"The Thing About More Tokens" — CJ Gustafson (Mostly Metrics)
Why this is in the vault
CJ's frame is the cleanest articulation I've seen of the operating-layer problem behind MAC: AI spend is starting to behave like a per-employee cost, not a project cost, and operators who don't tag it now will be flat-footed when the market starts asking "show me Rule of 40 net of AI." Maps directly to my COO posture and to MAC's "cost-routing as moat" pillar — if our pitch is "AI cost is a real P&L line, treat it like comp," CJ has already done the audience-prep work for us in startup CFO land.
⚠️ Sponsorship
Sponsored by Abacum (FP&A platform). Top-of-issue placement plus Summit recap and demo link. Abacum is in the FP&A automation space, so a piece arguing that finance teams should track AI cost per-employee is also a piece warming up the buyer for Abacum's product surface. Not a tactical conflict for our use (the frameworks stand independent of the sponsor), but worth surfacing: CJ has a structural incentive to push "more FP&A tooling needed" framing across the year.
Issue contents
The piece bundles a thesis essay, a benchmarking section ("Mostly Multiples" charts), and a sponsor block. Single-thread but multi-modal.
Core thesis
Token consumption is becoming the new "lines of code" — high-activity proxy that correlates with motion, not value. Top performers were already shipping before AI; their token bills go up but their throughput doesn't compound proportionally. CJ's recommendation: model AI spend at 20%+ of employee compensation, tag it by function and by individual at ingest, baseline before the optimization pressure shows up.
Frameworks
- AI cost = additive comp line. Budget like benefits: per-employee, with a function tag. Don't bury it in "infrastructure" or "R&D opex" where it disappears.
- Optimization-language signal. When earnings calls start saying "we're optimizing AI spend," the market interprets that as accountability time. Snowflake is CJ's referenced precedent for this language pattern.
- Revenue-per-employee and Rule of 40 are the real frames. Token count is vanity; RPE and R40 are what investors will ask about by end of fiscal year.
- Tag-then-optimize order. You can't optimize what you can't see. Get the per-person, per-function tagging in place before you start cutting.
- The "burning tokens" callback. CJ uses a Big-Short-style framing ("they're not confessing, they're bragging") to characterize teams that boast about high token consumption. Sharp re-frame.
Mapping against Ray Data Co
- Direct MAC validation. This is the third independent source (after Cole and the Practical-Data-Modeling cohort) treating AI cost as a finance-stack problem, not an infra problem. MAC's "cost-routing as the missing operating discipline" thesis is now defensible without us being the only voice.
- Sanity Check angle: the "tokens-as-LOC" comparison is a banger waiting to happen. Founder voice could land "everyone proud of token burn in 2026 is going to look like the 'I committed 10k LOC this week' person from 2014." That's a sub-2-paragraph SC piece with a clean re-frame.
- For the COO agent operationally. I should be tagging my own API spend per skill/function. Right now my Anthropic spend is one undifferentiated line. The fix is a wrapper that tags each Bash/MCP call with skill-name, then a monthly pulse against revenue-per-decision. Building this should be a Notion task.
- Contradicts our prior framing. I had been positioning AI cost as a unit-economics problem ("cost per task done"). CJ's framing is one level up: it's an HR-comp problem, because the unit doing the work is the human + AI bundle. That's a stronger frame for selling to CFOs. I should update MAC's pitch deck slide.
Related
- [[2026-04-30-backfill-discovery-not-boring]]
- [[2026-04-13-cole-100k-paid-newsletter-playbook]]
- [[../01-projects/mac/README]]