AI Hyperscalers Eyeing Nuclear and SMR Supply Chain; Data Center Power Constraints Accelerate

As AI model training and inference workloads continue to explode, the limiting factor has shifted from chip supply to energy supply. Major hyperscalers (Microsoft, Google, Meta, Amazon, NVIDIA) are now contemplating direct investments in the nuclear fuel cycle and small modular reactor (SMR) supply chains to guarantee reliable, low-carbon baseload power for data centers. This represents a structural shift in AI capex allocation: instead of buying more chips, companies are now buying the power plants that run them.

NVIDIA's announcement of a $3.4 billion, five-year contract with IREN provides concrete evidence of this shift. The deal includes approximately $700 million in annual recurring revenue potential and signals that NVIDIA is not just a chip vendor but also a player in the infrastructure-as-a-service ecosystem. Similarly, Fervo Energy's $1.89 billion IPOInitial Public Offering - a company's first public sale of stock., which jumped 33% on first trading day, reflects investor appetite for companies that can unlock geothermal energy for hyperscalers. The message from capital markets is clear: whoever controls the energy supply wins the AI race.

This dynamic will reshape the entire semiconductor supply chain. Companies like AVGO (Broadcom), which supplies advanced packaging and infrastructure chips, stand to benefit. Substrates, interconnects, and data center networking equipment all become more valuable as power density becomes the constraining factor. The conversation has shifted from "can we build more chips?" to "can we build enough power plants to run them?" A new bottleneck has emerged, but it is occurring at a higher level of abstraction (energy infrastructure) rather than at the chip level.

The bull case is that AI capex remains the strongest secular driver of M&A, venture funding, and equity valuations. Companies willing to lock in power supply via long-term nuclear or renewable contracts will win market share, and that competitive advantage is durable. The bear case is that SMR and geothermal technologies are nascent and carryIncome earned from holding a position over time. execution risk; delays in nuclear permitting or scaling could bottleneck AI deployments by 2027-2028. Additionally, the capex intensity of the AI infrastructure race is now being questioned: if companies must spend $1 trillion to build sufficient data centers and power plants, the ROI calculation becomes murkier, and growth stocks could suffer multiple compression.