ZDNet’s latest look at cloud repatriation opens with a familiar case study. 37signals, which had moved to the public cloud in 2006, began pulling workloads off AWS in 2022. By 2023 it had replaced its AWS and Google Cloud servers with company-owned hardware in colocation space. The company says a one-time hardware spend of $500,000 replaced more than $3.5 million in annual cloud bills.
The more telling details are further down. Platform9’s Sirish Raghuram told ZDNet his company moved about 40% of its workloads on-premises after cloud resource costs spiked in December 2025. A Cast AI benchmark of 2,100 organizations found Kubernetes clusters using an average of just 10% of provisioned CPU and 23% of memory. Flexera’s 2026 State of the Cloud report found that wasted cloud spend rose to 29%, the first increase after five years of decline, driven in part by the cost complexity AI introduces.
And the trend isn’t limited to startups with strong opinions. In Cloudera’s research, 63% of Asia Pacific enterprises said they had shifted AI workloads back to private cloud or on-premises environments to regain control.
Repatriation is a placement decision, not a protest
It’s easy to frame repatriation as a rejection of the cloud. That’s not what the data shows. As ZDNet notes, full public-cloud exits remain rare. Most organizations move some workloads to private or on-premises infrastructure and keep the cloud for the workloads that genuinely benefit from elastic scale.
The cloud remains a great fit for spiky, experimental, and fast-growing workloads. The question is what happens to workloads that aren’t any of those things.
ZDNet’s analysis identifies the deciding factor: a workload’s predictability and utilization. Workloads that run constantly, at consistent levels, keep the meter spinning around the clock. Over time, the economics of owning or reserving that capacity can beat paying for it by the second.
Why analytics and AI are at the center of this
Look at the profile of a mature enterprise analytics environment: thousands of users, hundreds of concurrent queries, nightly loads, always-on dashboards, and now AI retrieval and agentic workloads on top. That’s about as steady-state and high-utilization as enterprise computing gets, which is exactly why data warehouses keep coming up in repatriation conversations.
But moving these workloads is only worth it if the destination doesn’t give up what made the cloud attractive in the first place. Data teams still want elasticity, separation of storage and compute, modern operations, and the freedom to burst into the cloud when they need to. Repatriation that means a return to rigid, hard-to-scale legacy systems simply trades one problem for another.
That’s why the smartest repatriation strategies start with a hybrid-ready platform:
- One platform in both places. The same SQL, behavior, and tooling on-premises and in the cloud, so placement decisions don’t require application rewrites.
- Cloud-native operations everywhere. Kubernetes-based deployment and elastic compute, whether the hardware is in a hyperscaler region or your own data center.
- Economics you can forecast. For predictable workloads, predictable infrastructure costs.
Where Yellowbrick fits
Yellowbrick was designed so enterprises never have to choose between cloud-native operations and infrastructure they control. The same Yellowbrick SQL data platform runs on AWS, Azure, and Google Cloud, in customers’ own accounts, and on-premises, including on Dell infrastructure. It is a cloud-native, Kubernetes-based stack with separate storage and compute and elastic virtual compute clusters managed through SQL. That means a workload can move from the cloud to a private data center, or back, as costs, regulations, and performance needs change, without re-platforming the analytics built on top of it.
Cloud repatriation in 2026 isn’t about going backward. It’s about making a deliberate choice, workload by workload, about where each one runs best, and choosing a data platform that makes that choice reversible.
Related Resources
- Cloud Migration
- Enterprise Data Warehouse
- How to pull sensitive workloads back from public cloud without stepping back into legacy architectures
- The Hyperscaler Tax: What Snowflake’s $6 Billion AWS Commitment Means for Your Data Warehouse Bill
- Is There a Role for Cloud Repatriation in Data Warehousing?