01Why an operator reads "Graviton"
Most operators skip OpenAI's deeply technical papers — including "Graviton." They shouldn't. The infrastructure papers are the leading indicator for what becomes affordable to deploy in production six to twelve months later.
"Graviton" sits squarely in that category. Its specific contribution is engineering. Its broader implication is about the cost-per-operation curve — and that curve is the single most important variable for whether a given workflow is economical to put a model on.
02The wrong unit of analysis
Most teams budget for AI in cost-per-token. That's the wrong unit. The right unit is cost-per-decision, or cost-per-resolved-ticket, or cost-per-document-shipped — and those numbers move very differently from token cost.
A workflow where the model takes ten attempts to get one decision right is ten times more expensive than the per-token cost suggests. As per-token cost drops, cost-per-decision drops faster, because the affordable retry-and-verify budget grows. That's the curve that actually decides whether automation is economical, and almost no one is graphing it.
03What gets unlocked at each price point
We keep an internal mapping of workflow types and the per-decision cost at which each becomes economical to fully automate, partially automate, or leave manual. As Graviton-class infrastructure improvements hit production, the threshold moves and the map updates.
The clients that captured the last cost drop were the ones whose workflows were ready to receive it. The ones who tried to redesign the workflow on the way down spent the cost difference on consulting and ended up where they would have been anyway. We expect the same pattern over the next eighteen months.
"Infrastructure papers are the leading indicator. Operating-model readiness is the lag indicator that decides who captures the gain."