AI infrastructure power constraints to throttle cloud growth through 2028

The immediate shift from chip scarcity to power scarcity means streaming and AI firms can no longer rely on instant infrastructure scaling. As AWS, Microsoft, and Google hit utility limits, the cost of high-performance compute will likely rise, favoring companies that adopt frugal architectures like retrieval-augmented generation over massive model fine-tuning. This resource crunch forces a move toward hybrid cloud strategies where workloads are placed based on regional power availability rather than just latency. For the streaming ecosystem, this necessitates a disciplined approach to AI-driven personalization and encoding pipelines to avoid project blockers. Watch for a surge in municipal permit filings and micro-grid investments as cloud providers attempt to bypass traditional utility bottlenecks.

The scramble for electricity to feed AI workloads has pushed hyperscalers into unprecedented energy deals. In early 2025, Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1, a nuclear facility that will supply 835 megawatts of carbon-free power to its data centers starting in 2028. Amazon Web Services has taken a similar path, with Amazon acquiring a nuclear-powered data center campus in Pennsylvania from Talen Energy for $650 million in March 2024, and Google has committed to purchasing power from small modular reactor developer Kairos Power, targeting its first reactor to come online by 2030. These moves signal that the three largest cloud providers now treat energy procurement as a strategic bottleneck on par with GPU supply.

Regulatory friction is compounding the capacity crunch. In July 2025, the Federal Energy Regulatory Commission rejected a proposed deal between Talen Energy and Amazon Web Services that would have allowed AWS to draw power directly from the Susquehanna nuclear plant, citing concerns about cost-shifting to other ratepayers. That decision has forced hyperscalers to explore behind-the-meter arrangements and on-site generation, adding complexity and timeline risk to expansion plans. Meanwhile, PJM Interconnection, the largest US grid operator, warned in its 2025 capacity auction that reserve margins could fall below acceptable levels by 2027 as data center load growth outpaces new generation buildout. Municipal permitting delays for transmission upgrades further constrain where new facilities can be sited.

The technical response from cloud providers centers on efficiency gains at both the chip and architecture level. Google reported in its 2024 environmental report that its data center power usage effectiveness reached 1.09 on a trailing 12-month basis, among the lowest figures in the industry, while Microsoft has deployed liquid cooling systems across its AI-optimized server racks to manage thermal density. For streaming workloads specifically, the constraint favors inference-efficient approaches. Nvidia's H200 GPU delivers roughly 2x the inference throughput per watt compared to its A100 predecessor, meaning that encoding and personalization pipelines built on newer silicon can maintain output levels even as total available compute capacity tightens. Companies that adopt FlexSysAI workload orchestration platform and smaller fine-tuned models will face less exposure to the power ceiling than those relying on large-scale training runs.