For a decade, "cloud-first" was the default answer to every infrastructure question I heard in a boardroom. That reflex is breaking. Broadcom's Private Cloud Outlook 2026, a survey of 1,800 senior IT leaders, found 83% are considering or have already repatriated workloads from public cloud — up from 69% a year earlier — and half have already pulled something back. AI is the accelerant: public cloud's share of production inference workloads fell from 56% to 41% in a single year.
Cost Has Flipped the Equation
Security used to drive repatriation. Now it's money. For the first time, cost overtook security as the top public cloud concern in the study, and 97% of IT leaders say some of their public cloud spend is wasted — 52% put the waste above 25%. A separate Cloudian survey of 203 enterprises found 93% are repatriating or evaluating AI workloads, with 40% reporting cloud AI costs exceeded projections. The culprit is token shock: a GPU-heavy inference fleet running 24/7 with egress fees and consumption pricing stops making economic sense at scale.
Placement Is Becoming a Discipline
The forward-looking story isn't "cloud versus on-prem." It's workload placement as a core competency. The teams that will win are the ones treating infrastructure as a portfolio with per-workload unit economics — cost per useful token, per transaction, per active customer — and that can move workloads in both directions without re-architecture. I read vendor surveys with healthy skepticism; Broadcom sells private cloud, after all. But the EU's €180 million sovereign-cloud tender and the independent data point the same way. Religious cloud allegiance is finished.
What to Do Now
Build the repatriation muscle before you're forced to. Instrument the unit cost of every AI workload now, while the token bill is still legible. Treat data gravity and latency as placement inputs, not afterthoughts. The cloud isn't dying — but "cloud-first" as a reflex is.