The six largest hyperscale cloud operators are on course to spend more than $600 billion on infrastructure in 2026, a 36 percent increase over 2025, as they race to build capacity for AI workloads. The constraint holding back further acceleration is not capital or demand. It is power.
As of mid-2025, more than 36 data center projects representing $162 billion in investment were blocked or significantly delayed, largely due to power availability limits and local opposition. The global hyperscale pipeline stands at 770 facilities at various stages of planning, construction, or fit-out. Synergy Research Group projects that total hyperscale capacity will double in just over 12 quarters.
Each of the major operators is navigating the constraint differently. AWS is investing in new regional infrastructure across Saudi Arabia ($5.3 billion), Germany (EUR 7.8 billion committed through 2040), Chile, and New Zealand, with geographic diversification serving both latency reduction and power access goals. Microsoft's Fairwater AI campuses, beginning with Atlanta and followed by Wisconsin, deploy closed-loop liquid cooling systems that eliminate operational water consumption -- a design shift driven partly by growing regulatory and community resistance to water-intensive cooling in drought-prone regions.
Meta is targeting more than 10 gigawatts of total capacity by end of 2026, operating 30 data centers globally. Google is moving toward 24-7 carbon-free energy commitments across its data center fleet, though clean power availability at the scale required remains a structural bottleneck rather than a near-term solution.
The nuclear option is gaining traction as a long-term answer. Meta announced up to 6.6 gigawatts of nuclear energy projects in January. Three Mile Island was restarted to power Microsoft data centers. An Iowa nuclear plant is being evaluated as the next restart candidate, driven explicitly by AI data center demand. Dell'Oro Group projects data center capex reaching $1.7 trillion by 2030, with sovereign AI programs joining hyperscalers as a major demand source.
What the power bottleneck reveals is a structural tension in the AI infrastructure buildout: demand for compute is scaling faster than the grid can accommodate, and the most capital-efficient solution -- buying more GPUs -- is constrained by how much electricity those GPUs can draw. As NVIDIA unveiled the Vera Rubin platform this week with claims of 35x higher inference throughput per megawatt, the efficiency metric is becoming as strategically important as raw performance. The operators who solve the power problem first gain a durable infrastructure advantage that cannot be replicated simply by ordering more hardware.
Sources: Data Center Knowledge, Dell'Oro Group, Meta
--
By the Control Plane Editorial Team