on the market, your operations team already knows it, and the ecosystem — backup, DR, monitoring — is unmatched in depth. It fits large estates that consolidate onto dense hosts (fewer phantom cores under the minimums) and treat the renewal as a procurement event, not a formality.
Red Hat is the beneficiary of the moment. OpenShift Virtualization runs VMs on KVM alongside containers on one platform, the migration toolkit for virtualization automates vSphere imports at scale, and OpenShift Virtualization Engine is the VM-only edition aimed squarely at the “I just want a hypervisor without the drama” buyer. Its strengths: one platform for VMs and containers, and no hypervisor renewal shock. Its honest weaknesses: the operational model is Kubernetes-first — vSphere admins face a real learning curve — and the third-party tooling ecosystem around backup and DR, while growing fast, is younger than VMware’s. Fit: organizations already invested in OpenShift or containers, and mid-size estates willing to retrain.
The move: get both quotes. Even if you intend to stay on VMware, a credible OpenShift Virtualization quote is the best negotiating instrument Broadcom has ever given its customers.
What to do about it
- Inventory by utilization curve, not application age. Flat curves are repatriation candidates; volatile curves are not. This one query reframes the whole conversation.
- Price the full exit. Egress waivers help, but the real costs are replatforming labor and landing-zone licensing. A business case without a VMware or Red Hat quote attached is not a business case.
- Set thresholds and hold them. Utilization variance under ±20%, a three-year forecast you believe, GPU utilization above 60–70% — below the bar, stay in the cloud without guilt.
- Rebalance, don’t exit. The 8% doing full exits are mostly special cases. The defensible posture for 2026 is cloud for the spike, owned infrastructure for the base.
- Re-run the math annually. Hardware prices, GPU rates, and hyperscaler discounts all moved in the last eighteen months. A 2024 analysis is already stale.
Frequently asked questions
What is cloud repatriation?
Cloud repatriation is moving workloads from public cloud providers back to on-premises data centers, colocation, or private cloud. In practice it is almost always selective — specific workloads move while the rest of the estate stays in the cloud.
Are companies really leaving the cloud in 2026?
Mostly no — they are rebalancing. Barclays found 83% of CIOs plan to repatriate some workloads, but only around 8% of companies move entire workloads off the cloud. The dominant pattern is moving steady-state workloads back while keeping variable ones in the cloud.
Which workloads should move back on-premises?
Workloads with flat, predictable utilization over a 3–5 year horizon — stable databases, internal apps, batch, and archives — plus GPU inference running above roughly 60–70% sustained utilization. Spiky demand and workloads built on managed services should generally stay.
Does AWS charge egress fees when you leave?
As of mid-2026, AWS waives data transfer fees for customers migrating off the platform, following Google Cloud’s January 2024 move. The waivers are conditional — AWS coordinates them through support, and Google’s requires a full account closure within a 60-day window — so plan the exit before you start it.
Enterprise Techie publishes vendor-honest analysis like this daily — get the brief by email, free.