Three stories from the same week, all about the same thing: who pays for the electricity behind AI, and who has to get smarter about using it.

Congress: 417 to 3

On Wednesday the U.S. House passed the Ratepayer Protection Act by a vote of 417–3. The bill sets standards that state regulators can adopt for data centers drawing 100 megawatts or more: operators can be required to cover the cost of the new generation, transmission lines and grid infrastructure their demand creates, instead of spreading it across everyone's utility bill. The Senate hasn't acted yet, and states can decline to adopt the standards — but a 417–3 margin on anything in this Congress tells you where public sentiment sits. Nobody wants to subsidize a hyperscaler's power bill.

Industry: the data center as a good citizen

The same week, Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance, with Anthropic and utilities including AES, Constellation, National Grid and NRG joining as launch partners. The premise: a data center doesn't have to be a static, maximum-draw load. If it can pause non-critical work and shift compute when the grid is stressed, it becomes a controllable resource — and the alliance estimates that flexibility could let roughly 100 gigawatts of additional data centers connect to grids that would otherwise say no.

The proof point lands this year: Emerald AI says that later in 2026, with NVIDIA and Digital Realty, it will switch on a nearly 100-megawatt power-flexible AI facility in Virginia — designed to show that a data center can be dialed up and down like a thermostat.

Silicon: 7x more tokens per megawatt

SemiAnalysis's early testing of NVIDIA's Vera Rubin NVL72 rack shows up to 7x better token throughput per megawatt than Blackwell — on pre-release software, and NVIDIA's own marketing number is higher still. Take the exact multiple with a grain of salt; the direction is the point. The industry's key metric has quietly shifted from "tokens per second" to "tokens per watt." That's what happens when the constraint stops being chips and starts being substations.

The pattern

Policy is pricing the externality. Industry is building the control loop. Silicon is raising the efficiency ceiling. All three are the same message: the era of treating compute as free is over, and the teams that win the next few years will be the ones that treat energy — and its proxy, latency — as a first-class design constraint.

Why a web studio cares

Because the same discipline applies one layer up, on the products we build. Every unnecessary model call, every 4 MB hero image, every render-blocking script is a small version of the data center that draws maximum load whether it needs to or not. We've been running performance budgets on client sites for years for the user's sake; the energy story is the same practice with a bigger reason.

  • Cache aggressively; call models sparingly. The cheapest inference is the one you didn't run.
  • Ship the smallest asset that does the job. WebP, lazy loading, no autoplay video nobody asked for.
  • Measure work per watt, not just work per second. If NVIDIA can change its headline metric, so can your dashboard.

The House vote will get the headlines. The Virginia facility is the one to watch. If a data center can learn to behave, the software on top of it has no excuse.