Refuse to act on the AI spend number until it is decomposed - one-time hardware out, recurring API in - then treat the remainder as usage discipline rather than a spend cap

Situation

Steve Wallace raised rising AI costs at the end of the Aug 18 Engineering Weekly Sync. Peter did not accept the number as presented. He said the AI costs finance is counting include one-time purchases like the giant chassis just bought from NVIDIA for the 200s, and: I do not want to get hung up on one-time costs. I care way more about the recurring cloud costs and making sure people are being smart there. He set a meeting with finance for Thursday in LA to decompose it, and named what the response will be once the number is clean: model-tier discipline (teaching people not to do every single thing with Fable and Opus - Sonnet is still there for a reason and still really good at some things), possibly monthly per-person budgets, and a quality bar on AI-assisted work. Nathan raised that copy-pasting a prompt in and the answer out adds no value; Peter closed it with: we should be holding each other accountable and not accepting work at that bar.

Reasoning

Peter will not let a metric drive a decision until he is satisfied the metric measures the thing being decided about. A capital purchase and a per-token API bill are different problems with different remedies, and a blended figure would have produced the wrong remedy - most likely a blunt spending freeze. Once decomposed, the recurring number is a behaviour problem rather than a procurement one, so the levers he reaches for are teaching and accountability rather than a cap. The quality bar is the same move in a different register: the risk of cheap AI output is not the token cost, it is engineers shipping answers they do not understand, and that gets fixed by managers refusing to accept the work, not by a budget line.

Additional Context

Kelly Wall had asked Peter that same afternoon to set aside time Thursday in LA with her and Chris Baek about budget and costs, noting spend for AI and equipment has skyrocketed the last couple of months and they want a sanity check plus a cash projection for the rest of the year - the same conflation of equipment and AI that Peter separated in the sync. Steve Wallace then posted the Datadog AI Costs dashboard into #eng-management, where Nathan hit an empty-data view because it defaults to a one-hour window; Steve changed the default to 6 months.

Observed Evidence

Direct quotes from the Peter-recorded Engineering Weekly Sync transcript. Corroborated by Kelly Walls same-day request in the Baek mpdm, which bundles AI and equipment together in exactly the way Peter refused to reason about, and by the #eng-management dashboard thread that followed within the hour.

Matching Patterns

Confidence Breakdown

Reasoning Depth Analysis