The cost savings didn’t come from a single finding but from institutional alignment and standardization and political continuity. So seems that infra costs are often organizational problems as much as engineering ones

If the growth is increasingly led by investment (like AI, infrastructure) more than by the consumption then the cycle dynamic could be very different from the previous expansions

Seems to be an energy security trade, when oil goes up and geopolitics heats dependence gets priced again quickly

The limit is changing from scaling parameters to scaling datas quality however compute is still the big constraint

When critical infra becomes centralized then the reliability expectations shift from good enough to essential. Its normal that outages stop being simple issues and become production blockers

This is mainly a cost structure because AI replaces variable labor costs with costs of GPUs, energy, infra etc... ; so at current prices running models at scale can easily be more expensive than human pay for the tasks with a low marginal value. If that doesn’t change as expected the economics of automation becomes less intersting.

This seems a classic capital cycle problem, with huge upfront investment, unclear pricing power and everyone scaling supply at once and so that combination usually doesn’t ends with great returns

It's a big upfront capex with unknown monetization timelines : it's fine when the capital is cheap but it's much harder when it isn’t

It seems that AI is stopping being just a compute problem and becomes now an energy problem Power infrastructure scales really slower than chips so this could become a real important constraint.

This has been true for a few years now. US is gated by power. China is gated by advanced chips.

The US needs solar and wind hardware manufactured by China.

China needs advanced chips manufactured by ... Taiwan.

Yes but TSMC themselves rely on software and fab equipment/materials made by Europe and US. TSMC also needs stuff from China, mind you.

It's all connected.

> It's all connected

This seems to escape a surprising number of people.

... even when unsubtly reminded. Maybe we need a website with a decent set of infographics that show the various inter-dependencies of the world economy, and an overlay showing the timings of these inter-dependencies.

The 1.6T number is nice but also eye-catching and what matters most is how few parameters are active in practice, that’s what brings the most of the efficiency