Power, data centers, and bottlenecks
AI is already facing a power crisis—this isn’t a distant future.
The consensus he cited: by 2027, AI chips will require at least 15 gigawatts of power that will be in short supply.
AI chip output grows by about 40–50% per year. Power growth outside China is about 10–20%. A faster curve will overwhelm a slower one.
Companies like Google and Anthropic are already renting computing resources from SpaceX because SpaceX is building its own power plants. This is the only way they can quickly bring capacity online.
China has sufficient electricity, but a GPU export ban blocks the latest chip supply. The real constraint is power growth outside of China.
Opportunities for other countries: build lots of power infrastructure, host AI data centers, and tax/charge them reasonable fees.
AI as a growth engine
Countries should embrace new technology rather than cling to the past.
His rough estimate: digital AI alone could raise the global economy by 20–30%, roughly 20–30 trillion dollars per year.
By the end of next year, AI should be able to complete any digital task—any task that doesn’t require manually shaping atoms.
Software prediction: in about 12–18 months, AI writing software will reach “Stockfish level.” Humans won’t be able to compete, like the chess engine on phones has already been able to beat Magnus Carlsen.
Same time window: AI becomes exceptionally good at all engineering forms and any digital task, possibly at the same level.
He also promoted 𝕏 as the place for almost all serious AI discussions, saying it’s his way of keeping track of the field day to day.
Robotics and physical AI
Physical technologies always take longer than digital ones. Software can be copied instantly. Hardware requires a massive global supply chain and involves moving lots of atoms.
The practicality of humanoid robots is the product of three things: AI software × on-board AI chips × electromechanical dexterity (especially with two hands). All three are improving exponentially.
Once robots start making more robots, growth becomes recursive: slow at first, then explosive.
A 10-year forecast (he calls it a conservative estimate): more than 1 billion humanoid robots, each with productivity about 5x that of a human. That means these robots will surpass the total productivity of all humans combined.
He believes that physical infrastructure is a domain where the economy could grow 10x or more, not just 20–30%.
How do countries actually build new technologies?
New things should default to being legal, not illegal. Strict regulation (he cites the EU as an example) won’t kill progress, but it will slow it down a lot.
Startups are like saplings in a forest. Most governments over-support existing towering trees (existing giants) and under-support small trees.
Big companies can reach political leaders; startups can’t. Policy should intentionally tilt in favor of younger companies.
