A comprehensive analysis of 300 enterprise workloads finds that the true cost gap between on-premises and cloud is far narrower than most TCO models suggest, especially at scale.
Key Takeaways
For the better part of a decade, the conventional wisdom in enterprise IT has been clear: cloud is cheaper, more flexible, and strategically superior to on-premises infrastructure. Board presentations were built around it. Vendor-supplied TCO calculators reinforced it. And CIOs who pushed back were frequently told they were standing in the way of digital transformation. New research from IDC, covering 300 enterprise workloads across 14 industry verticals, suggests the picture is considerably more complicated, particularly for organisations operating at scale.
The study found that at the 500 or more VM threshold, mature on-premises deployments cost an average of just 12% more than equivalent cloud spend once all cost components are accounted for accurately. Below that threshold, cloud economics remain more favourable, particularly when factoring in the capital avoidance of hardware procurement and the operational overhead of facilities management. But the research makes clear that most organisations have been running TCO comparisons with a structural flaw: they consistently captured cloud's flexibility benefits while leaving significant cost line items off the ledger entirely.
The single largest systematic error in cloud TCO models is the treatment of egress costs. Data egress, the charge levied by cloud providers whenever data moves out of their network, is notoriously difficult to model in advance because it scales with data volume in ways that are hard to predict at the time of migration planning. The IDC research found that 67% of enterprises have materially underestimated their egress exposure, with the average mid-size enterprise incurring $240,000 in annual egress charges that were not present in the original business case. For organisations running data-intensive analytics pipelines, streaming workloads, or hybrid architectures that require frequent data movement between cloud and on-premises environments, the real-world figures can be substantially higher.
Data residency and sovereignty requirements add a further layer of cost complexity that vendor TCO calculators rarely surface. Regulatory frameworks across the EU, India, Brazil, and an expanding list of other jurisdictions now mandate that certain categories of data be stored and processed within geographic boundaries. Meeting these requirements in public cloud environments typically requires purchasing premium regional availability and, in some cases, deploying dedicated infrastructure within the provider's network at a significant price uplift. On-premises infrastructure in the required geography, by contrast, satisfies residency requirements as a baseline condition rather than a premium feature.
The staffing cost differential deserves more nuanced treatment than it typically receives. Cloud advocates correctly note that on-premises infrastructure requires personnel with deep expertise in hardware, virtualisation, storage, and networking. What is less frequently acknowledged is that cloud operations require their own specialist skill set, and the market for engineers with deep expertise in cloud cost optimisation, reserved instance management, and FinOps disciplines has become intensely competitive, with compensation premiums that frequently exceed the savings assumed in the original staffing cost model. The nature of the cost shifts rather than disappears.
Cloud vendors' own TCO calculators present a further analytical challenge. These tools are sophisticated, well-designed, and genuinely useful as directional guides, but they are built to produce outcomes favourable to migration. Independent analysis consistently finds that vendor calculators underweight egress, apply optimistic utilisation assumptions, exclude licence portability friction costs, and model on-premises depreciation against new hardware prices rather than the actual book value of existing assets that an organisation would retire. Using a vendor calculator as the primary analytical tool for a major infrastructure decision is, in effect, asking a sales organisation to adjudicate its own business case.
"The mistake most organisations made in the 2020s was treating cloud migration as a binary choice and running TCO models that only captured the easy-to-quantify costs. The egress bills, the licence portability restrictions, the re-platforming cost: these were either ignored or underestimated almost universally."
Dr. Kenji Watanabe, VP Infrastructure Strategy, IDC
The most useful output of the IDC research is not a verdict in favour of either model but a workload classification framework that allows infrastructure teams to make placement decisions with genuine analytical rigour. The framework evaluates each workload across five dimensions: compute demand predictability, data movement intensity, latency requirements, regulatory exposure, and organisational capability to manage the relevant infrastructure type. These five variables, when mapped together, produce a workload fitness score for cloud versus on-premises deployment that is meaningfully more accurate than headline cost comparisons alone.
Workloads with high compute demand predictability, significant data movement intensity, sub-5ms latency requirements, or hard data residency obligations score consistently better for on-premises placement. These include steady-state enterprise applications with stable user populations, large-scale analytics pipelines that process and move data continuously, manufacturing and industrial systems with real-time control requirements, and regulated financial workloads subject to data localisation mandates. The research found that organisations applying the workload fitness framework retained an average of 34% of their VM estate on-premises rather than migrating it, realising TCO savings that more than offset the cost of the analysis itself.
The opposite pattern holds for workloads with unpredictable or highly variable compute demand, minimal data movement requirements, no hard latency constraints, and no regulatory residency obligations. Development and test environments, early-stage applications with uncertain scaling trajectories, seasonal or event-driven workloads, and analytics applications that primarily read from rather than write to large datasets all display strong cloud fitness. For these workloads, the pay-as-you-go elasticity of public cloud delivers genuine cost advantages that on-premises infrastructure cannot match, regardless of how mature the infrastructure team is.
The emerging default architecture is neither pure cloud nor pure on-premises but a deliberately designed hybrid: cloud for the workloads where it genuinely wins on cost and flexibility, and on-premises or colocation for the workloads where scale, data intensity, or latency requirements make retention the more economical choice. This is not the hybrid cloud of the early 2020s, which often meant "cloud for new workloads and on-premises for everything we haven't migrated yet." It is a deliberate, continuously re-evaluated portfolio approach in which workload placement is treated as an ongoing optimisation problem rather than a one-time migration decision.
For IT leaders currently operating under a cloud-first mandate, the research does not argue for reverting to on-premises infrastructure. It argues for replacing the mandate with a workload-first analytical discipline. The organisations that have done this, applying structured fitness assessments before each migration decision rather than defaulting to cloud as the answer, report higher infrastructure satisfaction scores, more predictable cost trajectories, and significantly fewer instances of post-migration re-architecture driven by cost or performance problems. The fundamental lesson is straightforward: rigorous workload-by-workload analysis consistently outperforms wholesale strategy mandates, in both directions, regardless of which way the mandate points.
New deployment data shows SD-WAN has crossed the mainstream threshold. IT teams that delayed migration now face a compounding performance and cost penalty.
Edge deployments have moved from pilot to production across manufacturing, retail, and healthcare, with latency savings exceeding early projections.
