|
The speed limit on artificial intelligence is measured in gigawatts, not teraflops
For three years the artificial-intelligence story has been told in silicon: which company makes the fastest chip, who can buy the most of them, how high the capital budgets climb. This week alone, Nvidia’s record quarter and Broadcom’s report tonight will be parsed for exactly those numbers. But there is a physical fact underneath all of it that the chip headlines skip. A processor is inert until it is connected to the grid, and the grid is where the AI race is now stuck. A single AI query can draw up to a thousand times the electricity of an ordinary web search, and a rack of AI chips pulls ten to twenty times what a normal server rack does.
The numbers describe a wall more than a ramp. US data-center electricity demand has climbed from about 23 gigawatts in 2023 to roughly 42 today, and the Department of Energy expects it to triple by 2028. Globally, data centers will consume close to a thousand terawatt-hours this year, roughly Japan’s entire usage. The catch is timing: a power plant and its transmission lines take the better part of a decade to build, while a data center can go up in eighteen months. Supply and demand are moving on completely different clocks.
The result is a shortage with a price tag. Goldman Sachs estimates the US will be short about 9 gigawatts of the power it needs this year, a gap widening to 45 gigawatts by 2028. In many regions the wait to connect a new project to the grid now runs five to eight years. In the PJM market that covers the mid-Atlantic, the price utilities pay to guarantee future capacity rose more than 800 percent in a single year. In Ohio, the utility AEP simply stopped accepting new data-center connections, because it could not promise the power.
The bill lands on your kitchen table. Here is where it stops being an industry story. When data centers bid for scarce power, everyone on the same grid pays the higher clearing price. Retail electricity prices are up about 2.3 percent over the past year, with data-center demand named as a leading cause, and Goldman calculates that AI power use is adding roughly a tenth of a percent to core inflation this year and next. That is small in isolation and meaningful in a year when the Fed is already struggling to pull inflation down. The reader who owns the AI boom through an index fund is, at the same time, helping pay for it at the meter, and through an inflation rate that stays stickier for longer.
Why the fix is slow, and who gains. The industry’s answer has been to go around the grid. Every large hyperscaler has now signed at least one nuclear power deal: Meta’s twenty-year, 2,600-megawatt agreement with Vistra, Microsoft’s contract to restart Three Mile Island, the revival of the Palisades plant in Michigan. In March, seven AI companies pledged to fund the grid upgrades their projects require rather than push them onto households. Those moves have quietly turned a handful of unglamorous power producers, names like Constellation, Vistra, and Talen, into infrastructure sitting at the center of the AI trade. The market has begun to notice, and the risk in chasing it is that power forecasts, like all demand forecasts, have been wrong before.
Where that leaves you. The takeaway here is a lens for the whole AI trade, rather than a single stock or sector to buy. The bull case rests on a buildout of almost unimaginable scale, and that buildout runs into a physical limit no budget or faster chip can remove. Power plants take years, and permission to build them takes longer. Reading the AI story only through chip earnings misses the variable most likely to set its actual pace. And the same shortage that could slow the boom is quietly raising the cost of living for everyone plugged into the same wires. None of this is investment advice.
Nvidia reported a record last week, Broadcom reports tonight, and the money keeps pouring in. The limit on how far this goes will be measured in gigawatts, and by that measure America is running short. The chips are ready; the wires are not.
|