When the Wall Street Journal reported in February 2024 that Sam Altman was in discussions with investors to raise up to $7 trillion to build new semiconductor manufacturing capacity, the reaction ranged from astonishment to ridicule. For context, $7 trillion is roughly equivalent to the GDP of Japan.
But look past the headline number, and the underlying argument is more interesting than the reaction suggested.
The constraint on AI development is not software. It’s chips. Training and running frontier AI models requires specialised semiconductors that are produced in tiny quantities relative to the demand for them. The world currently makes roughly 100,000 to 200,000 of Nvidia’s H100 chips per month. The mismatch between supply and demand is the central bottleneck in AI development.
Building new semiconductor fabs — the factories that manufacture chips — is extraordinarily expensive. A modern leading-edge fab costs $20 to $30 billion to construct and takes four to six years to build. The specialised equipment required has a multi-year backlog.
The discussions ultimately did not produce a formal fund of anywhere near that scale. But they catalysed significant conversations among sovereign wealth funds, Middle Eastern governments, and technology companies about how to finance the infrastructure needs of the AI era.
Leave a Reply