AI scenarios to 2030

Four paths for AI

Who gets paid for AI depends on the path it takes into business, and how good it gets matters most for how much the world spends.

Huge innovation

$7bn a year in 2024. $193bn a year today.

World spending on AI models and AI work, a year, estimated.

Who wins? Start at the bank

The bank's decision

How should the bank answer this customer?

Hi, I was charged twice for the same card payment.
AI is typing

The bank gets

The bank gives up

Each choice can keep customer data private; they differ in who the bank trusts. The $100 is where AI spending goes in 2030, if AI keeps improving. By AI makers I mean the labs that build the models.

What it means for one AI maker

Take Anthropic. It earns about 45% of what the AI labs earn today. If it keeps that share, its revenue by 2030 depends on the path: about $94bn a year if AI gets cheap, and about $449bn if AI does the job.

Today's 45% share held to 2030. Revenue is the average of the model's two paths for what the AI labs earn: one tied to the jobs AI replaces, one following the labs' own plans. If AI keeps improving, with no crash. The model was built in Claude sessions, and Anthropic makes Claude.

How good does AI get?

How good AI gets is the second question. It matters most for how much the world spends and whether the data centres fill up.

Three roads to 2030: AI stalls, keeps improving or takes off.

Investors may pull back in 2027-29, whatever AI itself does. Markets fall, and the money for data centres dries up.

Speed also tilts who wins: if AI stalls, open models catch up and “AI gets cheap” gains; if it takes off, “AI does the job” gains.

Jobs lost by 2030

How many jobs AI takes by 2030 depends mostly on how good it gets. If AI keeps improving, the US loses about 3.3 million jobs and the EU about 3.0 million. If it stalls, about half that. If it takes off, about 5.2 million in the US and 5.0 million in the EU.

USEU

One number per road: the average over the four paths. Gross jobs displaced: the US from June 2025, the EU across all 208 million employed.

How this was made

My own scenario research. The scenarios and the $100 splits are my judgement, and nobody has measured the splits. The numbers come from my model, and the data-centre history from Epoch AI. I built the model in Claude sessions, and Anthropic makes Claude. Numbers frozen on 16 November 2026.

Not in the model: a chip shock, a halt or a serious AI incident, new rules on what AI may do, and AI running on your own phone or laptop. Any of them could change every road.

Where every number comes from