06-reference

not boring return on tokens

2026-06-10·reference·source: Not Boring·by Packy McCormick & Markie Wagner

"Return on Tokens (ROT)" — @packym & @markiewagner

Why this is in the vault

This lands the same week Fable 5 moves off subscription plans toward usage credits (June 23, per [[2026-06-10-stratechery-fable-5-anthropic-alignment-ai-tiers]]), which makes "what return do tokens generate" RDCO's own live operating question, not an abstract one. The essay names the backlash cycle against tokenmaxxing, proposes ROT as the replacement metric, and argues an architecture thesis — agents compile, code runs — that maps almost one-to-one onto RDCO's harness-engineering bet and the skillify pattern. It is also unambiguously go-to-market collateral for the co-author's startup, which is why it gets the bias treatment below before the argument gets taken at face value.

⚠️ Sponsorship

Two distinct layers here.

Layer 1 — explicit paid sponsor. The issue carries a standard Not Boring sponsor block for Deel (global hiring / Employer of Record). Deel is unrelated to the essay's argument; standard ad placement, low contamination risk.

Layer 2 — structural bias (the bigger one). This is the recurring Not Boring pattern where Packy hands the keyboard to a portfolio-adjacent founder. Markie Wagner's company Poetic is named explicitly, and the entire framework (ROT, the Thinking-Doing Ratio, "AI is a compiler, not a runtime") is the setup for Poetic's pitch: the back third is a product section with customer names (AIG, SoFi, Chime), a CEO testimonial quote, and a hiring CTA. The intro came via Genius Ventures, who Packy says back Poetic alongside Founders Fund, Kleiner Perkins, and OpenAI. Packy does not disclose any Not Boring Capital position in Poetic anywhere in the text — could not verify a position, treat as messenger. Read the framework as a vendor's argument for why you need their category, with the tokenmaxxing backlash anecdotes selected to make that case.

The core argument

Tokenmaxxing was this cycle's dumb metric. Maximizing token spend became a proxy for AI progress — leaderboards, token commits, employees celebrated for agent usage — the same way railroad miles, eyeballs, and gross revenue were prior cycles' proxies. The incentive chain (market → boards → leaders → managers) rewarded spend, and skeptics were dismissed with "Skill Issue." Wagner argues the labs engineered this: consumption-based pricing plus agents that act like lab employees holding the customer's no-limit credit card.

The spell broke. The named evidence chain: Uber's CTO saying the company burned its 2026 Claude Code token budget by April; a consultant's client accidentally burning half a billion dollars; Amazon shutting down its AI leaderboard; Legora's CTO and Ramp's Veeral Patel ("Token Casino") going public; Alex Karp comparing tokenmaxxing to "a porn addiction"; even Sam Altman conceding waste is a "huge issue." These are real signals but curated to set up the pitch.

ROT is the replacement metric. Return on Tokens = (Value of Output − Cost of Tokens) / Cost of Tokens × 100. Two levers: create more value per token, or spend less per unit of value. Companies start with cost because it's measurable — hence the routing wave (frontier models for hard reasoning, cheap Chinese open-source models for bulk work, visible in OpenRouter rankings). Wagner's counter: routing is a good start, but code is cheaper than any model.

Agents are the wrong architecture for most economic work. Three structural reasons: (1) agents improvise, so they can't hold the nines of accuracy that repetitive economic work (fraud, underwriting) requires — 80% accuracy is "0% usable" there; (2) engineers don't know the work, because thousands of tacit rules live in operators' heads far from San Francisco; (3) no goals — agents set loose on vague instructions burn tokens without a hill to climb.

The central claim: AI is a compiler, not a runtime. Software has a thinking step (compile goals into code) and a doing step (run it cheaply, deterministically, forever). Agents should replace the software company, not the software — do the thinking rarely and expensively, then emit deterministic code that does the doing at near-zero token cost, returning only when the rules change. Wagner pegs the current Thinking-Doing Ratio at roughly 1000:1 and quips that Silicon Valley "built AI assuming work is mostly thinking" when work is mostly doing. Agent spend is CapEx, not OpEx. Chat/support is the carve-out where genuine improvisation is the job.

Then the pitch. Poetic operationalizes this: ex-Palantir engineers go on-site to extract tacit rules, AI compiles them into code, the code runs deterministically and regenerates when the world changes — claimed 100x less token usage at 99%+ quality. "We tokenminn to ROTmaxx." The closing vision: every business re-founded as self-testing, evolving software, with humans defining what good looks like.

Mapping against Ray Data Co

Strong map, with one important carve-out RDCO sits inside.

Related