Enterprise infrastructure teams spent a decade chasing elasticity, and the public cloud delivered it generously. Procurement cycles shortened, GPU capacity arrived on demand, and finance teams stopped budgeting for hardware refresh cycles. That arrangement worked cleanly while workloads stayed bursty and unpredictable. Generative AI changed the underlying math, because inference and training rarely behave that way once they reach production scale. Continuous, always-on compute punishes rented infrastructure in ways that occasional analytics jobs never did. A counter-movement has formed quietly across data centre design teams, procurement departments, and compliance offices worldwide. Recent survey data captures the scale of this shift with striking clarity. A February 2026 study commissioned by Cloudian found that the vast majority of enterprises have already moved AI workloads away from public cloud or are actively evaluating the move. Nearly four in five respondents confirmed they had relocated at least some AI infrastructure already. Three converging pressures explain the pattern: unpredictable cost growth, tightening data sovereignty rules, and latency requirements that rented infrastructure cannot reliably satisfy. None of these pressures existed at this intensity five years ago. Together, they have turned a niche IT debate into a board-level conversation.
Cloud Cost Models Break Down Under Continuous AI Load
Public cloud pricing was designed for variable demand, not for GPU clusters running around the clock. A workload that trains or serves models continuously accrues hourly charges with no natural ceiling. Finance departments that budgeted for seasonal spikes now face compute bills that climb every quarter without warning. Nearly half of surveyed enterprises cite cloud cost unpredictability as a direct barrier to scaling their AI initiatives further. Forty percent report that actual AI spending has already exceeded original projections by a wide margin. Egress fees compound the problem further, since moving large training datasets between regions or providers adds recurring charges few teams anticipated. The 37signals case study illustrates this dynamic at a scale most enterprises can recognise. After nearly two decades as an AWS customer, the company watched its annual cloud bill climb past $3.2 million for largely predictable workloads. Leadership committed roughly $600,000 to Dell servers and NVMe storage across two colocation facilities instead. Annual infrastructure spend dropped toward $1.3 million within a single year of the transition. Projected savings now approach $10 million across a five-year horizon, according to public statements from the company. Few enterprises will replicate that exact arithmetic, yet the underlying lesson travels well beyond one firm’s specific circumstances.
Data Sovereignty Rules Are No Longer Optional
Regulators across multiple regions have shifted data residency from a guideline into an enforceable requirement. Financial institutions operating in the European Union now answer to the Digital Operational Resilience Act, which demands demonstrable infrastructure control rather than contractual promises alone. Healthcare and government sectors face comparable pressure in jurisdictions across Asia and the Middle East. According to Nutanix’s 2026 Enterprise Cloud Index, more than half of surveyed IT leaders feel compelled to keep infrastructure within a single country’s borders. That figure reflects a structural shift in how compliance teams evaluate vendor relationships. Sensitive training data and proprietary model weights increasingly need to stay inside a perimeter the enterprise fully controls.
Sovereignty concerns intersect directly with how organisations deploy large language models internally. Companies increasingly resist sending proprietary business context to general-purpose external models hosted on shared infrastructure. Executives at data and AI firms describe this as a closeness requirement between sensitive data and the systems processing it. Agent-based AI deployments raise the stakes considerably, since autonomous systems require continuous access to confidential operational data. Containment and containerisation inside an owned environment offer guardrails that multi-tenant cloud platforms cannot fully replicate. Compliance officers now sit at the same planning table as infrastructure architects when AI roadmaps get drafted.
Specialised Data Halls Replace General-Purpose Cloud Regions
The infrastructure receiving these repatriated workloads looks nothing like a traditional server room. Modern colocation facilities purpose-build data halls around dense GPU racks, liquid cooling loops, and high-bandwidth interconnects. These environments handle thermal loads that older general-purpose data centres were never designed to manage. Enterprises increasingly lease dedicated capacity inside specialised facilities rather than purchasing land and building from scratch. This approach combines the capital discipline of private infrastructure with the deployment speed historically associated with cloud providers. Colocation operators have responded by accelerating build-outs specifically tailored toward AI inference and training clusters.
Hardware refresh cycles have also evolved to support this transition more affordably. Refurbished enterprise-grade GPU servers now offer performance comparable to new units at a substantial discount. Procurement teams avoid factory lead times that can stretch several months for newly manufactured AI hardware. Rigorous stress testing and component verification processes give buyers confidence these systems will perform reliably under continuous load. Investment in dedicated GPU infrastructure frequently reaches payback within four to six months for steady-state workloads. After that threshold, ongoing costs shrink to electricity, cooling, and colocation space rather than recurring rental premiums.
Hybrid Architecture, Not a Wholesale Cloud Exit
None of this signals an abandonment of public cloud infrastructure. Gartner projects that forty percent of enterprises will adopt hybrid compute architectures for mission-critical workloads by the end of this year. That figure stood near eight percent only a few years earlier, marking a dramatic structural shift. Analysts consistently describe the trend as workload-by-workload rebalancing rather than a coordinated retreat from hyperscale providers. Bursty, experimental, or development-stage workloads still suit elastic cloud capacity reasonably well. Steady-state inference, regulated data processing, and continuous training increasingly belong on infrastructure the enterprise directly controls.
Open-source models have made this hybrid posture considerably more practical than it would have been two years ago. Industry estimates suggest open models like Llama and Mistral now handle the large majority of enterprise use cases at quality indistinguishable from proprietary cloud APIs. That parity removes a major dependency that once tied enterprises tightly to specific cloud-hosted model providers. Deloitte’s analysis found that on-premises AI delivers cost savings exceeding fifty percent over three years once token volume crosses a meaningful threshold. Microsoft itself responded by launching sovereign cloud capabilities built for AI models running fully disconnected from public infrastructure. Even hyperscalers now acknowledge that disconnected, sovereign deployment has become a durable category of enterprise demand.
What This Means for Enterprise Infrastructure Strategy
Infrastructure leaders now approach placement decisions with considerably more rigor than the lift-and-shift era ever demanded. A workload audit typically precedes any repatriation commitment, tagging each system by utilisation pattern, compliance status, and current spend. Teams then build a fully loaded total cost of ownership model spanning three to five years. That model accounts for hardware, colocation, power, staffing, and the often-overlooked cost of inaction. Organisations that skip this discipline tend to repeat the very mistakes that pushed workloads toward the cloud originally. Deliberate, workload-specific placement consistently outperforms ideological commitments to either extreme. Power availability has emerged as an unexpected constraint shaping where these specialised data halls can even be built. Grid capacity in established markets struggles to keep pace with GPU density requirements at scale. Developers increasingly target regions with available power first, before considering connectivity or labour costs.
This dynamic explains why colocation growth has accelerated in markets previously considered secondary infrastructure hubs. Enterprises planning multi-year AI roadmaps now factor energy procurement into vendor selection alongside traditional metrics. The data hall has effectively become the new unit of competitive advantage in enterprise AI deployment, replacing the cloud region as the primary planning reference point. The broader lesson extends well past any single survey or case study cited here. Enterprises spent years treating cloud and on-premises infrastructure as a binary, all-or-nothing choice. That framing no longer matches how sophisticated infrastructure teams actually operate today. Workload placement now follows cost, compliance, and latency signals rather than default vendor relationships. Specialised, purpose-built data halls have become the destination for AI workloads that no longer fit cloud economics. The repatriation wave, in that sense, reflects infrastructure maturity rather than a rejection of the cloud computing model itself.
