a16z General Partner Andrew Chen: The Scarce Resources of Startups Keep Changing
a16z General Partner Andrew Chen summarized that in 2010, the scarce resource for startups was "developer talent," and financing was aimed at "hiring people to develop products"; in 2016, it shifted to "traffic," with financing mainly used for "buying ad clicks"; by 2022, the focus turned to "GPU computing power," with financing used to "buy computing power"; and by 2026, many startups will again shift their spending towards "buying model tokens."
English venture capital and industry observations also show that this evolution is highly synchronized with infrastructure cycles: 2010 was marked by the "explosion of mobile apps and a shortage of development talent"; 2016 was characterized by the "monetization of traffic on social and short video platforms"; 2022 focused on "model training computing power and first-mover advantages in clusters"; and by 2026, many projects will revolve around AI model computing power tokens. Capital is no longer just burning money for backend capacity and user acquisition.
In this logic, the consensus between VCs and founders on "where the scarce resources are and what money must be spent on" essentially defines the "value migration path" of each cycle.
Source: Public Information
ABAB AI Insight
The "displacement of scarce resources" essentially redefines the "marginal growth drivers." When a factor becomes relatively cheap, practitioners will concentrate financing on "more expensive, harder to obtain, but more critical for long-term dominance" variables. In 2010, computing power and cloud infrastructure had become affordable for regular use, but developers who could write mobile apps were still scarce, so financing was for "hiring people"; in 2016, development tools matured, but traffic was concentrated on platforms, so financing became about "buying access"; in 2022, data and open-source models gradually became widespread, but GPU clusters and training costs remained high, so financing shifted to "buying computing power"; by 2026, deployment and AI generation will become simpler, while "protocols, tokens, and on-chain liquidity" will become scarce anchors, leading financing to turn towards "buying on-chain assets."
From a long-term capital structure perspective, this migration also continuously raises the "geographical location of the moat." In 2010, the moat was in "teams and technology"; in 2016, it was in "data and network effects"; in 2022, it was in "computing power and data flywheels"; and in 2026, it will again be in "tokens and protocol control." This shifts the narrative for entrepreneurs from "what we want to do" to increasingly focusing on "what resources we must control," and these resources can naturally be capitalized, repriced, and traded in secondary markets.
A deeper danger lies in the fact that once "buying tokens" becomes the mainstream model of the cycle, capital may divert funds that should be invested in "long-term cash flows and business models" into "speculatable on-chain assets," leading to a misalignment between the "real economy and financial assets." In extreme cases, startups could operate through a cycle of "buying tokens—telling stories—secondary market hype—refinancing" without having to genuinely prove their business and profit structure. This could distort capital efficiency and risk exposure again, driving a new round of "asset bubbles + project shells," waiting for the next cycle to be forced to "return to fundamentals."