Perplexity burn ratio
The data
Perplexity reportedly spent ~164% of its revenue on compute (API calls to foundation models) in 2024.
Comparable points in the same loop:
- OpenAI: ~50% of revenue on inference + ~75% on training (Zitron, 2024).
- Many application-layer startups: >100% of revenue flowing to foundation-model API costs.
Source
Edward Zitron, Better Offline / wheresyoured.at, 2024 reporting. Surfaced through zitron-circular-economics and synthesized in bear-case-synthesis.
What it shows
A unit-economics inversion at the application layer: revenue does not cover the cost of goods sold (compute). The bear reading: this is structural, propped up by VC funding rather than viable customer economics. The bull reading: a snapshot, with compute costs falling rapidly per unit of capability and enterprise pricing not yet matured.
Note the tension with cursor-50b-valuation and windsurf-acquisition-battle: those exhibit the opposite signal at the same layer (workflow-embedded distribution moats with healthy ARR multiples). The wiki’s interpretation: thin wrappers burn through revenue; workflow-embedded platforms do not. See middle-layer-defensibility.
How to cite
Perplexity reportedly spent ~164% of revenue on foundation-model API costs in 2024 (Zitron, 2024) — an example of the inverted unit economics at the application layer.
Connections
- circular-ai-economy — the loop this ratio sits inside.
- distribution-moat — Perplexity’s strategic response to its burn: own the user relationship, force model providers to compete on price.
- middle-layer-defensibility — Perplexity as the thin-wrapper-vs-workflow-platform inflection case.
- cursor-50b-valuation — the counter-example at the same layer.
- zitron-circular-economics — primary source.