5D Library › Subscription Unit Economics, Net Revenue Retention and AI Inference Scaling Laws · 中文
Preface: The Arithmetic Behind the Ratios
Note
This is not investment or financial advice; please read this first. This book studies the mathematics of subscription businesses and of serving AI models. It does not recommend investing in, lending to or valuing any company, and its numbers describe models, not any real business. All data are synthetic, generated by the programs printed in the book, unless a published study is cited. Accelerator prices and model sizes are illustrative and are not quotes for any product.
Subscription businesses are run by ratios: lifetime value over acquisition cost, net revenue retention, payback months, gross margin. Each ratio compresses a dynamic process into one number, and each compression hides an assumption. “LTV = margin / churn” assumes every customer churns at the same rate. “NRR above 100 %” is quoted as a constant. “LTV / CAC above 3” says nothing about when the money comes back. And software built on large language models has a cost per use that the ratios of traditional software never had to carry.
This book writes each ratio back out as the process it compresses, runs the process, and finds where the ratio holds and where it misleads. Like the other books of 5D Library, it disproves first and then sets bounds, and it reports negative results — including the cases where the model the book favours is itself wrong.
What the book contains
Chapter 1 replaces “LTV = margin / churn” with heterogeneous cohort survival and asks what six months of data can project. Chapter 2 writes net revenue retention as a differential equation with a threshold at 100 %, then shows why high NRR does not last. Chapter 3 computes the cash trough that LTV / CAC hides. Chapter 4 computes what a token costs to serve — prefill, decode, batching, the KV cache — and why agents pay to re-read themselves. Chapter 5 sets heavy-tailed usage against a flat price and examines routing between models. Chapter 6 compares Metcalfe’s with the of Odlyzko and Tilly, adds congestion, and shows how growth data can “confirm” the wrong law. Chapter 7 reads Sima Qian’s “the greedy merchant makes three, the modest merchant five” and Jiran’s “let no money lie idle” beside the models. Chapter 8 gives a four-step check and a pre-registered price test.
The book neighbours several others in 5D Library and only recalls them where they overlap: Book 6 treated the economics of attention and preferential attachment; Book 11 treated AI as a mirror of its users’ assumptions; Book 19 treated non-ergodicity and compounding; Book 33 treated pre-registration and decision rules (Chapter 8 follows it); Book 48 treated marketing attribution and funnels with Shapley values — this book does not repeat that material and contains no attribution model.
Claims corrected after checking
- No single churn number gives the right lifetime value. With heterogeneous customers, first-month churn understates discounted value by up to −59 %, a mature base’s churn overstates it, and the common pitch-deck number — undiscounted, with a mature base’s churn — overstates it by +16 % to +48 % in the examples; for Beta parameter the expected lifetime is infinite (p01).
- Heterogeneity models project better — until their structure is wrong. From six months of data a single churn rate values a cohort at 1364 against a true 1966; the beta model is close when the truth is a mixture and overstates by +33 % when every customer’s hazard falls with tenure (p02, p03).
- NRR of 100 % is a threshold, not a grade. With constant acquisition, revenue saturates below it (a ceiling of 538 at 80 % in the example), grows linearly at it, and compounds above it (p04).
- High NRR does not last, with nothing wrong. A cohort’s NRR falls from 144.1 % in year one to 95.7 % in year five as expansion fills its ceiling; a steady company’s blended NRR falls from 134.6 % to 100.2 % as its base ages (p05).
- LTV / CAC hides the cash. With identical unit economics, the funding needed rises from 912,868 at no growth to 8,320,737 at 10 % monthly growth; and of three businesses, the one with the best slogan ratio (4.96) needs the most money and has the lowest discounted ratio, 1.80 (p06, p07).
- Output tokens, batching and context set AI margins. At batch 1 the 8B model’s output costs 3.012 dollars per million tokens against 0.018 for input; batching cuts it to 0.072; long context removes the batch (p08). Agent loops re-send their history, so output is only 6.1 % of the cost of a hundred-turn conversation without caching (p09).
- The mean user, not the median, sets the margin, and a flat price selects heavy users. At a log-sd of 2.5 the gross margin is −36.1 %; as light users churn faster, margin erodes from +81.3 % to +77.0 % with nothing else changing (p11).
- Metcalfe’s law overstates, and growth data can mimic it. Tenfold users give a hundredfold value under Metcalfe but 11.60-fold under the Zipf law; a revenue-on-users fit returns an exponent of 1.91 when value follows and monetisation grows 50 % a year (p12, p14).
- “Three” and “five” as margin against turnover is an interpretation. The Shiji does not explain why the modest merchant earns more; the book’s reading is set out as a model and a hypothesis (p15).
How to read the boxes
Marked boxes separate published evidence, runnable Python calculations (16 programs, standard library only, fixed random seeds, each run at build time with its real output printed beneath it), classical sources, notes, and the author’s hypotheses, each with the result that would show it to be wrong.
Disclosure
The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and founded and runs Asia Influencer X and Yafu Global, which use AI tools in influencer marketing, digital marketing and content production; he therefore has a commercial interest in the subject of this book. The Yafu AI and community platforms the author runs include subscription software and knowledge services. The book promotes no course, card deck, software, model provider, investor or fund, uses no data from the author’s companies, and accepts no sponsorship. The price test in Chapter 8 is a design for any subscription business to use, to be run and analysed by people independent of the pricing decision.
Fan, Y. (2026). Subscription Unit Economics, Net Revenue Retention and AI Inference Scaling Laws: Cohort Survival Mixtures, Negative-Churn Differential Equations, Token Compute Margins and Metcalfe Congestion Limits [订阅制单位经济学、净收入留存与人工智能推理标度律] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090164 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).