Ben Horowitz on AI Infrastructure, Economics and The New Laws of Software

Ben Horowitz argues that a foundational rule of technology held for roughly 50 years, namely that you cannot buy your way out of a software problem by throwing money or engineers at it, no longer applies. With enough GPUs and the right data, companies can now compress years of software development into weeks, meaning a product that once had five to ten years of runway may now have only five weeks before being disrupted. Any CEO who continues viewing the business through the old rules rather than recognizing AI has changed them will, in his view, definitely fail.

The traditional software moats of migration pain, data lock-in, and user interface lock-in are largely gone because code is easy to replicate, data is easy to move, and AI agents are flexible on interfaces. Horowitz says the SaaS apocalypse is real in the sense that financial markets have genuine doubts about the terminal value of existing software companies. Features, products, and companies are becoming harder to distinguish because AI makes it easy to create features and extract data from established players. He argues the best companies are those that have hostages rather than customers through strong data lock-in, and uses Navon in corporate travel as an example of defensible complexity, since it requires explicit relationships with every airline, hotel, and train globally plus connections to corporate budgeting systems, and no major AI company including OpenAI or Anthropic has a channel to the corporate travel manager.

On infrastructure, Horowitz says the United States is effectively out of electricity capacity right now, not in 12 months, while China's electricity capacity is growing steeply and US capacity growth is flat. Token demand is growing vertically but infrastructure capacity is not. He predicts chips will likely reach sufficient supply before electricity does, making electricity the binding constraint on AI scaling. He also notes servers are currently shipping without RAM because memory supply has been entirely consumed by AI demand, and building a new DRAM factory would take approximately five years. Unlike the 1999 fiber overbuild where most fiber was dark, Horowitz says current GPU infrastructure is fully utilized. Andreessen Horowitz invested in a power transformer company on the basis that transformer technology has not meaningfully changed since electricity was invented.

Horowitz argues that AI-enabled personalized phishing at scale makes all communication potentially unusable, and that an AI deepfake on a Zoom call could instruct a finance team to wire large sums of money fraudulently. He says cryptographic signing of content and identity is the only reliable solution, and that AI will eventually be unable to detect other AI-generated video, making cryptographic provenance necessary. He favors a decentralized source of cryptographic truth not controlled by Google or Meta. He also contends that CAPTCHAs no longer function as human verification tools because AI can now claim to be a real person.

Horowitz argues that AIs functioning as economic actors will need a bearer instrument on the internet, likely crypto, since they cannot be credit card merchants. This frames crypto not primarily as a speculative asset but as necessary financial infrastructure for autonomous AI agents operating in the economy.

Andreessen Horowitz raised its first fund of 300 million dollars in 2009 and its most recent raise was 15 billion dollars across four of seven funds, with its international LP base growing from essentially zero to approximately 35 percent. Horowitz frames the current moment as comparable to the industrial revolution and offers several divergent scenarios for how it resolves, including a small number of very large companies monopolizing AI infrastructure, AI labs hitting an intelligence asymptote and being nationalized as utilities, electricity scarcity pushing computing to the edge with small efficient models, or AI enabling 8 billion people to execute ideas without capital gates, which could make venture capital far larger than it is today.

He predicts that in 15 years everyone in America and probably around the world will live better than the very best life available in 1980 in terms of luxury and access to information. He grounds this in the argument that Keynes was wrong about abundance satisfying human wants because human wants convert to perceived needs very quickly and people will always demand goods and services that do not yet exist.