Ben Horowitz on AI Anxiety, Big Tech Transitions & The Future of Startups | a16z
Summary
- Horowitz says AI has repealed two old laws of software: money can now buy speed, and possession no longer protects incumbents. With enough money, good data and GPUs, a laggard can “solve basically anything in software,” while replicable code, portable data and AI-operated interfaces erode migration, data and UI lock-in. Pricing must attach to genuinely distinct value because, as Rampell notes, a product’s runway might shrink from years to “five weeks.”
- The “SaaS apocalypse” is a terminal-value crisis, but Horowitz rejects a blanket death sentence for legacy software. CEOs must distinguish customers shifting spend elsewhere—which may require them to “cut deep and pivot”—from valuation compression masking a strengthening business. Navan remains defensible through global travel relationships and budgeting integrations; OpenAI and Anthropic do not naturally want to sell to travel managers, and the agentic travel experience is currently “much more complicated than one would think.”
- The investable scarcity is moving beyond GPUs into electricity, memory, rare-earth minerals, manufacturing and grid equipment. a16z raised $15 billion for four of its seven funds, versus a $300 first fund, partly because America must rebuild infrastructure “right now.” Horowitz expects NVIDIA may produce enough chips long before the system has enough electricity or memory; a new DRAM factory would take five years.
- AI-generated impersonation makes cryptographic identity and internet-native money core infrastructure rather than crypto side quests. Horowitz’s nightmare is an AI version of himself ordering a $500 million transfer; his required stack proves “are you a human,” “are you me” and “did I sign this content?” Blockchain could also provide people with payment addresses and let AI agents become economic actors through a bearer instrument on the internet.
- Venture capital has two radically different futures, and Horowitz refuses to pretend he knows which wins. AI could produce a few enormous companies, repeating the auto industry’s consolidation from roughly 300 manufacturers to the Big Three and pushing financiers into bank-like institutions. Or frontier models could plateau or become utilities, while power scarcity pushes computing to the edge and capable small models onto phones. Horowitz’s conclusion is unresolved.
- Horowitz’s answer to AI anxiety is abundance, tempered by honest uncertainty about the transition. He says 8 billion people who might have an idea can now get it out of their heads as code, music or film without a capital or idea gate. Technology has historically produced new kinds of work and expanding needs. His conditional forecast is that in 15 years nearly everyone in America—and probably worldwide—will live better, in terms of luxury and information access, than anyone did in 1980, though telling children what they should do remains “a hard one.”
Deep dive
1. AI destroys inherited software moats before it destroys incumbents
Horowitz’s first changed “law of physics”: the mythical-man-month constraint no longer holds. “You can throw money at the problem”—with enough capital, useful data and GPUs, a company can overcome software deficits that another thousand engineers could not.
The second casualty is possession. Code is easier to reproduce, data easier to move, and AI agents are flexible in how they use interfaces, leaving migration, data and UI lock-in “pretty much gone.”
Rampell sharpens the valuation problem: a strong product once offered five or ten years of runway; now it might offer “five weeks.” Staying private makes an existential pivot easier, but waiting too long can still leave the company worth zero.
Horowitz’s operating test is whether the business is strengthening while expectations reset. Navan’s valuation may imply “travel is dead,” yet travel still requires worldwide airline, hotel and rail relationships, budgeting connections and a channel to travel managers—one that neither OpenAI nor Anthropic naturally wants to build. The agentic travel experience is also currently more complicated than expected.
2. Physical capacity becomes the binding constraint on AI growth
a16z’s scale reflects the transition: its first fund was $300; it has now raised $15 billion for four of seven funds, with international capital rising from essentially zero to roughly 35%.
Horowitz’s infrastructure call is blunt: America lacks rare-earth minerals, electricity and manufacturing capacity, while existing chips consume too much power. The firm even invested in a physical power-transformer company because an old, essential component needs to become easier to manufacture and more efficient.
Rampell’s 1999 analogy—everyone said more fiber had to be built—meets an important distinction. Much of that fiber stayed dark because servers, application infrastructure and end users were not ready; today’s GPUs are lit, demand for tokens is “straight vertical,” and almost every supply-chain layer is constrained. NVIDIA may solve chips first, only to expose shortages in memory and electricity.
3. AI turns authentication and crypto payments into shared infrastructure
Horowitz’s early warning was an AI impersonator joining Zoom and telling his finance team to wire “$500 million to Nigeria.” His proposed rule is: “Don’t believe anything from me unless it’s got my cryptographic key on it.”
The required proof stack escalates from human-versus-bot, to personal identity, to signed content. Horowitz expects AI eventually will not reliably detect AI, making cryptographic provenance—not another classifier—the answer.
For that source of truth, he would rather trust “the mathematical game-theoretic properties of the blockchain” than Google, Meta or the U.S. government. A real Marco Rubio speech, for example, would carry cryptographically strong evidence that it was authentic.
Payments create a second opening. Horowitz says that, depending on the figures, around $450 billion of stimulus was stolen and argues everybody needs an address that can receive money; AI agents likewise need “internet money” because conventional merchant infrastructure may not work for a nonhuman. Rampell adds the harder wrinkle: verified humans can have their addresses used by machines, so CAPTCHA-style tests may need economic and game-theoretic enforcement, echoing Hashcash.
4. Venture capital could become banking—or become vastly larger
Rampell’s case for VC surviving automation is that founders are nondeterministic bets: investors judge whether someone can “materialize labor, capital and customers.” Horowitz notes the obvious dependency—if the venture-capitalist job survives, so must an entrepreneur job.
One precedent is industrial consolidation. Horowitz says roughly 20% of American workers worked for the auto industry in the 1930s; around 300 auto companies compressed into the Big Three, while venture capitalists associated with railroads and automobiles evolved into institutions such as JPMorgan Chase and Goldman Sachs.
Another future is that models hit an asymptote and the major labs become nationalized utilities, or “electricity-plus-plus.” Power scarcity could either entrench companies that capture the electricity and GPUs or push computing to the edge, where sufficiently capable small models run on phones. Horowitz’s conclusion is deliberately unresolved: “I don’t know.”
5. Abundance expands human wants faster than technology removes work
Horowitz’s optimistic framing is permissionless creation: “Now 8 billion people that might have an idea in their head can get it out of their head.” The output may be code, music or a movie; capital and institutional approval are no longer universal gates.
Their historical analogy is America shifting from 93% or 94% farmers in 1750 and 98% in 1789 to almost none today. A farmer then might have dismissed product marketing as absurd—“You’re not making any food”—yet new forms of work became normal as technology changed.
Horowitz says Keynes missed this adaptive demand when predicting abundance would reduce work to a maximum 15-hour week: one car became a car per person, while computers, televisions, elaborate vacations and tasting menus turned wants into needs. His 15-year forecast is that people will live “way, way, way better,” with luxury access and information exceeding the very best available to anyone in 1980, but he concedes the transition is disconcerting—especially when children ask what they should do.