Apr 2, 2026, 8:00 AM ETInterview, Fireside Chat, Conference Presentation

AI Exchanges: Power Problems?

Hyperscalers' combined capital and R&D budgets for 2026–2027 have increased by over $300 billion, driving a massive buildout of data center infrastructure.
Goldman Sachs has revised its 2023–2030 global AI and non-AI data center power demand growth forecast to 220% (up from a prior 175%).
This projected growth is comparable to adding the sixth-largest power-consuming nation to the global grid; previously, the sector was considered less energy-intensive during inference.
Demand growth is expected to outpace supply due to the emergence of "agentic" AI systems, which generate machine-to-machine traffic that is significantly more "token-heavy" than current human-AI interaction.
The sector currently remains in a "deeply supply-constrained" state, with no indication of oversupply or diminishing returns on capital similar to the post-2008 shale boom.

Goldman Sachs analyzes power demand dynamics through a framework of six factors: Pervasiveness, Productivity, Price, Policy, Parts, and People.
People (Labor): A shortage of skilled labor is identified as the primary constraint, requiring an estimated 500,000 new jobs in the U.S. alone (300,000 for power generation and 200,000 for grid transmission/distribution).
- A specific bottleneck exists in transmission and distribution, where electricians require four years of skilling; current U.S. energy apprentices (approx. 45,000) would need to increase by another 20,000–25,000 to meet 2030 targets.
Parts: Supply chain inertia is delaying the deployment of efficient natural gas combined-cycle generators, with availability not expected until 2029–2030.
Policy: Rising political and consumer resistance to rate hikes has led to state-level moratoria and legislation, though alignment is emerging where hyperscalers agree to "take or pay" contracts to isolate their costs from broader consumer grids.

Due to grid and labor delays, hyperscalers are increasingly adopting "behind-the-meter" power solutions, which are predominantly supplied by natural gas simple-cycle generators.
- These solutions are less efficient than combined-cycle plants but offer faster time-to-market to address immediate capacity constraints.
The projected 2030 power mix for data centers is estimated at roughly 60% thermal sources (natural gas) and 40% renewables (solar, wind, nuclear).
- Nuclear power is expected to play a more significant role in the 2030s, potentially through both new builds and re-mothballing existing plants, though large-scale deployment faces current hesitancy.
Utilities face distinct challenges compared to hyperscalers; while hyperscalers have strong balance sheets, utilities lack similar flexibility and must tap capital markets for project financing, creating a dependency on regulatory approval and investment viability.
The industry exhibits "super abundant chivalry" regarding nuclear deployment, where competitors delay projects hoping others will bear the initial technology risk and regulatory hurdles first.

Hyperscalers are redeploying approximately 87% of their 2026 operating cash flow plus R&D back into CapEx and R&D, a figure approaching the 120% peak seen during the shale boom.
- While financial flexibility is tightening, balance sheets remain robust with minimal net debt to EBITDA, and only a few outliers are nearing 100% cash flow utilization.
Goldman Sachs analysis indicates that shifting to green, reliable power sources (e.g., nuclear, solar + wind + battery) would cost hyperscalers approximately $40 more per megawatt-hour.
- Even if hyperscalers bore the full cost of this premium globally, it would only impact 2030 EBITDA by roughly 2.5% and reduce corporate return on invested capital by less than one percentage point.
Despite the higher costs, there is strong corporate alignment to adopt green solutions to mitigate political and community (NIMBY) friction, with many companies issuing public commitments to ring-fence data center costs from consumer rates.