Interview, Fireside Chat
Are AI models running out of power? | The Economist
- The AI technology stack faces a fundamental shortage as hardware demand outpaces supply, forcing companies to throttle access and reallocate scarce computing resources.
- Hardware capacity construction requires a timeframe of two to four years, creating a significant disconnect with software improvements that occur every few months.
- Lead times for essential equipment like transformers currently stretch between three to five years, while semiconductor manufacturers cannot produce GPUs in sufficient quantities.
- Five of the largest US cloud providers are spending close to 700 billion dollars this year to build data centers, yet TSMC's 60 billion dollar increase in capital expenditures may still be insufficient.
- Nvidia currently supplies over two-thirds of the world's AI processing power with sold-out chips, compelling some companies to utilize hardware that is two or three years old.
- Local opposition regarding electricity, land, and water usage is extending construction timelines, making the success of capacity expansion plans uncertain.
- Elon Musk plans to build a fabrication plant called TeraFab with an ambition to exceed all current combined capacities by 2030, though this would require between 5 to 13 trillion in capital expenditure and is considered unlikely to happen.
- A continued supply crunch poses a risk of price increases that could slow AI adoption, though some observers view the constraint as a necessary check on reckless spending.