Software margins split on usage caps, downshifting and copyability, not pass-through — resolves 2027-08-02

AI has added a model bill to software companies, and the market is arguing over which companies that cost will crush and which can carry it.

The answer shapes how investors pick software stocks and how software companies price, including whether to pass AI costs on to customers.

The original hypothesis, from the Gooaye podcast, drew the line at whether a company can pass costs through; after checking it, SecondSource draws the line with three variables and finds no company whose margins expanded because of pass-through.

SecondSource’s reading: the AI split in software is real, but the line is not whether a company can pass token costs (the unit models bill by) on to customers. It is three variables.

First, whether usage has a ceiling: an AI feature that generates once per click, or an agent (AI that runs long tasks on its own) left running with no cap. The capped group (Zoom, Duolingo, Adobe) has gross margins of 70% to 89%; the uncapped group (Cursor, Replit, Windsurf) is reported to run from negative to about 40%.

Second, room to downshift: whether work can move from the most expensive flagship model to a cheaper model that is good enough. Third, copyability: whether customers can build the same thing themselves with AI.

The real margin path is to absorb the cost first and win it back through optimization; SecondSource found no case of margins expanding because of pass-through.

Gooaye podcast (Hsieh Meng-kung), SoundOn

SemiAnalysis
“Anthropic's ARR has exploded from $9B to over $44B…”

SemiAnalysis
“AWS margins inflected this past quarter driven…”

AI Snake Oil / Normal Technology (Arvind Narayanan & Akash Kapur)
“we believe that model inference contains an…”

arXiv (Satoshi Matsuoka)
“a depreciation conveyor delivers newly amortized…”

20VC (Harry Stebbings) — YouTube
“I think we saw we we saw something bizarre…”

Checks: any public software company crediting margin gains to AI pass-through billing, not cost cuts; Zoom's annual non-GAAP margin under 75.8% or Duolingo's under 71%; OpenAI or Anthropic cutting live flagship list prices 50%+ without a new premium tier; a Fortune 500 replacing core SaaS with in-house AI; Benchmarkit's 2027 software margin median under 78%.

What would overturn it

Each check maps to a failure: pass-through expansion returning voids the no-evidence finding; the margin thresholds breaking voids the claim that they can absorb it; a 50%+ flagship cut devalues downshift freedom and revives the utility-bill view; in-house replacement widens the self-build scenario; a low median voids tiering and favors uniform compression.

Aug 2, 2027

Near-term checks: whether Sonnet 5 moves to $3/$15 when its promotion ends on 2026-09-01; Palantir’s second-quarter results on 2026-08-03 (level of disclosure and gross margin); and ICONIQ’s mid-2026 snapshot of measured AI-product gross margins against the 52-53% forecast.