A Simple Fix to the AI Energy Problem That Pays Households to Want More Data Centers...

I have been chewing on this idea from first principles for a while now, and I want to lay it out because I think it solves about four problems at once and the maths is clean... The set-up is the data centre electricity problem we are all watching. AI and robotics are going to need a vast amount of power over the coming decade, with every forecast getting revised higher every quarter as the build accelerates. Right now the cost of constructing the grid to deliver that power is landing quietly on household electricity bills. The Virginia ratepayer is paying more so OpenAI can train the next model, and that pressure is growing as the build accelerates. So here is the idea... You levy data centres an extra 10c per kWh on grid-pulled electricity. Simple to legislate, simple to administer. Then the move that does the heavy lifting. The tax collapses by 75% if the data centre builds its own power, so the rate falls to 2.5c per kWh on self-generated electricity. That single rule rewires the incentive set across the entire hyperscaler stack. Microsoft, Google, Amazon, Meta, xAI, Oracle and every hyperscaler downstream of them now has a strong tax-efficient reason to own their own generation co-located with the campuses. Different sites will suit different mixes: solar plus battery in the Sun Belt, paired gas turbines in the Midwest, SMRs at the Stargate-scale sites that justify the capex. The energy substrate gets built, financed by the people who use the power, with no federal subsidy money required. It is also extremely easy to measure and is almost impossible to game. Now the maths, which is the part that surprised me. A small levy on data centers... Today, US data centre electricity demand is running in the region of 250 to 300 TWh per year. At 10c per kWh on the grid-pulled share, with the rest collecting at the 2.5c rate, the levy pool is already in the order of $25bn a year. ...Gives a BIG rebate for Households and small business... Scoped to households and small business, that funds something in the range of a 15 to 25% reduction in electricity bills, depending on how tightly you target the relief. Now look forward, which is where it gets interesting. AI energy demand is on an exponential curve and the forecasts keep being revised up. Data centre electricity is expected to reach roughly 500 TWh in 2027, around 1,000 TWh by 2030, and 2,000 to 3,000 TWh by the middle of the next decade. The pool grows roughly in line with that, even as more share moves behind-the-meter (because the absolute volume is rising faster than the share-shift can offset it). By 2030 you are at $50 to $70bn a year, which gets the household bill reduction to around 50%. By the mid-2030s the pool is north of $150bn, which is enough to take the average household electricity bill to zero. Free residential power within roughly ten years is genuinely on the table here, and it scales precisely with how aggressive the AI build turns out to be. What I really like about this is that everyone wins... Good for the hyperscalers... The hyperscalers actually want it. Their unit economics on compute are compressing inside a Wright's Law curve so steep that the levy is a rounding number against the next capability cycle. They absorb the cost without feeling it. On top of that, they get political cover for the buildout and a tax-efficient on-ramp into owning their own energy infrastructure, which they were going to need to own anyway. The robotics layer benefits in the same shape, since robot production runs on industrial electricity and the 75% collapse rule pulls the same incentive into manufacturing siting. Amazing for the people... Households benefit immediately, and the benefit grows as AI demand scales. That is real cost-of-living relief landing in the same place that the costs are currently falling. And the new energy build accelerates rather than getting blocked. Solar, storage, gas turbines and nuclear all get pulled forward at the pace the AI substrate actually needs, without the federal subsidy fights and without the permitting battles that have stalled large-scale generation for the last two decades. It is a stunningly simple idea. The incentives line up across every party that matters, and the more AI demand grows the better the deal gets for everyone. I think it likely works and it appears that no one has suggested it yet...