Who Pays for the AI Boom?

The main bottleneck for AI replacing white-collar work at scale is that the best models are large and expensive to run. Every step up in capability costs a lot more compute, both in training and in inference. You can distill or quantize a model into something smaller and cheaper, but you lose reliability, and reliability is exactly what you need before a model can replace an employee instead of assisting one.

At the same time, the models are getting good at research. In October, OpenAI released 372 results on open mathematical problems, produced with an average of three hours of ChatGPT Pro compute each. One of them targets part of Hilbert's sixteenth problem, which has been open since 1900. Mathematicians are still reviewing the results and some won't hold up, but even if only a fraction survive, a model is now doing original research on problems humans have left unsolved.

The labs use their best models to build the next generation. It's only a matter of time before that loop produces architectures that fundamentally change how models are built. It has happened before: in 2017 the transformer replaced recurrent networks such as the LSTM, which had been the go-to architecture for sequence modelling for 20 years. When the next shift comes, it will move very fast, because the compute to exploit it is already racked in the data centres.

Blue-collar work is next. My guess is that jobs like warehouse work, factory work and tyre fitting go first. Robots taking them over face two blockers: unit cost, and models that can handle the physical world, which requires training data from real robots. Given how robotics has progressed over the past year, cost is on its way to being solved, and every robot put into production generates more training data. It's a flywheel.

The money behind it

AI capex is now one of the largest private investment waves in US history. By most estimates, only the railroad boom of the late 1800s (peak years at around 5–6% of GDP) and residential investment before the financial crisis (6.6% of GDP in 2005) were larger. The telecom and fibre build-out of the dot-com era, around 1.2% of GDP, is already well behind.

What's different this time is concentration. The railroads were built by hundreds of companies; AI infrastructure is being built by a handful of hyperscalers and labs. The largest tech companies' capex from 2023 through their 2026 budgets comes to roughly $1.5–1.6 trillion, and consensus expects hyperscaler capex alone to stay at about 3% of US GDP every year from 2027 to 2029.

So ask yourself: how do the investors expect to get a return on that? Through new services that benefit everyone, or through cost savings, which in practice means headcount reductions and wage pressure?

We need laws and regulation that protect people, not just the return on capital.