The post-cloud data center: Back in fashion, but not like before

For most of the last decade, I watched enterprise infrastructure strategy follow a simple arc: abstract complexity, speed up provisioning, move as much as possible into the cloud. That shift delivered real value. It shortened deployment cycles, empowered product teams and removed capital friction that had slowed change.

Cloud did not eliminate the need for physical infrastructure. It only postponed the moment when we would need to think about it again.

That moment is here.

In conversations with platform leaders, executive sponsors and IT leadership, questions have shifted from “how fast can we migrate?” to “where should this run and what risk are we taking if we choose wrong?” The catalyst is not nostalgia for owned data centers. It is the collision of AI, energy constraints, sovereignty expectations and GPU economics.

What appears to be a return is actually architectural maturity. We have entered the post-cloud data center era.

Why “back” is the wrong word

If you describe this moment as a reversal, you miss what changed. The first wave of cloud strategy optimized for velocity and elasticity. We wanted to escape procurement cycles, scale on demand and give more control to software shipping teams. That model is still the right answer for many workloads.

AI exposes the assumptions behind that universal default.

When models move from experimentation to daily operations, elasticity is no longer the dominant variable; GPU usage stabilizes; and data volumes grow rapidly. The cost curve becomes less forgiving. At the same time, boards and regulators ask more pointed questions: Where is data processed? Where are models trained? Who controls the infrastructure beneath it? What evidence exists when an auditor asks?

This is why I do not frame the shift as “cloud repatriation.” It is replatforming at the infrastructure layer. I am not arguing for an on-prem or colocation expansion in isolation. I am arguing for a deliberate placement model where cloud, colocation and on-prem each have defined roles, decision gates and an evidence package when you deviate. Placement is becoming situational, based on density, locality and governance, not ideology.

The survey data support the direction, even if every enterprise will land differently. In its 2024 Global Data Center Survey, the Uptime Institute reports that 64% of enterprise operators are growing their data center capacity, even as colocation and public cloud expand. That is not a mass retreat from the cloud. It is a signal that hybrid is hardening into a long-term operating model, especially as AI workloads mature.

In my architecture work, I see two triggers that bring physical infrastructure back into scope. First, sustained utilization changes the math. A steady, always-on inference pipeline behaves differently from spiky batch processing. If the workload is stable, the economic advantage shifts toward locations where you can control the unit costs of power, cooling and amortization.

Second, data gravity and accountability show up late in the cloud conversation and then dominate it. A proof of concept can run anywhere. A production system tied to regulated data, proprietary IP, customer confidence and board scrutiny rarely can.

Edge is now accountability

The biggest mindset shift I have had to make is changing what “edge” means.

Historically, edge computing meant spatial distance from the core: factories, stores, remote sites. In the AI era, edge now means proximity to accountability. Compute is moving closer to proprietary data, regulatory boundaries and operational decision-making. In practice, that often means enterprise facilities and colocation sites within defined legal and governance zones.

You can see the policy pressure building. Public sector briefings now treat data centers as part of national resilience and sustainability planning, not just private infrastructure. AI policy is also tied to governance and trust, which increases the burden of demonstrating where and how sensitive processing occurs, as reflected in the European Commission’s approach to AI.

This is where colocation has evolved from “outsourced real estate” to a deliberate architecture move. In several programs I have been pulled into, colocation is where the enterprise anchors GPU-dense clusters near power-rich regions, keeps sovereignty-bound workloads inside a controlled footprint and connects to multiple clouds without turning every workload into a single-provider dependency.

The key is control, not location.

If you are building AI capabilities that touch customer data, pricing models, supply chain enhancement or proprietary process know-how, the question is rarely “can the cloud do it?” The question is “can we prove, continuously, that we are operating inside the boundaries our risk owners will accept?” For many organizations, the cleanest proof lives in infrastructure they can audit end-to-end, whether owned or in colocation facilities.

This is also why “data locality” discussions are starting to sound like “data center” discussions again. Once you accept that some data cannot move freely and some models cannot train outside certain jurisdictions, placement becomes a design decision, not just a hosting preference.

What I ask leadership now

AI has pulled infrastructure decisions back into the executive agenda. I am seeing senior technology leaders and steering committees ask detailed questions about topics they delegated for years: rack density, power topology, cooling strategy, site selection and long-term capacity planning. That is not because they want to run facilities. These constraints now shape business outcomes.

GPU density is the forcing function. In NVIDIA’s GPU-ready data center guidance, liquid cooling and AI-optimized designs can enable roughly two-to-five times higher compute density than traditional air-cooled approaches, depending on GPU generation, cooling method and utilization targets. Treat that as a planning range, not a promise. It only works if power delivery, cooling and rack design are engineered together. Legacy enterprise sites were not built for that profile. Power, not square footage, becomes the limiting factor, changing which sites are viable and how quickly capacity can scale.

Energy is the second pressure point. In the cloud era, energy was bundled into pricing models. With AI, energy shows up as a hard constraint and a reporting obligation. Forecasting, securing and governing power capacity is now part of the technology plan, not a facilities footnote. The World Resources Institute has been explicit about the challenge of forecasting electricity needs amid the data center boom, which is exactly the problem AI workloads amplify.

Uptime Institute also highlights how operational constraints, resiliency and sustainability reporting are becoming first-class issues for data center operators, not optional extras. That matters because boards now treat AI as both an engine of growth and a source of risk, which means they will ask for evidence and discipline.

When I help executive sponsors make this practical, I use a short set of questions that forces clarity but without turning it into a cloud debate:

1. Which AI workloads are steady and which are bursty? If utilization is stable, treat the cost curve like a utility problem. If demand is spiky, cloud elasticity may still win.
2. What data can move, and what data cannot? Define the non-negotiables with risk and legal early. If you cannot move data, do not design as if you can.
3. What is the density plan? Document target rack power, cooling approach and upgrade path. If the answer is “we will figure it out later,” AI will arrive before the infrastructure can support it.
4. What is the evidence plan? By evidence, I mean the artifacts that survive audits and incidents: reference architecture, power and capacity model, security control mapping, runbooks, disaster recovery test evidence and cost telemetry.
5. What is the exit plan? Avoid permanent placement decisions where possible. Design for movement between cloud, colocation and on-prem as requirements evolve.

One more test I use is to ask whether our AI roadmap assumes power and cooling scale as fast as software. They do not. That mismatch creates the most expensive technical debt: business commitments built on infrastructure that cannot arrive in time.

To keep this from becoming an ad hoc fight between cloud and facilities, I push for a simple governance pattern: classify workloads by density and data sensitivity, map each class to approved landing zones (cloud, colocation or on-prem) and require an evidence package for exceptions. That keeps decisions fast and defensible, which sponsors and steering committees need as AI adoption grows.

This is the posture I push for in reviews: cloud, where it is elastic and safe, on-prem or colocation, where spatial density, locality and governance demand it, and a purposeful design for reversibility across all of it.

On-prem is back in fashion, but not like before. The story is not a return to the past. This is the moment infrastructure stopped being abstract and became a strategic constraint again.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?