跳到正文
原文
Tomasz Tunguz 博客·· 10 天前AI 评分68

AI 资金内循环有多大？Tomasz Tunguz 对比电信时代测算基建回报

How Large is the AI Circle?

文章要点
  • 1对比电信互联网时代，AI 应用收入达到基建开支 10% 只用了四年，电信时代用了七年
  • 2Apollo 分析 Ramp 数据称头部 1% AI 支出企业中 80% 有风投背景，风投支持企业占美国就业约 13%
  • 32026 年前九个月美国风投投出 5160 亿美元，超 2021 年全年总额 44%，AI 公司占交易额 82.7%
  • 4AI 前四年累计风投 1.24 万亿美元，约为互联网时代同期 2018 亿美元的 6.2 倍（2026 年美元口径）
决策要点

作者用电信时代的基建回报框架衡量 AI 周期，结论是应用收入追上投入的速度快了一倍，但资金内循环体量也远超当年。

  • 影响：2026 年风投交易额的 82.7% 流向 AI 公司，其他方向融资空间受挤压；头部 AI 支出方八成有风投背景，创业融资节奏变化会直接传导到上游 AI 供应商的收入。
  • 参考：可考虑把应用收入对基建开支的比例当作周期监测指标；评估 AI 供应商或制定自家 AI 预算时，值得先算清需求中有多少来自融资中的创业公司，再定扩张节奏。
  • 依据：收入与 capex 对比来自作者基于 SEC 文件、USTelecom 与 NVCA 数据的整理；八成头部支出方有风投背景出自 Apollo 对 Ramp 数据的分析，作者自注样本偏向科技业。
正文

In short : Every major build-out has two phases per Carlota Perez: infrastructure, then the applications that pay for it. The telecom internet build-out ran capex twelve times ahead of application revenue at the 2000 peak, & took seven years to get application revenue above 10% of capex, & ten to reach 48%. AI has reached that 10% mark in four years, roughly half the time.

Every major technology build-out requires two phases, according to Carlota Perez1: the infrastructure to power it, & then the applications to pay for it.

Between 1996 & 2005, American telecom companies spent $1.5t building the internet in today’s dollars.2 Meanwhile, applications riding on that infrastructure booked $180 billion of revenue.3

That’s an interesting ratio, isn’t it? Application revenue to infrastructure spend. It matters because the shorter the gap between the two, the sooner the infrastructure pays for itself. Let’s compare this to the AI era.

It would take a decade for dark fiber to fill its promise, flooded with the packets of application revenue. AI is achieving that mark in half the time.

Grouped bars comparing application revenue as a share of infrastructure capex by year since each build-out began, web era 1996-2005 against AI era 2023-2026

AI hit double digits in its fourth year. Plus, the unit economics from companies like SpaceX suggest that the paybacks on the infrastructure from this revenue are short, about a year.4

Grouped bars comparing US venture capital invested by year since each build-out began, web era against AI era, in 2026 dollars

US venture invested $516b in the first nine months of 2026. That is already 44% above the full-year record set in 2021.5 Inflation-adjusted, the run-up in venture investment is roughly two & a half times the size of the earlier era.6

In a very real sense, the entire ecosystem is hoping the gap between application revenue & infrastructure investment is significantly shorter on an inflation-adjusted basis. Cumulative venture investment in this wave is about six times the first four years of the web era.6

The major risk is whether the ecosystem is an ouroboros: is all the spend just within startups?7

80 percent of the largest AI spenders are venture-backed, against 13.1 percent of the median spender

Ramp’s data show that 80% of the largest AI spenders are venture-backed. Those firms now account for more than 13% of American employment, up from about 5% during the web era. The circularity is real, but the circle is a sixth of the economy.

That is a much larger circle.8

  1. Perez, Technological Revolutions and Financial Capital, 2002. Her central mechanism is a sequence: installation of infrastructure, then financial speculation, then a turning point, then deployment of the applications that justify the spending. ↩︎

  2. USTelecom historical broadband provider capex, all-in across wireline, wireless & cable. Nominal spend of $761.9b for 1996-2005, converted to 2026 dollars with CPI-U (2.116 to 1.699 across the period). ↩︎

  3. A 24-company panel from SEC 10-K and S-1 filings: Amazon, Yahoo, Google, eBay, Priceline, Expedia, Travelocity, Overstock, Netflix, DoubleClick, Overture, Ask Jeeves, LookSmart, Excite, E*Trade, Ameritrade, Schwab online, Monster, CNET, WebMD, Salesforce, MSN, Hotmail and Match. The 2001-2005 figures are a floor; four companies never broke out their online revenue. Google & Yahoo are gross of traffic acquisition cost. Nominal $100.9b, converted to 2026 dollars, or about $182b; rounded in the text to $180b. ↩︎

  4. SpaceX’s third-quarter S-1 states that “at current pricing, new compute deployments are generating a payback of less than a year.” Working through the figures in the GPU revenue ladder: a $46 per watt build costs $55,200 per card, and the $5.27 contract rate yields $46,165 of annual revenue against roughly $570 of cash operating cost, giving about a year at 1.21 years. At B200 spot prices it is 0.72 years. Reconciling the quote to a $46 per watt build implies an after-tax measure: at a 21% federal rate, bonus depreciation lifts first-year cash flow to $57,187 & shortens payback to 0.97 years, though that is a timing benefit rather than a cost reduction. ↩︎

  5. PitchBook-NVCA Venture Monitor, Q3 2026, as of 30 September. The 2026 figure is nine months, not a full year, & is plotted unannualized. The full-year 2021 record was $357.9b. AI companies took 82.7% of 2026 deal value, the highest share in the dataset. ↩︎

  6. Cumulative venture investment in 2026 dollars was $201.8b across the first four years of the web era (1996-1999) & $1.24t across the first four years of the AI era (2023-2026), a ratio of 6.2x. Measured against the web era’s entire decade ($630.3b) rather than its first four years, the same $1.24t is 2.0x. The four-year comparison is used in the text because it matches the two eras on elapsed time. Sources are NVCA/MoneyTree for 1996-2005, the NVCA 2026 Yearbook for 2023-2025, & the PitchBook-NVCA Venture Monitor for 2026, all inflated to 2026 dollars with CPI-U. The 2026 figure covers nine months, so the full-year multiple will be higher. ↩︎ ↩︎

  7. Apollo’s analysis of the Ramp AI Index, published 22 October 2026, finds that 80.4% of the top 1% of AI spenders are venture-backed, against 13.1% for the median spender. Ramp separately reports that 80% of spend on OpenAI & Anthropic is driven by 1% of firms. Three caveats matter here. The chart measures dollars spent, not tokens, so it cannot support a claim about token share. Ramp’s sample is 85,000+ US businesses skewed toward tech, drawn from card & bill-pay data, & excludes free tools & personal-account shadow AI, which Ramp estimates at roughly 40% of AI application spend. And venture backing is not the same as venture dollars paying for the tokens: the chart shows the funding status of high-spending firms, not the source of the spend. Apollo is using the number to advance a circularity thesis & Ramp sells spend-management software, so both have an interest in it being vivid. ↩︎

  8. Puri & Zarutskie put the employment share of venture-backed firms at roughly 4-7% across 1991-2005, using Census firm-level data; the Census Annual Business Survey and Longitudinal Business Database, as analysed by Ando (2025), put it near 13% for 2017-2021. The two estimates use different sources and slightly different definitions of venture backing, so this is a directional comparison across eras rather than a continuous series. Venture-backed firms remain about 0.64% of all US firms, so the employment share reflects their disproportionate size rather than their number. ↩︎

Get the next one in your inbox

The 1-minute read that turns tech data into strategic advantage.
Read by 150k+ founders & operators.

GP at Theory Ventures. Former Google PM. Sharing data-driven insights on AI, web3, & venture capital.

Bloomberg • WSJ • Economist

来源：Tomasz Tunguz 博客 · tomtunguz.com

#观点趋势#金融