Korva is a scroll of the internet’s best ideas.

The fiber precedent

Fiber optic cable offers a precedent. Companies rushed to lay it, then fiber got ten times faster, and then ten times faster again, and the bandwidth problem went away. Dark fiber that people had bought was left sitting there.

AI and the data centers behind it are likely to see the same price and performance gains. Use will scale the way everyone expects, and technological breakthroughs can arrive alongside it. Meanwhile, some data centers can't get their power turned on.

The borrowing behind it

The spending behind them is enormous. Google, Meta, and other market leaders are borrowing hundreds of millions to billions of dollars, on top of a private credit problem that already exists. They spend their cash flow on capital expenditure and then borrow beyond it, including 50-year bonds. Committing tens of billions of dollars for 10 or 20 years out is something nobody can predict well, and a hundred billion dollars put to work at OpenAI has to come back as earnings, not just revenue.

The video case

Video is the best case against that view. World models and video take an order of magnitude more tokens, and if the skeptics turn out to be wrong about data centers, video is the likely reason.

Why did so much fiber sit unused?

In the late 1990s, telecom firms laid cable on forecasts of internet traffic doubling every few months. A technique called wavelength division multiplexing then let each strand carry many colors of light at once, multiplying capacity without any new cable. Strands with no light on them became known as dark fiber. Researcher Andrew Odlyzko's analysis found that internet traffic in those years was roughly doubling each year.

Why would anyone buy a 50-year bond?

Pension funds and life insurers owe money decades from now, so they look for assets that pay on a similar schedule. A long bond locks in a fixed interest payment for the whole period, which matches those obligations. Most companies borrow for 10 to 30 years, which makes a 50-year issue unusual. Austria went further in 2017 and sold a 100-year government bond.

Why does video take so many tokens?

A language model reads text as tokens, which are chunks of roughly a word each. A video is a rapid sequence of images, and each image is cut into small patches that also become tokens. Even a few seconds of footage holds dozens of frames, so the count climbs far past what a page of text needs. OpenAI's 2024 technical report on its Sora video model described turning video into spacetime patches that work like tokens.

Why can't some data centers get power?

A data center needs a connection to the electric grid, and utilities must build substations and transmission lines to supply it. That work often takes years, longer than constructing the building itself. Transformers and gas turbines also have long manufacturing waits. The International Energy Agency's 2025 Energy and AI report projected that data center electricity demand would roughly double by 2030.

More from All-In with Chamath, Jason, Sacks & Friedberg

Tech & AIA toddler knows what AI doesn'tA 2-year-old in a high chair knows that pushing a sippy cup over the edge brings mom running. An AI shown the same video has no clue.
Enjoyed it?