Could the race to build out generative AI’s infrastructure outpace the ability to finance and power it? That was the underlying tension when OpenAI’s CFO Sarah Friar likened today’s AI boom to the advent of electricity-arguing that the industry has “turned on the lights for the first time” and must now build the grid to keep them on. The problem is that this grid is no metaphor. It is a vast network of semiconductor supply chains, hyperscale data centers, and gigawatts of energy capacity-each requiring billions in capital, years of lead time, and careful risk management.
The scale of the build-out is staggering. McKinsey estimates $5.2 trillion in global AI-related data center investment will be needed by 2030, with 156 gigawatts of AI-specific capacity. That includes about $3.1 trillion for chips and computing hardware, $1.3 trillion for power generation and cooling systems, and $800 billion for land, construction, and site development. In the United States alone, hyperscalers like Amazon, Google, Microsoft, and Meta are on track to spend more than $350 billion in 2025 on AI and cloud infrastructure, pushing total tech capex toward half a trillion dollars.
The semiconductor layer is the first choke point. AI workloads depend on accelerators like NVIDIA’s GB200 NVL72, which delivers a 25-fold leap in energy efficiency and 30 times higher inference throughput compared with the H100 for large language models. Memory vendors such as SK Hynix and Micron are supplying high-bandwidth memory optimized for these GPUs, while hyperscalers are developing custom silicon Google’s TPUs, Amazon’s Trainium to match particular workloads. But chip fabrication is capital-intensive and slow to scale, and any disruption in supply can stall entire data center projects.
The second constraint is energy. AI training and inference are among the most resource-intensive computing tasks ever devised. In 2024, U.S. data centers consumed 4.4% of national electricity, a share projected to triple by 2028, with AI alone potentially consuming as much electricity annually as 22% of U.S. households. The carbon intensity of data center electricity is on average 48% higher than the U.S. grid overall, due to clustering in fossil-fuel-heavy regions like Virginia. Even with aggressive renewable build-outs, intermittent generation cannot yet meet the 24/7 demand of AI clusters, making nuclear and natural gas essential for near-term reliability.
The physical realities of power infrastructure create an imbalance with AI’s speed of deployment. Projects for transmission take seven years from permitting to completion, while it takes just 18 months to build data centers. In some regions, interconnection queues are longer than construction timelines, encouraging developers to pursue on-site generation and microgrids. Dominion Energy, whose Northern Virginia territory hosts the world’s largest concentration of data centers, is investing $5 billion in transmission upgrades to support 7.5 GW of new capacity-a signal of the scale of grid reinforcement needed.
Financial modeling for these projects is complex. Long-term offtake agreements, in which tenants pay for capacity whether they use it or not, are standard for hyperscale facilities, reducing vacancy risk. But uncertainty over AI’s technology cycles how long a GPU generation remains competitive, how quickly inference overtakes training, and whether efficiency gains will offset demand makes underwriting difficult. As KKR’s infrastructure team notes, “unit economics are more important than hype,” with returns after power and capital costs being the critical metric.
Friar’s mention of a federal “backstop” for chip financing hit a political nerve. She later said OpenAI is not looking for bailouts, but the prospect of US government guarantees for AI infrastructure does raise many industrial policy questions. David Sacks, an adviser on AI to President Trump, emphatically says, “there will be no federal bailout for AI,” although he favors streamlining permitting and increasing power generation to speed build-outs without raising residential rates. The dispute highlights a strategic tension: the US wants to win the AI race without tying up taxpayer dollars in speculative ventures. Sustainability introduces yet another layer of complexity.
As Mahmut Kandemir, among other experts, underlines, without greener hardware, optimized models, and renewable integration, AI’s environmental footprint spanning electricity, water, and e-waste will rise unabated. Solutions are beginning to emerge in the form of liquid cooling to reduce thermal loads, neuromorphic chips for reduced power consumption, and aligning compute workloads with periods of peak renewable availability. But these require coordination across chipmakers, data center operators, utilities, and policymakers an ecosystem not known for moving in lockstep.
To investors and executives, the path forward is a calculus of speed, scale, and risk. Investing too much risks stranded assets in case demand slows or technology shifts; investing too little risks ceding competitive advantage in what many see as the backbone of the next industrial revolution. The lights are on-but keeping them on will require unprecedented alignment between capital markets, engineering innovation, and energy infrastructure.
