Baseload vs Intermittent Power

The distinction between baseload (dispatchable, always-available) and intermittent (variable, dependent on weather) electricity is the central tension in the AI power story. It explains why a decade of falling solar and wind costs has not solved the hyperscaler power problem, and why nuclear is commanding a premium.

Baseload generation produces power continuously at high capacity factors: nuclear (92-95%), large hydro (40-60%), natural gas combined-cycle (50-85% when dispatched). It is "dispatchable" — the operator can command it to run. Intermittent generation — solar (25-35% capacity factor), wind (35-45%) — produces power only when the sun shines or wind blows. You cannot dispatch it on demand.

For AI datacenters, intermittency is unacceptable on two levels:

1. Operational: A multi-week AI training run drawing 100+ MW cannot tolerate interruptions. Power must be available 24/7, not just when renewable conditions are favorable.
2. Accounting: Hyperscalers committed to 100% renewable or carbon-free energy goals need to match their consumption hourly, not just annually. Buying a $35/MWh solar PPA that generates 0 MWh at 2 AM does not satisfy a 24/7 carbon-free energy (24/7 CFE) commitment.

Google pioneered 24/7 CFE matching — committing to match every hour of consumption with carbon-free generation in the same region in the same hour. This standard, if applied industry-wide, is essentially a nuclear/baseload mandate: the only low-carbon sources that can reliably generate in every hour are nuclear, hydro, and to a lesser extent geothermal.

Gas as the bridge: Natural gas combined-cycle (CCGT) is dispatchable and can back up intermittent renewables. GE Vernova's 100 GW gas turbine backlog reflects the reality that gas is the near-term solution for AI datacenter reliability while nuclear (existing fleet and future SMRs) scales up. Gas is not zero-carbon, but it is firm power.

Key Claims

• Nuclear capacity factor: 92-95% vs solar 25-35% and wind 35-45%. Evidence: strong (EIA)
• 24/7 CFE matching standard — Google's framework for hourly carbon-free energy matching — effectively requires baseload carbon-free generation. Evidence: strong (Google Energy)
• A baseload PPA specifies a fixed, continuous electricity volume — the seller guarantees delivery, using storage or supplementary generation to fill gaps. Evidence: strong (South Pole PPA Buyers Guide 2026)
• Gas turbines are the near-term AI power bridge: GE Vernova's $163B backlog ($100 GW gas turbine) reflects hyperscaler demand for firm, dispatchable generation while nuclear scales. Evidence: strong (GE Vernova Q1 2026)
• Hyperscalers represent 76% of all new corporate energy contracts (2024) — largely driven by 24/7 reliability requirements pushing toward nuclear/gas. Evidence: strong

The 24/7 CFE Hierarchy (Cleanest to Dirtiest, Baseload Only)

1. Existing nuclear (cheapest and cleanest baseload)
2. New nuclear restarts (higher cost, still carbon-free)
3. Large hydro (geographically constrained)
4. Enhanced geothermal (Fervo; nascent)
5. Fusion (2030s+ if commercialized)
6. Natural gas + CCS (expensive, CCS at scale unproven)
7. Natural gas CCGT (dispatchable; not carbon-free)

Intermittent sources (solar, wind) are cheap but must be paired with storage or overbuilt to approach 24/7 reliability — which drives cost up substantially.

Open Questions

• Can grid-scale long-duration energy storage (Form Energy's iron-air 100-hour batteries, etc.) make intermittent renewables baseload-equivalent? What is the LCOS threshold?
• Will hydrogen (green H2 from surplus renewables) serve as a baseload bridge fuel?
• As AI inference (vs training) grows, does the strictness of 24/7 reliability requirements soften?
• Do regulatory carbon markets eventually require hourly CFE matching from all large buyers?

Related Concepts

• Datacenter Power Demand — the demand that requires baseload
• Conventional Nuclear — the preferred baseload solution
• Grid Interconnect Queue — constraint on adding any new generation
• Power Purchase Agreements — the vehicle for securing baseload

Changelog

• 2026-06-15 — Initial compilation from web research (Google Energy, South Pole, GE Vernova Q1 2026, EIA)