Sam Altman Wants $7 Trillion

Uses Sam Altman's $7-trillion ask as a lens on AI-scaling economics. Each GPT has cost ~30x the last (GPT-4 ~$100M; projecting GPT-6 ~$75B, GPT-7 ~$2T), so Scott breaks the cost into three inputs: compute (GPT-7 would need ~1/3 of all the world's computers), energy (GPT-7 ~ fifteen Three Gorges Dams, needed locally -- hence Altman's fusion interest), and training data (we're nearly out of text; synthetic data works for chess/math/games but nobody can do it well for text yet). Frames the whole thing as an exponential like a pandemic R: if each generation is exciting enough to fund, and smart enough to cheapen, the next one (R>1) you get fast takeoff; otherwise (R<1) it fizzles to a slow, decades-long, decentralized rollout. Reads the $7T as wanting the centralized/fast version -- which contradicts Altman's own past compute-overhang (go-gradual) safety argument, leaving safety-minded observers feeling betrayed.

Quality 72 · Strong. Strong. A clear, numerate explainer that makes AI-scaling limits legible (the compute/energy/data breakdown and the R>1/R<1 framing) and lands the Altman safety-contradiction. Bounded and back-of-envelope by design, so mid-Strong.

Claude’s paradigm shift 48 · Moderate. Moderate. Synthesises known scaling trends; the mild fresh handle is the pandemic-R framing of the funding/cheapening feedback loop.

Real-world impact 2 · Minor. Minor. A popular AI-scaling explainer within the AI-watching discourse; no material reach. Within-niche.