5D Library › 订阅制单位经济学、净收入留存与人工智能推理标度律 · English

第二章 净收入留存:100% 处的门槛,以及它为何不能持久

一个方程

净收入留存(NRR)比较的是:一年前的客户现在付给你的钱,与他们当时付的钱之比,已离开的客户按零计。高于 100% 时,存量客户自己就在增长;投资者称之为「负流失」。本章把它写成一个微分方程,然后问:高 NRR 能持续多久。

推导

模型。 令 R(t)R(t) 为月经常性收入,AA 为每月靠获客新增的月收入,ee 为存量客户的月扩张率,cc 为其月收缩加流失率: dRdt=A+(e−c)R,R(0)=0。\frac{dR}{dt} = A + (e – c)\,R, \qquad R(0) = 0 。 令 k=e−ck = e – c,解为 R(t)=Ak(ekt−1)R(t) = \frac{A}{k}\left(e^{kt} – 1\right),当 k→0k \to 0 时趋于 AtA\,t。年度 NRR 为 e12ke^{12k}。

  • 若 e<ce < c(NRR<100%\text{NRR} < 100\,\%),R(t)→A/(c−e)R(t) \to A/(c – e):一个天花板,无论企业以同样速度获客多久都会触及。不动点 R*=A/(c−e)R^* = A/(c-e) 是稳定的。
  • 若 e=ce = c,R(t)=AtR(t) = A\,t:线性增长。
  • 若 e>ce > c,R(t)R(t) 以速率 e−ce – c 指数增长;不动点移到了负收入一侧并变为排斥。

e=ce = c 处的变化是不动点的分岔——这就是「负流失」是一种质变、而不只是一个更好的数字的含义。

推导

# Chapter 2. Net revenue retention as a differential equation. Recurring
# revenue R(t) (monthly) grows by new business at a constant rate A (new
# customers per month times their starting revenue) and changes on the
# installed base at net rate (e - c):
#     dR/dt = A + (e - c) R,
# where e is expansion and c is contraction plus churn, per month. Annual net
# revenue retention is NRR = exp(12 (e - c)). Solution, with R(0) = 0:
#     e < c:  R(t) = A / (c - e) x (1 - exp(-(c - e) t))   -> saturates at A /
#     (c - e) e = c:  R(t) = A t                                    -> grows
#     linearly e > c:  R(t) = A / (e - c) x (exp((e - c) t) - 1)     -> grows
#     exponentially
# A = 10 (thousand a month of new monthly revenue). Same acquisition in every
# row.
import math

def R(t, A, k):            # k = e - c per month
    if abs(k) < 1e-12:
        return A * t
    return A / k * (math.exp(k * t) - 1)

if __name__ == "__main__":
    A = 10.0
    print(
        "   annual NRR   e - c / month  R at 12 mo   36 mo    60 mo  120 mo "
        " ceiling"
    )
    for nrr in (0.80, 0.90, 1.00, 1.10, 1.20):
        k = math.log(nrr) / 12
        ceil = "none" if k >= 0 else "{:.0f}".format(A / -k)
        print(
            "   {:7.0%}      {:+8.5f}      {:7.0f}   {:7.0f}  {:7.0f} "
            " {:7.0f}   {}".format(
                nrr,
                k,
                R(12, A, k),
                R(36, A, k),
                R(60, A, k),
                R(120, A, k),
                ceil,
            )
        )
    print("\nTime for revenue to double from month 60, by NRR:")
    for nrr in (0.90, 1.00, 1.10, 1.20):
        k = math.log(nrr) / 12
        target = 2 * R(60, A, k)
        t = 60.0
        while R(t, A, k) < target and t < 2000:
            t += 0.25
        print(
            "   {:5.0%}: {}".format(
                nrr,
                (
                    "never (ceiling below double)"
                    if t >= 2000
                    else "{:.0f} months".format(t - 60)
                ),
            )
        )
    print(
        "\nReading: below 100 % NRR a business that keeps acquiring at the"
        " same pace"
    )
    print(
        "still stops growing - revenue approaches a ceiling A / (c - e). At"
        " 100 % it"
    )
    print(
        "grows linearly; above 100 % the installed base compounds on its own."
    )
#    annual NRR   e - c / month  R at 12 mo   36 mo    60 mo  120 mo  ceiling
#        80%      -0.01860          108       262      362      480   538
#        90%      -0.00878          114       309      466      742   1139
#       100%      +0.00000          120       360      600     1200   none
#       110%      +0.00794          126       417      769     2007   none
#       120%      +0.01519          132       479      980     3417   none
#
# Time for revenue to double from month 60, by NRR:
#      90%: 135 months
#     100%: 60 months
#     110%: 40 months
#     120%: 31 months
#
# Reading: below 100 % NRR a business that keeps acquiring at the same pace
# still stops growing - revenue approaches a ceiling A / (c - e). At 100 % it
# grows linearly; above 100 % the installed base compounds on its own.

表中获客保持不变,只改变 NRR。80% 时,无论企业做什么,月收入都不会超过 538;90% 时天花板为 1139,从第 60 个月的水平翻倍需要 135 个月。100% 时线性增长,60 个月翻倍。110% 与 120% 时存量复利:到第 120 个月分别为 2007 与 3417,而线性情形为 1200——销售投入完全相同。

管理上的后果很尖锐。NRR 低于 100% 时,一个保持获客不变的增长计划,就是一个停止增长的计划;要继续增长,获客本身必须增长。高于 100% 时,存量客户成了第二台引擎。

获客完全相同、NRR 取五个水平时十年间的月经常性收入:低于 100% 饱和,等于 100% 为直线,高于 100% 复利增长。

高 NRR 为何不能持久

模型把扩张当作永远不变的月度速率。真实的扩张是有界的:客户买的席位不能多于它的人数,用量不能多于它的工作量。第五段程序让每个账户按逻辑斯蒂曲线向其起始收入的四倍扩张,月流失为 0.01。

推导

# Chapter 2. Why NRR above 100 % does not last. Each account starts at revenue
# 1 and expands logistically toward a ceiling of 4 (it cannot buy more seats
# than it has people): r(t) = 4 / (1 + 3 exp(-0.06 t)), t in months since
# signing. Accounts churn with probability 0.01 a month whatever their age. A
# cohort's NRR in year y is its revenue at the end of year y divided by its
# revenue at the end of year y - 1 (accounts that churned in between count as
# zero, as NRR is defined). A business that signs the same number of accounts
# every month has a blended NRR over its whole base; we compute it at company
# ages 2 to 10 years.
import math

CAP, K, CHURN = 4.0, 0.06, 0.01

def rev(t):
    return CAP / (1 + (CAP - 1) * math.exp(-K * t))

# expected revenue of a cohort of 1 account
def cohort_revenue(t):
    return rev(t) * (1 - CHURN) ** t

if __name__ == "__main__":
    print("A. One cohort's NRR by year since signing")
    for y in range(1, 8):
        print(
            "   year {}: {:6.1%}".format(
                y, cohort_revenue(12 * y) / cohort_revenue(12 * (y - 1))
            )
        )
    print(
        "   long-run limit: (1 - churn)^12 = {:.1%}".format((1 - CHURN) ** 12)
    )
    print(
        "\nB. Blended NRR of a business signing one account a month, by"
        " company age"
    )
    for age_y in (2, 3, 5, 7, 10):
        T = 12 * age_y
        # signing months of accounts live at T-12
        signed = range(0, T - 12)
        before = sum(cohort_revenue(T - 12 - m) for m in signed)
        after = sum(cohort_revenue(T - m) for m in signed)
        print(
            "   company age {:>2} years: blended NRR {:6.1%}".format(
                age_y, after / before
            )
        )
    print(
        "\nReading: expansion is front-loaded, so each cohort's NRR falls"
        " toward the"
    )
    print(
        "survival rate as accounts fill their ceilings, and a company's"
        " blended NRR"
    )
    print(
        "falls as its base ages - with no change in the product or the team."
    )
# A. One cohort's NRR by year since signing
#    year 1: 144.1%
#    year 2: 127.5%
#    year 3: 112.7%
#    year 4: 102.1%
#    year 5:  95.7%
#    year 6:  92.2%
#    year 7:  90.4%
#    long-run limit: (1 - churn)^12 = 88.6%
#
# B. Blended NRR of a business signing one account a month, by company age
#    company age  2 years: blended NRR 134.6%
#    company age  3 years: blended NRR 125.6%
#    company age  5 years: blended NRR 112.0%
#    company age  7 years: blended NRR 104.8%
#    company age 10 years: blended NRR 100.2%
#
# Reading: expansion is front-loaded, so each cohort's NRR falls toward the
# survival rate as accounts fill their ceilings, and a company's blended NRR
# falls as its base ages - with no change in the product or the team.

A 段跟踪一个队列。它的 NRR 第一年为 144.1%,第二年 127.5%,第五年降到 100% 以下(95.7%),并趋近存活率 88.6%。扩张集中在前期;流失则不然。

B 段跟踪一家每月签下同样数量账户的企业。它的混合 NRR 两岁时为 134.6%,五岁时 112.0%,十岁时 100.2%——产品、价格与团队都没有任何变化。只是客户基数变老了。所以,一家成熟中的公司 NRR 下降,本身并不是出了问题的证据;而一家年轻公司 130% 的 NRR,也不是它会继续复利的证据。

注

这对估值意味着什么。 外推一家年轻公司的 NRR——在上面的方程中把一个恒定的 e>ce > c 代入很长的时间——等于假设永远处于指数分支上。第五段程序说明默认路径恰恰相反:随着基数老化,混合 NRR 向存活率衰减,于是企业会跨回饱和分支,除非它不断找到新的扩张空间(新产品、新席位、提价)。

假说 · 尚未验证

作者的假说(尚未验证): 在连续多年披露 NRR 的订阅制公司中,控制提价与新产品发布后,混合 NRR 随客户基数按收入加权的平均年龄而下降。什么结果会证明它错:两者无关,或者 NRR 随基数年龄上升。

小结

获客恒定时,经常性收入服从 dR/dt=A+(e−c)RdR/dt = A + (e – c)R:NRR 低于 100% 时饱和于 A/(c−e)A/(c – e),等于 100% 时线性增长,高于 100% 时存量复利——e=ce = c 处是一个分岔。但扩张是有界的,所以每个队列的 NRR 都向其存活率下降(例子中从第一年的 144.1% 降到第五年的 100% 以下),而一家公司的混合 NRR 随基数老化而下降(从两岁的 134.6% 降到十岁的 100.2%),企业本身并无任何变化。高 NRR 是公司生命中的一个阶段,不是一个可以外推的常数。

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).