Tariffs and oil are changing AI economics

Training an artificial intelligence model runs on two inputs: chips and electricity. In April 2026, both got more expensive at once. (whitehouse.gov) (eia.gov) The White House said on January 14, 2026 that semiconductor imports threaten national security and said it may impose “significant tariffs” on semiconductors, chipmaking equipment and related products after a Section 232 investigation. U.S. Customs and Border Protection said in a February 2026 tariff overview that new tariffs are being applied under both Section 232 and the International Emergency Economic Powers Act. (whitehouse.gov) (cbp.gov) The U.S. Trade Representative also published a December 2025 notice on semiconductors from China that set an initial tariff level at 0 percent starting December 23, 2025, with a higher rate scheduled for June 23, 2027. That leaves importers facing a mix of current duties, threatened duties and changing country-specific rules. (ustr.gov) (congress.gov) The energy side moved just as fast. The U.S. Energy Information Administration said on April 7 that Brent crude averaged $103 a barrel in March 2026 and could peak at $115 in the second quarter after disruption tied to the Strait of Hormuz. (eia.gov 1) (eia.gov 2) Oil does not power most data centers directly, but it feeds into broader power and cooling costs through fuel markets, transport and grid prices. The International Energy Agency said in 2025 that artificial intelligence workloads are pushing data-center electricity demand sharply higher, while the Energy Information Administration said in January 2026 that U.S. power demand is heading for its strongest four-year growth since 2000, fueled by data centers. (iea.org) (eia.gov) That changes the math for startups more than for the biggest cloud companies. A single Nvidia H100 graphics processing unit typically costs about $25,000 to $40,000 to buy in 2026 pricing guides, and an eight-chip server can run well above $200,000 before networking, racks and cooling. (docs.jarvislabs.ai) (gmicloud.ai) When hardware prices jump, small companies usually cannot spread the hit across many products or lock in long-term supply deals. The International Energy Agency said data centers used about 460 terawatt-hours of electricity in 2024 and projected that figure would rise to more than 1,000 terawatt-hours by 2030 in its base case, making power procurement a larger line item for artificial intelligence operators. (iea.org 1) (iea.org 2) The tariff case is still contested. The White House argues import dependence threatens national security and domestic production capacity, while trade analysts at the Congressional Research Service describe a year of tariff truces, framework deals and ongoing negotiations that left businesses navigating policy changes across 2025. (whitehouse.gov) (congress.gov) The immediate result is not that artificial intelligence stops getting built. It is that founders now have to budget for geopolitics the same way they budget for servers: as a recurring cost that can change between one quarter and the next. (cbp.gov) (eia.gov)