Casey Handmer on Dwarkesh Patel: Why Solar Beats Gas for AI

China massively outproduces the U.S. in solar panels and other industrial goods, but the U.S. still has superior natural resources, automation capability, capital markets, and geopolitical security. If it chooses WWII‑style mobilization, it can rapidly stand up large‑scale domestic solar and energy infrastructure in a few years.

For large AI labs, electricity is a tiny fraction of total cost; GPUs dominate capex and the value of cognition produced per kWh is enormous. They could tolerate 10–100× higher electricity prices and still be profitable, so their real bottleneck is getting guaranteed megawatts when and where they need them.

Gas turbines and existing grid capacity can handle a gigawatt‑scale build‑out, but as demand rises to tens or hundreds of gigawatts, limits on gas supply, turbine manufacturing, and grid capacity will bite. Solar has a ~43% learning rate and is much easier to manufacture, making it the only plausible way to “firehose” energy at massive AI data centers by the late 2020s and 2030s.

Transmission lines are slow, litigious, union‑ and regulation‑heavy, and underbuilt relative to projected needs. Batteries can perform temporal arbitrage (shifting solar from noon to night) just as the grid performs spatial arbitrage, enabling large data centers and heavy industry to run on captive solar + batteries with only fiber connectivity to the wider world.

Handmer argues that NEPA and related environmental and permitting regimes make it far harder to build solar in the U.S. than to do more environmentally damaging things like paving land or parking leaking cars. Texas is outbuilding California on renewables largely because its regulatory environment is less hostile to deployment.