The Reversal in Numbers
Published: July 20, 2026 | Category: Technology | By Mahesh
For most of the past fifteen years, the direction of enterprise infrastructure strategy was settled. Everything moves to the cloud, eventually, and asking why was treated as a sign a company had not modernised yet. That consensus has cracked. According to a Barclays survey of enterprise CIOs, 83 percent now report plans to repatriate at least some workloads from public cloud back to private infrastructure, and IDC separately found that 70 to 80 percent of companies are repatriating data annually.[1] For AI workloads specifically, the shift is even sharper. Cloudian's Enterprise AI Infrastructure Survey 2026, conducted in February across 203 enterprise IT decision-makers, found that 93 percent of enterprises have already repatriated AI workloads, are actively doing so, or are formally evaluating it, with 79 percent having already moved at least one AI workload off public cloud.[2] This is not, despite how the trend gets discussed casually, a wholesale retreat from the cloud. It is a far more interesting and more permanent shift: enterprises are finally pricing their infrastructure decisions properly, workload by workload, instead of defaulting to the cloud because that was the safe recommendation a decade ago. This piece examines what triggered this reversal, the actual economics behind it, why AI workloads specifically are leading the charge, and what the shift means for how technology leaders should be thinking about infrastructure through the rest of 2026.
What Actually Broke the Cloud-First Consensus
Three structural forces, not a single dramatic event, pushed cloud repatriation from a fringe conversation into a mainstream boardroom topic. The first and most direct is simply the scale cloud bills have reached. The average enterprise now spends $13.7 million annually on cloud infrastructure according to Flexera's 2025 State of the Cloud report, and a meaningful share of that spend covers workloads that have run at consistent, predictable utilisation for three to five years with no real elasticity benefit to show for the premium being paid.[3] Cloud pricing was built around the value of elasticity, the ability to scale instantly up or down as demand fluctuates. For a genuinely bursty workload, that value is real and repatriation makes little sense. For a workload running at steady, predictable load around the clock, the enterprise is paying an elasticity premium for a capability it never actually uses, month after month, for years.
The second force is a cost dynamic almost nobody discusses at the board level until they see it on an invoice: egress fees. Cloud egress charges, the cost of moving data out of a public cloud environment, are cited by AgamiSoft's 2026 analysis as the single most unexpected cost driver that triggers a formal repatriation review, with enterprises reporting bills ranging from $50,000 to $500,000 tied specifically to data transfer out of the cloud.[4] These charges are structurally asymmetric by design. Moving data into a cloud provider's environment is typically free or heavily discounted. Moving it back out is expensive, which is precisely why the fee structure functions as a form of soft vendor lock-in that becomes visible only once an enterprise tries to leave or shift a meaningful volume of data.
The third force is regulatory and reflects a shift that has nothing to do with cost at all. According to the Nutanix Enterprise Cloud Index 2026, 57 percent of IT leaders now feel a genuine need to run infrastructure within a single country, driven by intensifying regulatory pressure in financial services, healthcare and government, alongside continued uncertainty around the EU-US Data Privacy Framework.[5] For enterprises in regulated industries, keeping specific categories of data on-premise or within strict jurisdictional boundaries is moving from a best-practice recommendation to a hard compliance requirement, a shift that no amount of cloud cost optimisation can address because the constraint is legal rather than financial.
The Cost Math That Changed the Conversation
Broadcom's internal infrastructure analysis, drawn from its own operations after the VMware acquisition, found that modern private cloud delivers 40 to 50 percent lower total cost of ownership for steady-state workloads compared to equivalent public cloud deployment.[2] Broadcom itself moved critical workloads off public cloud database-as-a-service offerings and reported savings exceeding $10 million as a direct result. That is not a hypothetical case study from a vendor with something to sell. It is a company with genuinely enormous infrastructure scale making the calculation and acting on it.
The average repatriated workload, once on-premises hardware amortisation, colocation fees and ongoing operational overhead are all properly accounted for, saves 32 percent of its annual cloud infrastructure cost according to Gartner's 2025 analysis, specifically for the workload categories where the repatriation economics genuinely work.[4] That qualifier matters enormously and is the single most important nuance in this entire trend: repatriation economics do not apply universally. They apply specifically to workloads that are high-utilisation, predictable, compliance-sensitive or performance-critical, exactly the profile that gains the least from cloud elasticity in the first place and pays the steepest ongoing premium for a flexibility benefit it structurally cannot use.
Andreessen Horowitz's widely circulated analysis of what it termed the cloud paradox adds a valuation dimension that goes well beyond simple cost savings. Public cloud spending, the firm found, can weigh down a software company's gross margins by 50 percent or more, and repatriating the right workloads can meaningfully improve, in some cases roughly double, the market valuation multiple investors are willing to apply to the business.[1] For any founder or CFO thinking about infrastructure purely as an operating expense line, this is the reframe that matters. Cloud spend that depresses gross margin is not just a cost problem. It is a valuation problem every time the company raises capital or considers an exit.
Why AI Workloads Specifically Are Leading the Exodus
If cloud repatriation as a general infrastructure trend is significant, its concentration in AI workloads specifically is the part of the 2026 data that deserves the closest attention, because it inverts the assumption most technology leaders still carry: that AI, being new and compute-intensive, obviously belongs in the cloud where GPU capacity is elastic and immediately available.
The raw cost of continuous AI inference in the public cloud is the starting point for understanding why this assumption is breaking down. A single continuous, always-on 8-GPU cloud instance, the AWS p5.48xlarge being a representative example, can cost over $270,000 annually according to infrastructure cost analysis from PracticalLogix's work with enterprise clients.[6] For a workload running large language model inference around the clock rather than in short bursts, that is not an elastic, pay-for-what-you-use cost profile. It is a fixed, predictable, extremely large recurring bill, precisely the profile where owning the hardware outright starts to make overwhelming financial sense. PracticalLogix's direct client observation across 2025 and 2026 engagements found that migrating from persistent AWS p5 instances to a dedicated on-premise cluster typically eliminates up to 70 percent of annualised compute costs for continuous AI inference workloads specifically.[6]
Deloitte's independent analysis reaches a strikingly similar conclusion through a different lens: on-premise AI delivers 50 percent or more in cost savings over a three-year period compared to relying on cloud API alternatives, once an enterprise's token volume crosses a specific usage threshold that most organisations now cross far earlier than they expect given how quickly internal AI adoption has scaled.[2] Two developments made this shift technically viable in a way it simply was not two or three years ago. Open-source models including Llama and Mistral now handle 85 to 90 percent of enterprise AI use cases at quality that is genuinely indistinguishable from proprietary cloud APIs for most business applications, according to the same Deloitte analysis, removing the technical dependency on a cloud-hosted proprietary model that used to be the strongest argument for staying in the cloud.[2] And refurbished enterprise GPU hardware, offering performance identical to new units at a 40 to 60 percent discount while eliminating the three to six month factory lead times that used to make on-premise AI deployment impractically slow, has made the capital outlay for on-premise AI infrastructure dramatically more accessible than it was even eighteen months ago.[6]
There is also a timing dimension to this decision that most enterprises are not yet factoring properly into their planning. Cloud prices are rising 5 to 10 percent annually, while hardware prices, driven by continued demand pressure across the semiconductor supply chain, are rising considerably faster at 15 to 25 percent annually according to PracticalLogix's 2026 analysis.[6] That gap creates a genuine strategic timing question: enterprises that complete their repatriation procurement during 2026 lock in today's hardware prices and capture the full projected savings, while those that delay the decision into 2027 or 2028 may find that hardware price inflation has quietly eroded a meaningful share of the economic case that looked compelling when the analysis was first run.
This Is Not a Retreat From the Cloud
The most important nuance in the entire repatriation trend, and the one most likely to get lost in a headline, is that this is fundamentally not a rejection of cloud computing. The Flexera State of the Cloud Report found that 42 percent of workloads are moving from public cloud to private cloud or on-premise infrastructure specifically to optimise cost, while cloud adoption as a whole continues to grow in parallel.[1] Broadcom's own Private Cloud Outlook survey found 69 percent of respondents are actively considering workload repatriation, with a clear and consistent pattern in which workloads move where: unpredictable, bursty workloads stay in the public cloud where elasticity genuinely earns its premium, while stable, high-compute applications, the kind that run at consistent utilisation month after month, move in-house.[7]
As Jon Toor, CMO at Cloudian, put it in the company's 2026 Enterprise AI Infrastructure Survey release, enterprises are not abandoning the cloud so much as getting genuinely smarter about where specific AI workloads belong.[2] That framing captures what the underlying data actually shows more precisely than the "great repatriation" headlines that tend to accompany this story. The cloud-first decade, in which the default answer to nearly every infrastructure question was simply to lift and shift to a public cloud provider, is over. What has replaced it is a more deliberate, workload-by-workload evaluation in which the question is no longer whether to use the cloud but specifically which workloads earn their place there and which do not.
What This Means for Technology Leaders Making Infrastructure Decisions Right Now
Four practical implications follow directly from this data for any technology leader currently reviewing infrastructure spend or planning AI deployment for the second half of 2026.
Start by auditing utilisation patterns before evaluating any repatriation decision on cost projections alone. The entire economic case for repatriation rests on a workload being high-utilisation and predictable. A workload with genuine elasticity needs, where demand spikes and drops meaningfully, will not deliver the 32 percent average savings Gartner found, and may in fact become more expensive on dedicated infrastructure sized for its peak load rather than its average. The audit, not the vendor pitch or the industry headline, should drive the decision.
Treat continuous AI inference workloads as the highest-priority candidates for this analysis specifically. The economics here are the most extreme and the most consistently validated across independent sources, Deloitte, Broadcom, PracticalLogix and Cloudian all converging on savings in the 50 to 70 percent range for continuous, always-on AI inference moved off public cloud APIs onto dedicated infrastructure.[2] If your organisation is running large language model inference at meaningful, sustained scale rather than in occasional bursts, this is very likely the single highest-value infrastructure review available to you this year.
Factor egress fees into every cloud contract renewal conversation explicitly, not as an afterthought. The $50,000 to $500,000 in unexpected egress charges that AgamiSoft found triggering repatriation reviews are rarely visible until an enterprise actually attempts to move meaningful data volume. Modelling the true cost of leaving a cloud provider, not just the cost of staying, should be part of every major cloud commitment from the outset rather than a surprise discovered years into the relationship.
And build the timing question into any capital planning discussion around infrastructure this year. With hardware prices inflating two to three times faster than cloud pricing, a repatriation decision that pencils out clearly in 2026 may look considerably less attractive by 2028 purely due to the shifting cost of the hardware itself, independent of any change in cloud pricing or workload characteristics. This is precisely the kind of infrastructure decision where treating the analysis as a one-time exercise rather than continuous review carries a real, quantifiable cost, and it connects directly to the discipline explored in our earlier analysis of supply chain resilience strategy, where the businesses that model these decisions ahead of time consistently outperform those reacting to a cost or capacity shock after it has already arrived.
Common Questions
Sources
- Novoserve. Cloud Repatriation: Trends and Statistics of Enterprises Leaving Cloud, citing Barclays, IDC, Andreessen Horowitz and Flexera research. novoserve.com
- PracticalLogix. The Great Cloud Repatriation of 2026: Why 93% of Enterprises Are Pulling AI Workloads Home, citing Cloudian Enterprise AI Infrastructure Survey 2026 and Deloitte analysis. practicallogix.com
- AgamiSoft. Cloud Repatriation: Why Enterprises Move Back 2026, citing Flexera State of the Cloud 2025. agamisoft.com
- AgamiSoft. Cloud Repatriation: Why Enterprises Move Back 2026, citing Gartner 2025 workload savings analysis. agamisoft.com
- mrc's Cup of Joe Blog. What's Driving Cloud Repatriation in 2026, citing Nutanix Enterprise Cloud Index 2026 and Broadcom internal analysis. mrc-productivity.com
- NewServerLife. Cloud Repatriation in 2026: Moving AI and Heavy Workloads On-Prem, citing PracticalLogix client engagement data. newserverlife.com
- NewServerLife. Cloud Repatriation in 2026, citing Broadcom Private Cloud Outlook survey. newserverlife.com
Read More
- The Global Semiconductor Race in 2026: Why Chips Are the New Strategic Resource
- Supply Chain Resilience in 2026: Why Disruption Is Now Permanent
- Cybersecurity in 2026: What Every Startup and Business Must Know
- E-Commerce in 2026: Inside the $6.88 Trillion Market Being Rewritten by AI Agents
Article by Mahesh | Depth Grid