5D Library › Books › 5DL-B-058

订阅制单位经济学、净收入留存与人工智能推理标度律

队列生存混合模型、负流失微分方程、Token 算力毛利与网络效应拥堵边界

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

  • 范雅福 Fan Yafu
  • Version 1.0.0 · 2026-11-07
  • CC-BY-NC-ND-4.0
  • 5DL-B-058
DOI10.5281/zenodo.24090164

摘要

订阅制企业靠比率经营,而每个比率都把一个动态过程压缩成一个数字、藏起一个假设。本书把 LTV、NRR、LTV/CAC 与 AI 产品的毛利还原为它们所压缩的过程并实际运行,找出比率在哪里成立、在哪里误导。本书不是投资或财务建议;全部数据为程序生成的合成数据。

第一章用移位贝塔-几何队列存活取代「LTV = 毛利 / 流失率」:客户异质时,首月流失率最多把折现价值低估 −59%,而不折现、用成熟基数流失率的路演数字高估 +16% 到 +48%;贝塔参数 a ≤ 1 时期望寿命为无穷大。用六个月数据,单一流失率给出的队列价值为 1364,真值为 1966;异质性模型在结构正确时接近、在风险率随使用时长下降时高估 +33%,所以应报告跨模型的区间。第二章把 NRR 写成 dR/dt = A + (e − c)R:获客恒定时,低于 100% 饱和(80% 的天花板为 538),等于 100% 线性,高于 100% 复利——e = c 处是分岔;但扩张有界,一个队列的 NRR 从第一年的 144.1% 降到第五年的 95.7%,一家稳定公司的混合 NRR 随基数老化从 134.6% 降到 100.2%。第三章说明 LTV/CAC 藏起了现金:单位经济学相同,所需资金从不增长时的 912,868 升到月增长 10% 时的 8,320,737;口号比率最好(4.96)的企业需要最多的钱,折现比率却最低(1.80)。

第四章从第一性原理计算 token 成本:批量为 1 时 8B 模型的输出为每百万 3.012 美元、输入 0.018;批处理降到 0.072;长上下文拿走批处理;智能体循环重发历史,一百轮对话中输出只占成本的 6.1%,缓存把 0.6137 降到 0.0708。第五章说明路由以三分之一到一半的成本买到大模型的大部分准确率;固定价格下决定毛利的是平均用户,对数标准差 2.5 时毛利率为 −36.1%,且轻度用户更快流失使毛利率从 +81.3% 侵蚀到 +77.0%。第六章比较梅特卡夫的 n² 与 Odlyzko-Tilly 的 n ln n(十倍用户分别带来 100.0 倍与 11.60 倍价值),加入拥堵后价值先升后降,并说明当变现能力每年增长 50% 时,收入对用户的拟合会给出 1.91 的指数而「证实」错误的定律。第七章与《史记·货殖列传》「贪贾三之,廉贾五之」与计然「无息币」并读,把利润率与周转率的读法明确标为诠释。第八章给出四步核查与一项预注册价格检验(每组 7,908 位注册用户)。书中以标注框区分已发表的证据、可运行的 Python 计算(16 段程序)、原典与作者的假说。

关键词:订阅制经济学;客户终身价值;流失率;队列分析;移位贝塔几何模型;异质性;生存分析;净收入留存;负流失;微分方程;分岔;获客成本回收期;现金流;单位经济学;大模型推理;预填充;解码;KV 缓存;批处理;提示词缓存;模型路由;重尾用量;毛利率;梅特卡夫定律;网络效应;拥堵;混杂;A/B 测试;预注册

Abstract

Subscription businesses are run by ratios, and each ratio compresses a dynamic process into one number and hides an assumption. This book writes LTV, NRR, LTV/CAC and the margins of AI products back out as the processes they compress, runs them, and finds where each ratio holds and where it misleads. It is not investment or financial advice; all data are synthetic, generated by its programs.

Chapter 1 replaces "LTV = margin / churn" with shifted-beta-geometric cohort survival: with heterogeneous customers, first-month churn understates discounted value by up to −59 %, while the undiscounted pitch-deck number using a mature base's churn overstates it by +16 % to +48 %; for Beta parameter a ≤ 1 the expected lifetime is infinite. From six months of data a single churn rate values a cohort at 1364 against a true 1966; a heterogeneity model is close when its structure holds and overstates by +33 % when hazards fall with tenure, so a range across models should be reported. Chapter 2 writes NRR as dR/dt = A + (e − c)R: with constant acquisition, revenue saturates below 100 % (a ceiling of 538 at 80 %), grows linearly at 100 % and compounds above it — a bifurcation at e = c — but expansion is bounded, so a cohort's NRR falls from 144.1 % in year one to 95.7 % in year five and a steady company's blended NRR falls from 134.6 % to 100.2 % as its base ages. Chapter 3 shows that 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 the business with the best slogan ratio (4.96) needs the most money and has the lowest discounted ratio (1.80).

Chapter 4 computes token costs from first principles: at batch 1 the 8B model's output costs 3.012 dollars per million against 0.018 for input; batching cuts it to 0.072; long context removes the batch; agent loops re-send their history, so output is only 6.1 % of the cost of a hundred-turn conversation, and caching cuts 0.6137 to 0.0708. Chapter 5 shows that routing buys most of a large model's accuracy at a third to a half of its cost; that against a flat price the mean user sets the margin, which is −36.1 % at a log-sd of 2.5; and that faster churn of light users erodes margin from +81.3 % to +77.0 %. Chapter 6 compares Metcalfe's n² with Odlyzko and Tilly's n ln n (tenfold users giving 100.0-fold against 11.60-fold value), adds congestion so that value peaks and falls, and shows that with monetisation growing 50 % a year a revenue-on-users fit returns an exponent of 1.91 and "confirms" the wrong law. Chapter 7 reads the *Shiji*'s "the greedy merchant makes three, the modest merchant five" and Jiran's "let no money lie idle", marking the margin-against-turnover reading as an interpretation. Chapter 8 gives a four-step check and a pre-registered price test (7,908 sign-ups per arm). Marked boxes separate published evidence, runnable Python calculations (16 programs), classical sources and the author's hypotheses.

Keywords: subscription economics; customer lifetime value; churn; cohort analysis; shifted-beta-geometric model; heterogeneity; survival analysis; net revenue retention; negative churn; differential equations; bifurcation; CAC payback; cash flow; unit economics; LLM inference; prefill; decode; KV cache; batching; prompt caching; model routing; heavy-tailed usage; gross margin; Metcalfe's law; network effects; congestion; confounding; A/B testing; pre-registration

目录 · Contents

  • 序:比率背后的算术English 中文Preface: The Arithmetic Behind the Ratios
  • 第一章 客户终身价值不是毛利除以流失率English 中文Chapter 1. Lifetime Value Is Not Margin Divided by Churn
  • 第二章 净收入留存:100% 处的门槛,以及它为何不能持久English 中文Chapter 2. Net Revenue Retention: a Threshold at 100 %, and Why It Does Not Last
  • 第三章 现金低谷:为什么单位经济学很好也会把钱烧光English 中文Chapter 3. The Cash Trough: Why Good Unit Economics Still Run Out of Money
  • 第四章 一个 Token 的成本:预填充、解码与内存墙English 中文Chapter 4. What a Token Costs: Prefill, Decode and the Memory Wall
  • 第五章 重度用户、固定价格与模型路由English 中文Chapter 5. Heavy Users, Flat Prices and Routing
  • 第六章 网络效应:n²、n log n,以及用户越多越糟的那一点English 中文Chapter 6. Network Effects: n², n log n, and the Point Where More Users Hurt
  • 第七章 「贪贾三之,廉贾五之」:利润率、周转率与「财币欲其行如流水」English 中文Chapter 7. "The Greedy Merchant Makes Three, the Modest Merchant Five": Margin, Turnover and Money That Flows
  • 第八章 四步核查与一项预注册的价格检验English 中文Chapter 8. A Four-Step Check and a Pre-Registered Price Test

How to cite · 引用

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

BibTeX
@book{fan2026subscription,
  title = {{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},
  author = {Fan, Yafu},
  year = {2026},
  publisher = {5D Library},
  edition = {Version 1.0.0},
  doi = {10.5281/zenodo.24090164},
  url = {https://doi.org/10.5281/zenodo.24090164},
  note = {Chinese title: 订阅制单位经济学、净收入留存与人工智能推理标度律}
}
DOIhttps://doi.org/10.5281/zenodo.24090164
DOI(全部版本 all versions)https://doi.org/10.5281/zenodo.24090163
ORCIDhttps://orcid.org/0009-0004-3519-1203
5D Library ID5DL-B-058
版本 Version1.0.0 · 中英双语版 Bilingual edition
出版日期 Date2026-11-07
网址 Onlinehttps://5dlibrary.com/books/subscription-unit-economics-net-retention-ai-inference-scaling/
SHA-256 · subscription-unit-economics-net-retention-ai-inference-scaling_bilingual_v1.0.0.pdf915edbdc822fe16673ed2c87471341b8f51bdc95586d1611ad901e30ed288d7e
SHA-256 · subscription-unit-economics-net-retention-ai-inference-scaling_en_v1.0.0.pdf214ed220cb0652d1b96f058f6bf483c95b82c3263293e1ff6aa9a9f1bfe2b854
SHA-256 · subscription-unit-economics-net-retention-ai-inference-scaling_zh_v1.0.0.pdf2c5f82e69b409554199f71c72b965fb1596dcec6cf81424b8fbd97b752f29ee0
Licence · 许可Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0). https://creativecommons.org/licenses/by-nc-nd/4.0/

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.