The cloud comes back down to earth: why the cloud calculus has changed in 2026

As AI drives unprecedented demand for compute, power and data-centre capacity, the cloud is being forced back down to earth. For telecom operators, the question is who controls the physical infrastructure powering the next phase of the digital economy?

For years, the cloud industry operated on the assumption that compute could scale almost without limit. That assumption is now being challenged as data-centre deployment becomes constrained by the physical realities of power availability, land, cooling, grid capacity, fibre connectivity and political consent. Rising energy demand, environmental opposition, geopolitical instability, supply-chain pressures and growing concerns around digital sovereignty are making cloud infrastructure harder to build, power and operate.

The arrival of AI-scale workloads is amplifying many of these pressures. Training and running large AI models requires vast quantities of compute power, electricity, cooling and high-capacity networking. According to Goldman Sachs, AI is predicted to drive a 165% increase in power demand for data centres by 2030.

Driven by the AI boom, the four largest hyperscalers are expected to spend around $700 billion in capital expenditure this year, more than double the annual investment of the entire global telecom sector. Yet concerns are growing that parts of this buildout may be running ahead of practical reality.

For telecom operators, this changes the public, private and hybrid cloud debate. The question is no longer which model offers the best combination of cost and scalability, but which workloads need to remain under direct control for reasons of resilience, sovereignty, compliance and service continuity. In an environment where networks must process enormous quantities of data and deliver services in real time, the balance between public and private infrastructure is becoming more important.

Large-scale AI infrastructure projects now face delays linked to power availability, grid capacity, transformer shortages, rising operational costs, planning and permitting challenges, labour shortages and environmental resistance. According to the European Data Centre Association (EUDCA), 67% of European data-centre operators now identify access to power as their single biggest operational challenge.

In fact, a growing proportion of planned facilities depend on electricity infrastructure that does not yet exist and may take years to approve or construct. Some projects will secure power because they are backed by hyperscalers with enormous financial and political leverage. Smaller speculative campuses may not, and parts of the industry are at risk of creating “dark data centres,” echoing the dark-fibre overbuild of the dotcom era.

According to the International Energy Agency, electricity demand from data centres rose by 17% in 2025, with AI-focused facilities growing even faster and significantly outpacing broader global electricity demand growth.

In the UK, reports suggest that around 140 data centres are currently waiting for grid connections, with projected demand estimated at roughly 50GW. For perspective, Great Britain’s electricity demand typically sits at around 30GW, including power supplied through interconnectors. This is forcing a difficult trade-off between rapid AI infrastructure buildout and environmental commitments; over 100 new data centres are planning to use natural gas for power.

Nowhere is the data-centre campus boom more stark than with Stratos, a hyperscale campus in Utah, associated with Shark Tank investor Kevin O’Leary. The project, once expected to approach twice the size of Manhattan, and consume more than double the state’s electricity use, was scaled back after intense public backlash.

Elsewhere, several hyperscale developments have been delayed or reconsidered because of power constraints and lack of infrastructure.

But technical and economic limitations are only part of the story. Opposition from local residents is now becoming a major barrier.

And is it any wonder? Massive data-centre campuses are creating sweltering “heat islands” around them, contaminating water supplies, and pushing up levels of noise pollution for those living nearby.

And what are the local residents getting in exchange for putting up with this? Far fewer jobs than hyperscalers claim.

Google notably dropped a proposed development in Abilene, Texas after local opposition intensified. Meanwhile, residents of Monterey Park, California successfully overturned the construction of a 250,000 square foot campus. Even OpenAI’s highly publicised Stargate programme is now facing funding and deployment challenges as the company attempts to assemble the eye-watering $500 billion required for full deployment.

The original edge-computing narrative is becoming more complicated. AI economics strongly favour large, centralised compute clusters located near abundant and reliable power sources, potentially weakening the case for widely distributed edge-cloud architectures that many operators previously viewed as strategically important.

However, that does not mean edge infrastructure disappears; edge data centres are more likely to evolve into smaller, geographically distributed facilities positioned closer to factories, cities, transport systems and telecom networks. These deployments may face less political resistance than giant hyperscale campuses, but they also introduce significantly greater operational complexity and cost.

The likely outcome is not a simple shift from centralised cloud to distributed edge, but a more selective edge model: centralised clusters for training and large-scale workloads, with distributed edge infrastructure used where latency, sovereignty, resilience or local data processing justify the added complexity.

Some operators are already beginning to reposition telecom infrastructure as AI infrastructure; MTN Group recently outlined plans to transform parts of its tower network into a distributed AI inference grid, deploying GPU-based compute infrastructure directly at tower and edge locations. This points less to a replacement for centralised AI training clusters than to a more distributed model for inference and localised AI services.

Recent geopolitical instability has also highlighted how vulnerable digital infrastructure can become when it depends heavily on regional energy systems, fuel supply chains and transport corridors. Reports of attacks affecting data-centre infrastructure in Ukraine and the Middle East underline how cloud services can be exposed to physical conflict, even when the services themselves appear virtual.

For countries that have positioned themselves as attractive hosts of hyperscale data centres, this is a sharp reminder that low-cost power and land are only part of the equation. Political stability, energy security and physical resilience are becoming just as important.

Even where facilities themselves are not directly targeted, cloud operations remain highly dependent on reliable energy generation and international logistics. Disruptions to either can increase operating costs and threaten service continuity.

These developments strengthen the case for private and sovereign telecom infrastructure. Operators that maintain control over distributed facilities retain greater flexibility to prioritise emergency traffic, meet lawful-intercept obligations, and preserve service continuity during periods of instability. In that sense, the renewed interest in private cloud reflects one of telecoms’ oldest assumptions: some workloads are simply too critical to place entirely within someone else’s infrastructure footprint.

Rather than replacing public cloud, however, the industry is moving towards hybrid models that combine the strengths of both approaches. Public cloud continues to offer scalability, innovation and rapid service deployment, while private cloud remains essential for network control, regulatory compliance, sovereignty and resilience. The competitive advantage will depend on how effectively operators orchestrate across both environments.

Data-centre infrastructure is therefore no longer a separate cloud-sector issue for telecom operators. It is beginning to compete directly with networks for power, sites, fibre routes and executive attention. The real strategic assets are not simply network coverage or cloud capacity, but deployable infrastructure: powered, connected, resilient sites that can support the next generation of digital services.

As a result, OSS platforms are gradually evolving from network management systems into broader infrastructure orchestration platforms. Operators may require systems capable of understanding not only network topology and service provisioning, but also power availability, cooling constraints, generator capacity, electricity pricing and grid limitations.

This also elevates the strategic importance of inventory management. For years, telecom inventory systems were viewed as necessary but relatively unglamorous operational tools. But in an AI-driven infrastructure market, operators cannot monetise, protect or optimise assets they cannot accurately see. Detailed knowledge of fibre routes, exchanges, ducts, rooftop rights, substations, backup power systems and available land becomes commercially critical.

Operators with fragmented or inaccurate infrastructure records may find themselves at a serious disadvantage as hyperscalers search aggressively for deployable sites and available capacity.

Enterprise telecom services are also likely to become more infrastructure-centric. Instead of simply selling connectivity, operators may offer integrated AI environments tied to geography, sovereignty, resilience, latency guarantees and power availability.

This places pressure on traditional BSS/OSS models, many of which were designed around relatively stable products and predictable service chains. AI infrastructure is considerably more fluid, requiring real-time coordination between networking, compute, facilities and energy systems. However, static or siloed catalogue models may struggle to support that level of orchestration. Operators will need product, service and resource catalogues that can interact more dynamically with inventory, assurance, charging and infrastructure orchestration.

Meanwhile, hyperscalers already operate highly integrated software stacks that combine networking, compute orchestration, telemetry, automation and consumption-based billing at enormous scale. Many telecom platforms, in contrast, remain fragmented and operationally slow.

Operators continue to improve power-usage effectiveness (PUE) and water-usage effectiveness (WUE), but efficiency gains alone are unlikely to offset rising demand. Improvements in model design and hardware could reduce the amount of power needed per task, but lower computing costs tend to increase overall usage because new applications emerge as technology becomes cheaper and more accessible.

Nevertheless, a full transformation of telecom operators into integrated AI infrastructure platforms remains unlikely. Telcos historically excel at operating critical infrastructure but move more slowly in software and platform transitions. Hyperscalers possess the opposite strengths.

What appears more likely is a tighter convergence between telecom, cloud and energy operations, alongside a growing emphasis on infrastructure-centric monetisation. Hyperscaler operating models will have a growing influence on telecom architecture, while AI operations are layered on top of existing systems rather than replacing them outright.

The cloud has not stopped being strategic, scalable or essential. But it is no longer detached from the constraints of the physical world. Power, sites, cooling, fibre and resilience are becoming operational and commercial variables in their own right. The result is a growing recognition that the digital economy remains fundamentally dependent on physical infrastructure.

Over time, telecom operators may evolve into orchestrators of fibre, power, edge facilities and strategic infrastructure assets, while hyperscalers continue to operate the dominant software and AI platforms that sit above them. The telcos that succeed in the AI era won’t be those that build the biggest models, but those that understand their infrastructure best, orchestrate it effectively, and monetise it intelligently.