Tokens Per Watt: Why Your Context Window Is a Power Decision | HackerNoon

On an H100, tokens per watt drops 12x between 4K and 64K context. Agents live at the fat end of that curve. The fix comes from semiconductor architecture.