For the last three years, every AI headline followed the same script: a new model drops, benchmarks get smashed, commentators argue about whether it’s “true” intelligence. That script is running out of runway. The real constraint on how much AI the world can actually run has stopped being cleverness. It’s electricity.
This isn’t speculation from outside critics. It’s showing up in the infrastructure decisions AI companies are making with their own money, which tends to be a far more honest signal than any keynote.
Chips Convert Power. They Don’t Create It.
When pressed on why labs don’t simply manufacture more chips to solve the compute crunch, one OpenAI executive gave an answer that cuts through a lot of noise: more chips are useless without more electrons to run them. A GPU sitting in a warehouse produces zero intelligence. It only becomes useful once it’s plugged into a grid that can actually feed it. That single distinction, compute versus power, is quietly reordering the entire industry’s priorities.
The Scale Involved Makes the Point on Its Own.
OpenAI’s five-year cloud agreement with Oracle assumes 4.5 gigawatts of power running continuously, roughly enough electricity for four million homes. Its Stargate site already runs at 1.2 gigawatts. Meta’s Louisiana facility is projected to cross 2 gigawatts. For context, training GPT-3 back in 2020 consumed a total of about 1,300 megawatt-hours across the entire run. Frontier labs are now planning individual sites that would dwarf that number many times over, and doing it as a baseline, not a stretch goal.
Zoom out further and the trend gets harder to write off as one company’s ambition. Forecasters surveyed by the Forecasting Research Institute put the median estimate at 7.4% of total US electricity consumption going toward training or running AI systems by 2030. Global data centre electricity use, sitting around 415 terawatt-hours in 2024, is projected to roughly double to 945 terawatt-hours by 2030, with AI-specific workloads driving most of that growth. That’s not a sustainability footnote tucked into a corporate report. It’s a resource constraint that will decide which companies can even attempt to compete at the frontier.
Why Power is a Harder Moat than any Model
Here’s the part that gets buried under the bigger, flashier numbers: a model architecture can be reverse-engineered, replicated, or open-sourced within months of release. A gigawatt-scale power contract cannot. Securing that kind of energy commitment takes years of negotiation, capital, and regulatory approval, the same way building a new chip fabrication plant typically takes three to five years from ground-breaking to output. Neither can be rushed with a bigger funding round.
That’s why data centre siting decisions increasingly follow watts, not talent. A facility doesn’t go where the best engineers want to live anymore. It goes wherever a region can guarantee firm, cheap power and a grid connection that won’t take two years of utility approval to secure. The old logic of clustering near research hubs is quietly losing out to the logic of clustering near cheap hydro, gas, or nuclear capacity. It’s an unglamorous variable, but it now shapes company strategy as much as any research roadmap does.
The New Scoreboard: Tokens Per Watt, Not Benchmark Scores
The efficiency race happening underneath all this is arguably the more interesting story. Serving a single H100-class GPU node under real inference load draws around 10 kilowatts, and running it at low utilization for single requests burns almost as much power as running it at high utilization with smarter batching, just for a fraction of the output. Engineering teams have found that simply batching requests more efficiently can lift tokens generated per watt by 5 to 10 times on the same hardware, with no new chips involved at all.
Context length adds its own tax. Independent benchmarking has shown that tokens-per-watt roughly halves every time the context window doubles, which means every “give the model more context to work with” product decision is quietly also an energy decision. That’s a tradeoff most product teams outside the infrastructure world have probably never had to think about, and it’s about to become unavoidable.
Chip design is bending toward the same pressure. Custom inference silicon built in-house at OpenAI reportedly runs at around 700 watts per chip, compared to 1,200 to 1,400 watts for comparable off-the-shelf alternatives. That’s not a raw performance play. It’s a direct response to the fact that watts saved now matter as much as flops gained, because watts are the scarcer resource of the two.
None of this means model research has stopped mattering. It means it’s stopped being the whole story. The companies best positioned for the next few years aren’t necessarily the ones with the most elegant architecture. They’re the ones who locked in power capacity and efficient silicon early, before energy access became the industry’s actual bottleneck instead of a footnote in a sustainability slide.
For smaller players and product teams building on top of these models, the takeaway is more practical than philosophical: efficiency is no longer a nice-to-have optimization, it’s the constraint that will decide what’s even possible to ship at scale. The next meaningful leap in AI capability might not come from a research paper at all. It might come from an energy engineer figuring out how to squeeze more useful output out of the same watt.
