US senator floats data centre tax for AI disruption

Senator Mark Warner says entry-level job postings in the United States have fallen 35% since 2023, and he wants the data centres behind the AI boom to help pay for the disruption. Speaking at the Axios AI Summit, Warner told TechCrunch he is exploring a tax on data centres, arguing it is the “easiest place” to extract a “pound of flesh” from the industry. The proposal comes as lawmakers such as Bernie Sanders and Alexandria Ocasio-Cortez push for a federal moratorium on data centre development, citing local backlash over noise, pollution, water use and electricity prices.

The argument quickly runs into a measurement problem. “AI’s energy footprint” is not a single number; it is an accounting choice that depends on what you count, where you draw system boundaries, and which time window you use. Data centre operators commonly report Power Usage Effectiveness (PUE), the ratio of total facility energy to IT equipment energy, but PUE is a site-level efficiency metric, not an attribution tool for a particular model. A facility can post an impressive annual PUE while still drawing large peaks at the worst hours for the grid, or shifting heat and water burdens to nearby communities.

For AI workloads, the split between training and inference matters. Training runs are sporadic, can be scheduled, and often occur in a few specialised clusters; inference is continuous, latency-sensitive, and spreads across regions closer to users. That changes both the “load factor” (how steady demand is) and the marginal electricity mix that actually meets the extra demand. Marginal power is rarely the same as average grid power: at night it may be baseload; during peaks it may be gas turbines or imported electricity. Without granular metering—hourly consumption, location-specific grid intensity, and cooling mode—headline claims about “clean” AI can be true on paper while communities experience higher local prices and stressed distribution equipment.

Water is the second blind spot. Cooling systems range from air-cooled designs to evaporative cooling towers and hybrid setups; the same megawatt of compute can consume very different volumes depending on climate and design. Reporting “water used” without distinguishing between withdrawals and consumption, or without seasonal context, tells residents little about whether their summer restrictions will coincide with a facility’s hottest operating period.

Warner says he supports strict requirements to prevent data centres from passing water and power costs onto residents, and he points to localities such as Henrico County, Virginia, where data centre tax revenue has funded projects including affordable housing. But a tax regime that cannot distinguish between a lightly loaded facility and a high-utilisation AI cluster risks rewarding good reporting rather than good behaviour.

If the policy goal is to make communities whole, the first step is not a new levy but a shared spreadsheet: time-stamped electricity draw, cooling energy, and water consumption that can be audited by someone other than the operator.