Three words kept showing up this week, from three people who don't coordinate their talking points.
Chamath Palihapitiya on the All-In Podcast: "Energy equals intelligence."
Nick Turley, OpenAI's VP of Product, on BG2Pod: "GPUs are zero sum. An unlimited AI plan is like having an unlimited electricity plan."
Jensen Huang, in a blog post ahead of GTC: a five-layer cake of AI value creation. The bottom layer is energy.
An investor building data centers. The product chief of the biggest AI company on the planet. The CEO who makes the AI chips. Same conclusion, arrived at independently: the binding constraint on AI is not algorithms or talent or data. It is energy.
The cost of intelligence is now denominated in watts.
What a Gigawatt Buys
Chamath described building a one-gigawatt data center in Arizona. When he greenlit that project, he budgeted $4 or $5 billion. The number climbed to $10 billion. Then $15 billion. Then $20 billion. It is now approaching $50 billion.
Every gigawatt translates to roughly $10 billion in annual AI revenue. A five-year payback just to break even.
The demand backing those numbers is real. Anthropic grew from $1 billion to $14 billion in annual run rate in 14 months. OpenAI went from $2 billion to $20 billion in 24 months. Brad Gerstner pointed out that AI spending has crossed a threshold: it no longer competes with IT budgets. It competes with labor budgets. That is a fundamentally larger pool of money.
From inside OpenAI, Turley confirmed it: token consumption per user keeps climbing even as prices fall.
The $120 Billion Graveyard
Here is where the story turns.
About 40% of all protested data centers in America get canceled. In 2025 and 2026, roughly 12 gigawatts of capacity were killed by local opposition. At $10 billion per gigawatt in annual revenue, that is $120 billion in lost capacity per year.
Senator Bernie Sanders is calling for a full moratorium on data center construction. Pew found that more Americans believe data centers hurt their environment, energy costs, and quality of life than believe they help.
Stanford asked a simpler question: do you think AI will be more beneficial than harmful? In China, 80% said yes. In the United States, the number was in the low 30s.
NBC News polled Americans on AI as a category. AI ranked below ICE. Immigration and Customs Enforcement is more popular than artificial intelligence.
I want to be clear about something: the people protesting these projects are not irrational. They are responding to an industry that has done a terrible job explaining why any of this matters to them. If your neighbor tells you a giant facility is going up next door that will consume as much electricity as your entire city, raise your power bill, and truck in thousands of diesel generators during construction, you would protest too. I would.
The industry earned this backlash. The question is whether it can recover fast enough to matter.
Meanwhile, Texas has had zero data center cancellations and 150 gigawatts of capacity requests in the pipeline.
This Has Happened Before
Every generation of technology has had a moment where the physical layer became the bottleneck. In the 1990s, it was fiber. In the 2000s, it was data center power and cooling. In the 2010s, it was edge compute and bandwidth to mobile.
Each time, the companies that understood the physical layer won. The ones that treated infrastructure as someone else's problem lost.
We are in that moment again. The models are extraordinary. The research is moving at a pace nobody predicted. But none of it works if you cannot power it, cool it, and connect it.
Energy equals intelligence. The country that builds the most intelligence infrastructure wins. The one that lets local politics and disinformation kill its buildout loses.
Every energy policy is now an AI policy. Every grid upgrade is an intelligence upgrade. Every data center permit is a vote on whether your country competes or falls behind.
The infrastructure has always been the real game. The rest of the industry is just starting to notice.
John Engates writes about agentic AI, infrastructure, and the history of the Internet at exagentica.ai.