The AI boom just got a price tag that makes the moon landing look like a rounding error.

The Summary

The Signal

PwC just quantified what everyone building in AI already feels: the infrastructure requirements are staggering. $31.6 trillion over 25 years works out to $1.26 trillion annually, more than the entire global semiconductor market today. This is not incremental cloud expansion. This is rewiring the world's computational backbone.

The "no precedent in history" framing from PwC matters. The Manhattan Project cost $28 billion in today's dollars. The Apollo program, $257 billion. The Interstate Highway System, $500 billion. The entire buildout of the commercial internet through 2020, maybe $4 trillion if you're generous with accounting. AI infrastructure will eclipse all of them combined, then keep going.

"This is rewiring the world's computational backbone, not upgrading it."

What drives this number is training costs colliding with inference at scale. Training GPT-4 class models already runs into hundreds of millions. But training is one-time cost. Inference is forever, compounding with every user, every agent, every API call. When ChatGPT hit 100 million users in two months, OpenAI's compute costs became a meme. Now imagine billions of AI agents running 24/7, each one querying models thousands of times per day.

The spending breaks into three waves:

  • 2026-2030: Foundation model training clusters and early inference infrastructure
  • 2031-2040: Mass deployment as agents become standard business tools, edge inference scales
  • 2041-2050: The agent economy runs on autopilot, compute becomes utilities-level infrastructure

The Implication

Follow the capital. $31.6 trillion flows somewhere, and wherever it lands shapes the next 25 years. Nvidia and the chip makers get the headlines, but the real value capture happens in power infrastructure, cooling systems, real estate near fiber backbones, and the specialized construction firms that know how to build at this scale and speed.

For builders, this is clarifying. If you're launching an AI company and your unit economics assume cheap inference forever, you're modeling fantasy. Compute costs will stay high or go higher as demand outpaces supply. The winners will be those who build compute-efficient architectures from day one, not those who optimize later.

Sources

Bloomberg Tech