The Great Repatriation: Why Enterprise AI Is Moving Back On-Prem

The Great Repatriation: Why Enterprise AI Is Moving Back On-Prem

Apr 11, 2026 - 7 Min read

The Pendulum Swings

For the better part of a decade, the default answer to every infrastructure question was “put it in the cloud.” It didn’t matter what the workload was, what the data sensitivity looked like, or whether the economics actually made sense. Cloud was the answer, and anyone who questioned it was legacy.

That era is ending. Not because the cloud failed, but because AI workloads are different — and enterprises are doing the math.

New research from Broadcom and the Cloud Security Alliance shows that 59% of enterprises are now adopting hybrid cloud strategies, with 16% moving to fully private infrastructure. Private cloud revenue growth has doubled year over year, jumping from 13% to 25%. The tide isn’t just turning. It’s pulling entire industries back toward infrastructure they own.

This isn’t a regression. It’s a correction.

The Numbers Don’t Lie

The shift is showing up in every data set that matters.

A March 2026 global study found that AI is the primary driver behind hybrid and private cloud adoption. It’s not legacy migrations or cost optimization exercises. It’s organizations looking at what AI workloads actually require — sustained GPU compute, sensitive training data, predictable throughput — and concluding that the public cloud model doesn’t fit.

The same study found that 82% of organizations rate data sovereignty as extremely or very important when it comes to AI deployments. Not “somewhat important.” Not a line item on a compliance checklist. The vast majority of enterprises consider it a primary decision factor.

In regulated industries — finance, healthcare, defense, government — the picture is even starker. 60% of enterprises in regulated sectors prefer private cloud or sovereign deployment options for their AI workloads. These aren’t fringe holdouts. These are the industries that process the most sensitive data, operate under the strictest oversight, and face the biggest consequences when something goes wrong.

What Changed

Three forces converged to make this inevitable.

Regulation got teeth

The EU AI Act is now in effect. GDPR enforcement has intensified. In the United States, state-level data privacy laws are multiplying — California, Colorado, Connecticut, Virginia, and counting. Each jurisdiction adds new requirements about where data can reside, how it can be processed, and who can access it.

For AI workloads, these aren’t abstract compliance concerns. Training data, inference inputs, model outputs — all of it falls under regulatory scrutiny. When your AI processes customer data, patient records, financial transactions, or classified information, the question “where does this run?” has a legal answer now, not just a technical one.

Sending that data to a shared cloud environment in another jurisdiction isn’t just risky. In many cases, it’s non-compliant.

The economics tipped

Here’s the number that matters: when cloud costs exceed 60-70% of equivalent on-premise acquisition costs, on-prem becomes the more economical option.

For traditional cloud workloads — variable, spiky, hard to predict — the cloud pricing model makes sense. You pay for what you use. The elasticity is worth the premium.

AI workloads are different. Training runs are planned. Inference workloads — which now represent the majority of enterprise AI compute — follow predictable patterns. You know how many requests you’ll process. You know what throughput you need. You can forecast GPU utilization with reasonable accuracy.

When the workload is predictable, the cloud premium is just a tax. And it’s a tax that scales linearly with adoption. Every new model, every new use case, every expansion of AI across the organization adds to the bill. At enterprise scale, the numbers get uncomfortable fast.

Organizations are running the TCO analysis and arriving at the same conclusion: for steady-state AI workloads, owned infrastructure is cheaper. Often significantly so.

The vendors showed their hand

The most telling signal isn’t what enterprises are doing. It’s what the cloud and infrastructure vendors are doing.

Microsoft launched Sovereign Cloud with support for large AI models running fully disconnected from the public internet. Not limited models. Not hobbled versions. Full-scale AI, air-gapped.

VMware (now under Broadcom) is positioning VCF as the foundation for private AI , explicitly citing data sovereignty as the driving factor. Their 2026 predictions report doesn’t hedge — it calls private cloud the future of enterprise AI infrastructure.

A new class of “neocloud” providers has emerged, building sovereignty-first, AI-optimized infrastructure from the ground up. These aren’t traditional hosting companies rebranded. They’re purpose-built for organizations that need GPU compute without sending data to someone else’s environment.

When the vendors pivot this hard, it’s not speculative. They’re responding to what their enterprise customers are demanding.

This Isn’t Anti-Cloud

Let’s be clear about what this is and isn’t.

This isn’t a return to the data center era of 2010. Nobody is arguing that every workload should run on iron you own. The cloud remains the right answer for a lot of things — elastic web applications, SaaS delivery, global content distribution, dev/test environments, burst capacity.

But AI workloads have specific characteristics that break the cloud-first assumption:

High sustained compute — GPU workloads aren’t ephemeral web requests. They’re long-running, resource-intensive processes that benefit from dedicated hardware.
Sensitive data — AI models ingest and process the most sensitive data in the organization. Training data, inference inputs, and outputs all carry risk.
Predictable patterns — inference workloads have forecastable demand. The cloud premium buys elasticity you don’t need.
IP concentration — model weights, fine-tuning data, and prompt engineering represent significant intellectual property. Hosting that IP on third-party infrastructure introduces risk that’s hard to price.

The right framework isn’t “cloud vs. on-prem.” It’s: what does this specific workload require, and what’s the best infrastructure to support it?

For many AI workloads, the answer is infrastructure the organization controls. That might be a private cloud. It might be on-premise hardware. It might be a sovereign deployment in a specific geography. It might be fully air-gapped.

The point is that the decision should be driven by the workload’s requirements — data sensitivity, regulatory posture, cost profile, performance needs — not by a default assumption that everything belongs in the public cloud.

What This Means in Practice

The organizations getting this right aren’t making binary choices. They’re building flexible infrastructure that matches deployment to workload.

Development and experimentation might run on cloud GPUs where speed matters more than cost. Production inference for sensitive workloads runs on owned infrastructure where data never leaves the perimeter. Fine-tuning with proprietary data happens in isolated environments. Compliance-critical deployments go into sovereign or air-gapped configurations.

This requires AI platforms that aren’t architecturally wedded to a single deployment model. If your tooling only works when it can phone home to a cloud API, you’ve already lost the flexibility that the market is demanding.

Calliope was designed around this reality — deployable on-prem, in private cloud, or fully air-gapped, with the same capabilities regardless of where it runs. Not because we predicted a trend, but because the requirements of enterprise AI have always pointed in this direction.

The Repatriation Accelerates

Broadcom’s 2026 predictions suggest this is still early innings. As AI adoption deepens, as regulatory frameworks mature, and as enterprises get better at modeling the true cost of cloud AI at scale, the movement toward private and hybrid infrastructure will accelerate.

The organizations that will navigate this best are the ones making infrastructure decisions based on workload characteristics rather than ideology. Cloud where cloud makes sense. Private where private makes sense. Air-gapped where the requirements demand it.

The great repatriation isn’t about rejecting the cloud. It’s about finally treating infrastructure as an engineering decision rather than a religious one.