The data center boom: ‘All the cheap power is gone’

Episode Summary

Executive Summary: The episode examines how the AI/data-center boom is colliding with grid limits, rate design, and clean-firm supply. Jigar Shah and Catherine Hamilton argue the demand surge is real despite DeepSeek, but the key question is who pays for new infrastructure and whether utilities can plan fast enough. In a second conversation, Peter Fried says the industry is entering a narrow window to deploy flexible grid upgrades, batteries, geothermal, and nuclear while data-center load ramps sharply.

Main Topics: AI and data-center load growth is still massive despite DeepSeek (Priority: 5/5): The hosts argue that even if models become more efficient, demand for compute and associated infrastructure remains enormous. Jigar stresses that the grid is already facing credible new load requests and that the AI boom is only one part of broader electrification. Who pays for grid and gas infrastructure upgrades (Priority: 5/5): A central theme is cost allocation. The speakers debate whether data centers should pay incremental costs directly or whether utilities should socialize upgrades across all customers, with concern that ordinary ratepayers could subsidize trillion-dollar tech companies. The economics of clean firm power and the 'everything is $100/MWh' thesis (Priority: 5/5): Jigar repeatedly argues that once storage, backup, and interconnection costs are included, many seemingly cheap solutions converge around roughly the same high cost. This frames the real competition as a choice among expensive options, not cheap ones. Data-center siting, size, and grid constraints (Priority: 4/5): Catherine and Peter discuss why mega-campuses are hard to build and why 200 MW blocks remain the dominant practical unit. Local land, water, transmission, gas, and permitting constraints shape where projects can go and how large they can be. Clean firm technologies: nuclear, geothermal, batteries, and grid-enhancing technologies (Priority: 4/5): The conversation highlights the technology stack needed to serve load: batteries for flexibility, grid-enhancing technologies for squeezing more out of existing infrastructure, geothermal where geography allows, and nuclear for longer-term clean firm supply. Regulatory reform and demand-response market design (Priority: 4/5): The guests emphasize that current tariffs, interconnection processes, and market rules are not built for speculative, fast-moving load. They call for clearer price signals, demand-response participation, and new financing or partnership models. Trump-era policy, energy dominance, and national security framing (Priority: 3/5): The episode places the discussion in the early Trump administration, suggesting AI infrastructure will be framed as a national security and China-competition issue. That could accelerate permitting and fossil buildout, but also create openings for nuclear and geothermal.

Key Arguments: DeepSeek may lower compute per model, but overall demand still rises because cheaper models can expand deployment and inference usage, leaving a large net load growth problem. The grid is not designed for 1,000 MW-scale loads in one place; even 200 MW data-center blocks create significant local constraints and require serious planning. The real bottleneck is not just generation but who pays for transmission, gas pipelines, batteries, and backup needed to serve a small number of extremely large customers. Many apparently low-cost energy options become expensive once storage and backup are included, so the market should stop assuming there are cheap solutions left. Utilities and regulators should use clearer cost-causation rules so data centers pay for the infrastructure they trigger rather than shifting costs to households. Demand response, flexible compute, and co-located batteries can reduce grid stress, but they are most effective when there is a transparent price signal. Nuclear, especially large-scale reactors, may fit AI/data-center load well because both need high uptime and firm power; geothermal is attractive but geography-limited. The current window to build new clean firm capacity is roughly 2027 to the early 2030s, so near-term grid optimization and longer-term generation buildout must proceed in parallel. Tech companies will likely continue procuring clean energy, but they may become more flexible about gas if the political and business environment rewards speed over purity.

Data Points: Stargate project pledge: $500 billion - Trump, OpenAI, Oracle, and SoftBank announced a massive AI infrastructure and power buildout. Global data-center investment outlook: $1.3 trillion over five years - The episode cites expected tech spending on data centers globally. Credible new load by 2030 or earlier: 25,000 MW - Jigar Shah says roughly this amount of new load is signing contracts now. Projected electricity demand from data centers by 2030: ~350 TWh - The discussion says load roughly doubles again by 2030 after rising during the Biden years. Data-center load growth during Biden administration: ~85 TWh to ~170 TWh - Stephen frames recent growth as a doubling over four years. Potential upper-end demand forecast: 50 GW - Catherine cites McKinsey’s high-end estimate for data-center demand. Natural gas plant installed cost: $1,700 to $2,000 per kW - Jigar uses this to argue new gas is very expensive. Battery storage cost: ~$800 per kW for four-hour storage - Used to argue batteries can be cheaper than new gas capacity for similar service. Natural gas turbine project cost: $2,000 per kW installed - Jigar says Louisiana gas for Meta/Entergy is among the most expensive options. Henry Hub gas price move: About $2.50/MMBtu to $3.50-$4/MMBtu - Jigar notes early Trump-era gas price volatility. Natural gas generation share: 42% to 62% of the grid - Jigar warns that moving to more gas would increase volatility and risk. Renewable deployment by tech companies: Over 100 GW - Peter Fried says tech companies have already deployed massive variable renewable portfolios with storage. Possible additional nuclear capacity through uprates: ~8 GW - Peter identifies uprates as a major near-term source of extra nuclear supply. Battery and data-center timing window: 2027 to early 2030s - Peter frames this as the period when load growth hits before new clean firm supply arrives. Oversubscribed nuclear funding round: $800 million - Jigar says bids for initial nuclear grant funding were oversubscribed by projects tied to data-center load. Load-block size: 200 MW - Discussed as the common practical data-center building block. Potential megacampus size: 1 to 5 GW - The conversation contrasts huge campuses with more realistic 200 MW blocks. Natural gas backup used in some projects: 8 to 12 hours of battery storage plus gas backup - Jigar says even curtailed renewable sites require expensive storage and backup to support load. Data-center/building plan horizon: 3 to 4 years - Peter says data centers typically take this long from deployment to meaningful load online. Power grid claim: 100 hours to 200 hours of annual constraints - Jigar uses this as the scale of reliability/peaking issues data centers must ride through. Clean transition tariff example: 100% of incremental cost - Peter describes Google’s Nevada approach with Fervo geothermal.

Pivotal Quotes: ""It's $100 a megawatt hour, to be crystal clear."" — Jigar Shah: Repeated refrain to argue that once backup, storage, and transmission are included, many power solutions cost roughly the same. ""The wall is a tiny little hump that we're now jumping over."" — Peter Fried: Peter describes the scale of today’s AI/data-center demand surge compared with Meta’s earlier power-planning challenges. ""I think that we will find ourselves in a moment... where many things that got kicked off in earnest in 2024 will begin breaking ground in the beginning of 2025."" — Peter Fried: On the likely timing of data-center deployments and why the industry has a limited window to act.

Implications: Utilities, regulators, and tech firms need clearer cost allocation and faster planning now. Expect more demand-response, batteries, gas, and some nuclear/geothermal bets—but also more rate pressure and policy fights over who funds the AI buildout.