Open any AI podcast from the last six months and you will hear Carlota Perez quoted within fifteen minutes. We're in the installation phase. Exuberant, capital-intensive, productive in parts. Be patient. The deployment phase will repay everyone.
It is the consoling line of 2026. It is also a misreading.
What Perez actually said
Her framework describes how each technological revolution unfolds. Not what it owes anyone. Financial capital overshoots. Infrastructure gets built ahead of demand. The bubble bursts. A fraction of what was built becomes the foundation of the next era. Deployment begins.
That is the pattern. She describes it as historically observed, not as a contract anyone signed. Nothing in her work says the people paying for the installation phase are the same people who collect the deployment benefits. That part gets read into the framework by people who want it to mean something else.
The historical record reads the opposite way.
Read the railway ledger
Between 1850 and 1871 the federal government transferred 129 million acres of the public domain to private railroad companies. Add 51 million from the states. The loans, $64.6 million at six percent, got mostly repaid. So on paper, the public got its money back. End of debate, according to a certain kind of economist.
The trouble is that the public domain wasn't a federal asset in any honest accounting. It was land taken from Native American tribes, often along the very corridors the railroads wanted to lay track. The route maps and the dispossession maps are the same maps.
Stanford, Hill, Harriman: deployment phase. The settlers who bought cheap farms a decade later: smaller deployment phase, real but second-tier. The earlier inhabitants: no deployment phase at all. The framework treats them as a footnote because the framework is about capital flows, not about ledgers of who paid what.
Perez described a pattern. She did not endorse it.
The 2026 invoice
Now AI. Private capex on AI in 2026 is forecast to exceed $700 billion. Most of that builds compute. But compute needs electricity, and electricity needs grid. Goldman estimates $720 billion of grid investment through 2030 to keep up with data center load. That grid spend doesn't come out of OpenAI's budget. It comes out of yours.
In the PJM region, the territory from Illinois to North Carolina, 67 million residents are watching residential bills rise around 15% in 2026. Maryland households pay about $18 more per month. Ohio about $16. The increase lands hardest on lower-income households, because electricity is a bigger share of their spending. (None of those households are getting equity in the model labs.)
This is the part that should make a serious person uncomfortable. The compute that runs Anthropic, OpenAI, Google DeepMind is being subsidized by a Maryland nurse who has never used Claude.
The honest version
The installation phase isn't a temporary tax on the future. It is a permanent transfer that happens to be amortized across two phases of a chart. Perez describes the pattern. She doesn't endorse it. And when the deployment phase arrives, whatever it ends up looking like, it doesn't reverse the transfer that already happened. The bill doesn't get refunded. It stops appearing on next quarter's invoice.
If you are reading this on a screen in Ohio: you are not the next Carnegie. You are the public domain.
— Pan