Compute is the new oil: megaclusters, GPUs and the infrastructure race
September 27, 2026 · 16 min read
Discussions of Stargate, xAI's clusters, NVIDIA and TSMC, together with Altman's and Musk's comments, all treat raw computing power as the deciding resource in AI. Understanding this bottleneck explains the energy demands, chip geopolitics and huge capital spending now shaping the industry.
The rules, numbers and steps worth keeping in front of you.
- Altman: compute is the currency of the future; demand for intelligence grows as its price falls.
- Altman: energy is the hardest part of supply; his answer is nuclear.
- Musk's bottleneck order: chips now, then step-down transformers, then electricity.
- Musk told utilities to prepare for demand to roughly triple.
- Musk: electrifying transport and heating will outweigh AI in the short term.
- Musk: brute force is temporary; a brain's thinking runs on under 10 watts.
- Patel: xAI's Memphis cluster has 200,000 GPUs; clusters of 500,000 to 700,000 are coming.
- Patel: data centers could reach 10% or more of US electricity by the end of the decade.
- Stargate's first phase is about 100 billion dollars in total cost, not 500 billion in hand.
- Leading-edge chips depend on TSMC and Taiwan; relocating would take a decade and a trillion dollars (Patel).
- Reasoning models make running AI costly: output tokens cost about four times input.
- Cost of a given level of intelligence fell about 1,200-fold in roughly three years, yet demand rose.
- Lambert: data quality is the top determinant of model quality.
- Amodei: frontier training costs heading from about 1 billion toward 100 billion dollars by 2027.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.