Enterprises Shift to Hybrid-First Stacks for AI, Cost

Companies are replacing cloud-first strategies with hybrid-first stacks that place workloads across hyperscalers, private data centres, sovereign clouds and edge sites. IT teams now decide where each application should run to control costs, meet data-location rules and support AI workloads with required latency.

Rising public-cloud bills, governance gaps and stricter data-sovereignty rules coincided with rapid AI growth. Early cloud moves often tied applications to vendor services; some organisations are reorganising infrastructure to reduce long-term dependency on a single provider.

Irin Rahman, CTO at Audiences, described hybrid-by-design as planning how environments are built and operated from the start. She warned that overlapping SaaS and vendor platforms create operational complexity that can grow without standardisation and governance.

Portability and standardisation are central to the approach. Teams use containers and Kubernetes to keep software consistent across multiple platforms. Joe Baguley, Broadcom’s EMEA CTO, recommended open standards and a layered architecture: “If you want to avoid lock-in, you must stratify your architecture.” He noted that different Kubernetes configurations for networking, security and storage can reintroduce operational complexity.

Workload placement is now driven by economics and control. Moving large datasets between environments increases costs and raises security and compliance exposure. Leo Derikyants, CEO of Mind Simulation Lab, described “data gravity” as the tendency for heavy datasets to remain where they are, and said compute should be moved to that data when possible.

Some steady-state workloads are returning to private infrastructure. Long-running AI inference systems and large databases can cost less and be easier to manage on owned hardware, while hyperscalers continue to provide burst capacity and environments for rapid experimentation.

Edge computing and AI are increasing demand for distributed deployments. Low-latency processing and geographically spread endpoints require local compute on factory floors, retail sites, vehicles and remote locations. Kate Obiidykhata, Percona’s cloud-native lead, cited research showing 66% of organisations host generative AI inference on Kubernetes.

Organisations are building internal platforms, automation and platform engineering to reduce operational overhead and enforce governance. Derikyants pointed to GitOps and automation as methods to improve resilience: “With a GitOps approach, your whole system is described upfront, and the infrastructure just follows it.”

IT teams are creating formal frameworks to match applications with the environments best suited for jurisdictional, performance and cost requirements. Open technologies, containerisation and orchestration are being used to keep those multi-environment setups consistent and manageable.