AI is breaking the economic logic of the public cloud

For over a decade, enterprise cloud strategy followed a clear trajectory. Organizations moved workloads to the public cloud to gain scalability, flexibility and cost efficiency. Hyperscalers such as Amazon Web Services, Microsoft Azure and Google Cloud Platform became the default foundation for digital transformation.

That model is now starting to break. The same force driving the next wave of investment, AI, is exposing its structural limits, particularly as organizations confront the growing unpredictability and scale of AI infrastructure costs.

AI workloads require sustained, high intensity compute, operate on large volumes of data and are often latency sensitive. At scale, they disrupt the economic and architectural assumptions that made the public cloud attractive in the first place. Compute costs become high and difficult to predict, large-scale data transfer becomes inefficient and regulatory constraints increasingly limit where data can reside and be processed.

What is emerging is not a retreat from the public cloud, but a shift away from treating it as the default. Infrastructure decisions are becoming workload-specific, cost-aware and deliberately governed. The question is no longer whether to use the cloud, but when, where and for which workloads it creates real value.

Why cloud-first no longer works

Cloud first simplified decision-making when most workloads benefited from elasticity and on-demand infrastructure. Under AI-driven demand, that assumption no longer holds.

A new model is emerging, centred on workload economics. Instead of asking where infrastructure is available, organizations are asking where each workload can run most efficiently, given its compute profile, data gravity, latency sensitivity and regulatory constraints.

This is not multi-cloud in the traditional sense. It is not about distributing workloads across providers for redundancy or vendor diversification. It is about intentional placement.

Public cloud remains optimal for elastic and globally distributed workloads. Private or on-premises environments are becoming more viable for sustained, high intensity compute and data-intensive processing. Sovereign and regional environments are increasingly required where regulatory and jurisdictional constraints apply.

Specialised providers are also emerging to address high-performance AI compute, cost-optimized GPU access and workload profiles that hyperscalers do not efficiently serve.

Consider a large-scale AI training workload running continuously over several weeks. In a public cloud environment, the combination of sustained GPU usage, data transfer and storage can result in costs that are both high and difficult to predict. The same workload, executed on dedicated or co-located infrastructure closer to the data, can offer significantly greater cost stability and performance consistency.

The difference is not marginal. It fundamentally changes the economic profile of the workload.

No single environment optimises for cost, performance, compliance and control simultaneously. The objective is no longer standardization. It is the ability to make consistent, economically sound placement decisions.

The real constraint is execution discipline

The limiting factor in modern cloud strategy is no longer access to infrastructure. It is the ability to apply consistent economic discipline to how that infrastructure is used.

Managing a portfolio of environments is materially more complex than operating within a single platform. It requires clear ownership of workload placement, visibility into cost and performance trade offs, and alignment between architectural choices and business objectives.

Most organizations are not equipped to do this well.

Teams optimise locally. They prioritise speed, adopt the most accessible tools and accelerate development. This works in isolation, but it does not hold at scale. AI workloads deployed for speed in public cloud environments often reveal, over time, cost and data locality constraints that make those architectures unsustainable.

The result is predictable. Fragmented architectures, escalating costs and inconsistent outcomes.

The problem is not infrastructure. It is the absence of a coherent model for governing how infrastructure decisions are made.

In this context, cloud performance becomes a function of organizational capability. High-performing organizations treat cloud as a governed business capability. They define clear accountability for workload placement, establish visibility across environments and align infrastructure decisions with strategic priorities.

Others experience the opposite. Growing complexity, limited control and rising cost without proportional value.

Why infrastructure strategy is becoming more deliberate

This shift is not optional. It is being driven by structural constraints around data, regulation and infrastructure.

Private cloud is resurging as a response to constraints the public cloud cannot efficiently accommodate. Sovereign cloud is becoming a strategic requirement rather than a niche consideration.

Data sovereignty now extends beyond storage location. It includes who can access data, how it is processed and under which legal frameworks it operates. This introduces architectural segmentation by necessity, with workloads distributed across jurisdictions, each with distinct constraints.

Hyperscalers are adapting with sovereign offerings. Regional providers are gaining traction by addressing specific regulatory and performance requirements. Specialised platforms are emerging to support high intensity AI workloads more efficiently.

The result is not fragmentation for its own sake. It is a shift towards more deliberate, constrained and economically grounded infrastructure decisions.

The cloud becomes a compositional model

The cloud market is evolving into a compositional model, where different environments serve distinct roles.

Public cloud remains critical, particularly for capabilities that are difficult to replicate, such as managed services, global distribution and integrated tooling. However, it is no longer sufficient as a single, dominant environment.

Alternative providers are carving out specialised roles. Some focus on high-performance AI compute. Others prioritise regulatory compliance or cost efficiency. Private infrastructure continues to play a role where control and predictability are essential.

This is not a winner-takes-all market. It is a compositional one, where value is determined by how effectively organizations combine different environments.

Fragmentation is not a weakness. It is a response to the increasing diversity of workload requirements.

The new question for CIOs

The central challenge for CIOs is no longer cloud adoption. It is the ability to operate across multiple environments in a consistent and economically rational way.

Most organizations have already chosen their providers. The more difficult problem is determining how to use them.

This requires a shift in mindset. Architecture must take precedence over platform selection. Governance must take precedence over convenience. Capability must take precedence over tooling.

AI has not diminished the importance of cloud. It has exposed the limits of using it indiscriminately.

The organizations that succeed will not be those with the best cloud platforms, but those with the discipline to decide, consistently and at scale, where each workload should run.

In this environment, cloud strategy becomes less about infrastructure and more about economic control.

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