Are AI models running out of power? | The Economist

Demand Constraints and Corporate Responses

• AI firms are currently throttling tool access to manage overwhelming demand.
• Anthropic altered service terms to discourage capacity usage during peak times.
• OpenAI suspended Sora, its video generation tool, to reallocate scarce computing resources toward more profitable ventures.
• Inference (user query processing) consumes significant processing power, creating a direct link between user growth and hardware strain.

Capital Expenditure and Infrastructure Investment

• Five major US cloud providers (including Amazon, Meta, and Microsoft) plan to spend approximately $700 billion on AI data centers this year.
• Well-funded model developers like Anthropic and OpenAI are also heavily investing in securing external computing resources.
• Industry strategy focuses on aggressive "build-out" of data centers, though physical construction faces delays due to local opposition regarding electricity, land, and water usage.

Hardware Supply Chain Shortages

• The tech industry faces a critical shortage of basic network components (transformers and switches), with lead times extending between three to five years.
• GPU manufacturing cannot keep pace with the demand required to equip new data centers.
• The hardware supply chain is characterized by long construction timelines (2–4 years for excess capacity) which fundamentally conflict with the rapid, months-long software improvement cycles.
• Many companies are resorting to using chips that are two to three years old due to severe scarcity of modern hardware.

Semiconductor Monopolies and Choke Points

• Nvidia controls over two-thirds of the world's AI processing power, leaving its current chip inventory effectively sold out.
• TSMC is the sole major manufacturer fabricating most AI chips, with capital expenditures increasing by $60 billion this year.
• TSMC's current expansion is described by OpenAI CEO Sam Altman as insufficient to meet market needs.
• Elon Musk announced "TeraFab," a proposed project to build a single fabrication plant with capacity exceeding all current global plants by 2030, estimated to require $5–13 trillion in capital.

Market Impact and Future Outlook

• Analysts project that the supply crunch could force AI firms to raise prices, reversing the previous trend of inference costs dropping by half every six months.
• Rising costs may slow AI adoption and pressure firms to curb "reckless spending."
• The disconnect between software agility and hardware constraints threatens to slow the overall pace of AI development.