Key Takeaways

  • The United States runs on roughly 500 gigawatts of average electrical load (479 gigawatts last year), generating between $50 and $65 of GDP for every single watt of continuous power.
  • A single continuous gigawatt of electrical power translates to $60 billion to $65 billion of annual economic output across the US economy.
  • Frontier AI labs like OpenAI and Anthropic are generating annual recurring revenue per gigawatt that matches the broader US GDP-to-power ratio almost dollar for dollar.
  • The AI bull case relies on a stark asymmetry: hardware compute capacity has been tripling annually while top lab revenues have been multiplying tenfold each year.
  • Lab revenue functions like GDP value-added math, distributing cash across Nvidia, AWS, Oracle, power utilities, and specialized engineering payroll.

The Power Grid Anchor for Software Revenue

For decades, software companies pretended the physical world did not exist. Code shipped over fiber, server instances spun up on demand, and margins hovered near 80 percent. That illusion is over. Modern foundation models require dedicated electrical substations and industrial-scale energy contracts.

On the macro level, the math behind energy and production is remarkably stable. John Coogan points out that the total electrical load across the entire US grid sat at 479 gigawatts last year. Dividing total national economic output by that continuous baseline yields a clean ratio: “A single gigawatt generates about 60 to 65 billion dollars of GDP per continuously consumed gigawatt. So what's interesting is that the labs are tracking the frontier labs opening anthropic are tracking almost perfectly in line with that.”

When a frontier lab secures 100 megawatts or one full gigawatt of dedicated capacity, skeptics often question whether enterprise software spend can justify the capital expense. Yet at current monetization rates, OpenAI and Anthropic are extracting value from every electron at the exact same efficiency as the aggregate US economy.

The Real Margin Expansion: 3x Compute vs 10x Revenue

The central anxiety in artificial intelligence investing is energy exhaustion. If adding model capability requires exponential power increases, revenue growth will eventually hit a hard physical wall.

Coogan highlights why the macro data tells a different story: “And that's been the real like AI bullcase which is that compute has been tripling every year but revenue has been 10xing every year.”

As long as top-line revenue outpaces raw compute consumption by a factor of three, the unit economics hold up. The revenue captured by labs does not stay locked in a single bank account. Just like national GDP figures, it cascades through a massive supply chain. As Coogan explains, “GDP measures value added. So if you look at openanthropic revenue includes payments that ultimately flow to Nvidia, AWS, Oracle, electricity generation, employees etc.” Every dollar earned by a model developer validates capital expenditure for chip designers, cloud vendors, and utility operators.

Where the Model Breaks

This macro equation faces two major tests over the next three years. First, if algorithmic efficiency stalls, labs will need to buy raw megawatts simply to keep performance flat. That will drag their revenue-per-watt metric down toward commodity hosting levels.

Second, grid infrastructure has multi-year procurement delays. Even if enterprise demand supports a $60 billion return on a new gigawatt cluster, utility interconnect queues and transformer shortages can delay cluster activation by years. Software growth rates are now chained directly to industrial construction schedules.

What to Do With This

Audit your company's compute spend against gross revenue today. If your cloud inference bill is growing faster than your top-line expansion, your software is burning watts without capturing value. Re-architect your model routing so expensive frontier calls are reserved only for high-margin, revenue-generating tasks.