The AI bill has left the earnings call and arrived at the meter. The Belfer Center says U.S. data-center demand could rise from 176 terawatt hours in 2023, about 4.4 percent of national electricity use, to 325 to 580 terawatt hours by 2028. [1]
That is the infrastructure side of the paper's May 30 account of SpaceX filing language meeting Musk's compute lease caveat. The earlier story asked whether compute obligations in AI filings could be read as durable evidence. Today's story follows the wire from contract to grid: every compute promise eventually asks a utility, a regulator, and a household to pay for capacity. [1]
The public argument often treats AI capex as a corporate race among Amazon, Microsoft, Google, Meta, Anthropic, OpenAI, and xAI. Belfer makes the quieter point: data-center expansion changes the electric system around it. In some regions, AI-driven demand is already outpacing available capacity, pushing companies toward delayed projects, direct power contracts, and on-site natural-gas generation. [1]
That is how a model release becomes a household story. EESI reports that utilities faced at least 700 gigawatts of data-center connection requests in 2025, more than the 477 gigawatts of electricity the United States consumed in all of 2023. Many of those projects will never be built. The problem is that utilities and regulators have to plan before they know which requests are real. [2]
Planning is expensive even when demand is uncertain. EESI says utilities requested more than $29 billion in rate increases in the first half of 2025, double the first half of 2024, and that those increases were expected to affect 40 million customers. The article ties the data-center boom to transmission, transformers, generation, and the affordability crisis already visible in power bills. [2]
Belfer names the risk in technocratic terms: stranded costs. If utilities build wires, gas plants, substations, and transformers for demand that does not materialize, consumers can be left paying for infrastructure that served a speculative buildout. If utilities underbuild, data centers can strain reliability and force emergency workarounds. Neither outcome belongs in a press-release sentence about AI leadership. [1]
The reliability story is not imaginary. Belfer cites a July 2024 voltage fluctuation in northern Virginia that disconnected 60 data centers at once and created a 1,500-megawatt power surplus requiring emergency adjustments. The event shows why data centers are not just big customers. They are grid actors large enough to change operating conditions when they trip offline. [1]
The fossil-fuel story is just as concrete. EESI says natural-gas plants are being built and coal retirements delayed to meet data-center demand, while some coal-plant refurbishments can cost up to $1.3 billion and operating costs rose 28 percent from 2021 to 2024. AI capex can therefore reappear as higher bills, higher emissions, and a slower retirement schedule for old plants. [2]
This is where X and mainstream coverage often talk past each other. Online boosters frame data centers as national destiny and a necessary cost of AI sovereignty. Local critics frame them as noise, water, land, and ratepayer burdens. The policy story is less theatrical and more dangerous: regulators must decide who pays for infrastructure before anyone can prove which AI workloads will be profitable enough to justify it. [1] [2]
The supportable conclusion is not that data centers should stop, nor that every bill increase is an AI tax. The sources say something narrower. Data-center power demand is now large enough that utilities, households, and regulators inherit risks once treated as private AI-company strategy. The capex race has become a public-utility case. [1] [2]
-- HENDRIK VAN DER BERG, Brussels