Ben Horowitz on Investing in AI: AI Bubbles, Economic Impact, and VC Acceleration

Summary

• Horowitz’s investing filter is concentrated excellence, not all-around competence. The recurring mistake is getting “too wrapped around the axle” over a weakness instead of asking whether a founder or company is “literally the best in the world at a thing”; being “pretty good at a lot of things” is generally worse. Because venture outcomes lag 10–15 years, he judges GPs at the point of attack—sourcing, winning, and underwriting quality—not by waiting for realized portfolios.
• a16z scales judgment by keeping investing groups near basketball-team size. A late colleague identified in the transcript as Dave Swanson/Swinsson’s five-starter analogy drove verticalization as “software was eating the world”; across seven current verticals, overlapping teams cross-attend meetings and the full GP group spends two or three days together twice yearly. The operating principle is that “clarity, not correctness” keeps an organization moving, while politicking is explicitly deincentivized.
• A vertical earns capital when entrepreneurial density and technological change can produce multibillion-dollar companies. Horowitz rejected ESG as a separate lens because investing is hard enough without criteria beyond whether a company can become huge and profitable; American Dynamism was narrowed from a broad marketing story into a tighter fund thesis around real change in defense, public safety, energy, mining, and supply chains. “I want to know what the fund idea is… how do I make money? We have investors. We’ve got to make money.”
• The firm casts technology investing as giving builders “a real shot at life,” with American competitiveness as the mechanism. Ben argues America must win economically and militarily, therefore technologically; a junior employee’s push led to a Mexico meeting about helping secure the border, defense manufacturing, and energy. His cultural call: “If you want to change the world, you have to believe you can change the world.”
• The AI stack is proving more plural and application-specific than the “giant brain” thesis of three or four years ago. Foundation models remain infrastructure, but the fat tail of human behavior must be modeled per use case: Cursor uses 13 AI models and released its own coding foundation model alongside OpenAI or Anthropic options. That makes application complexity a potential moat, weakens simplistic benchmark readings, and gives incumbents reason to “acquire the DNA of the future” through M&A.
• Rapid valuations alone do not settle the AI-bubble question: adoption, revenue growth, and demand are also exceptionally strong. Horowitz calls demand “very intense,” says even NVIDIA’s multiples are not historically outrageous against growth and earnings, and frames AI as a larger technology market than any he has seen. He still hedges—“we’ll see how it plays out”—but thinks there will be more $1 billion and $10 billion companies because AI is a new computing platform with an enormous design space.
• Lean companies have not erased the value of a true company-building partner or necessarily broken venture ownership economics. Recent a16z deals often reach 20% or better; exceptional companies with special founders can have different terms, but those businesses may become valuable so quickly that lower ownership “has been fine.” Against 3,000-plus VC firms, Ben says few can actually help a company succeed, while Speedrun moves earlier because new tools turn ideas into products much faster.

Deep dive

1. World-best strength matters more than a tidy scorecard

• Ben’s management problem is unlike running a conventional company: many GPs are former operators with unusually concentrated talent, so he gives someone like Martin Casado—“probably the best architect in networking software in the last 20 years”—little direct instruction. His work is shaping the investment conversation, risk calibration, conflicts, personnel, and closing.

• The most common underwriting mistake is obsessing over a company’s weakness rather than determining “how great they are.” The winning question is whether they are “literally the best in the world at a thing”; a company that is merely “pretty good at a lot of things” is generally a worse investment.

• Accountability therefore starts “at the point of attack”: can a GP find and win opportunities, and was the decision high quality when made? Waiting 10 or 15 years for portfolio proof permits too many mistakes and missed winners. Ben also watches whether investors remain deep in the technology; when someone “runs out of gas,” the firm has to make a change.

2. Small vertical teams preserve conversation, while clarity creates speed

• A late colleague identified in the transcript as Dave Swanson/Swinsson said in 2009 that an investing team “shouldn’t be too much bigger than a basketball team”—five starters—because investment judgment must remain a real conversation. Verticalization let a16z preserve that intimacy while expanding as software entered more markets.

• Connectivity is deliberately engineered: adjacent groups such as AI infrastructure and AI applications attend each other’s meetings, there is a management meeting for the group, and the full GP group goes away for two or three days twice a year with little agenda.

• Jen’s observation is that incentives favor opportunities “at the intersection,” not territorial protection. Ben says politicking either gets rewarded until coups and infighting follow, or it gets deincentivized; a16z chose the latter.

• Ben stays informed by talking to people “at the point of attack,” where organizational knowledge actually lives—not merely their managers. As a founder, he also invites escalation when something breaks; an issue may take him 14 seconds to resolve. Often people need clarity more than correctness: “If you have clarity, you can move.”

3. A market must work as a fund before it works as a story

• Verticals are products for markets: a cluster of entrepreneurs capable of building multibillion-dollar companies must have distinct needs a specialist team can serve. Crypto, bio, and American Dynamism differ materially; timing is “a little bit of an art,” and entering the right market does not guarantee winning it.

• Ben called American Dynamism a good marketing idea but demanded a “fund idea” grounded in real technological change. Jen says the team found that change in supply chains, defense, and government engagement, then ultimately zoned in on three core vertical areas. He rejected standalone ESG because extra “do good” criteria can distort the investment question: can this become a giant, profitable company?

• The wider purpose is to give people “a shot” to contribute, not promise utopian equality. Ben places the past 250 years’ gains in wealth, lifespan, and population alongside the rise of free markets, capitalism, and rule of law; America must now win economically and militarily by winning technologically. A junior colleague’s conviction catalyzed a Mexico meeting on securing the border, defense manufacturing, and energy.

4. Application behavior is not fully subsumed by the foundation model

• AI threatens every incumbent, so Ben expects substantial M&A: companies will need to “acquire the DNA of the future” and reconstruct how they operate if they are going to survive.

• Three or four years ago, people expected foundation models to become “giant brains that could do anything better than anybody.” Ben says that has not quite happened: large models provide essential infrastructure, but each use case still contains a fat tail of human behavior that applications must model and understand very well.

• Cursor is his load-bearing example: it uses 13 AI models for different aspects of programming and programmer interaction, then released a coding foundation model that can replace or complement OpenAI and Anthropic models. The approach has gotten great adoption, suggesting application complexity is not subsumed by the largest GPU-trained model; benchmarks can therefore mislead, and different video use cases may likewise require different models.

5. Unprecedented demand complicates the bubble verdict

• Ben will not declare the eventual winner count—“we don’t know quite what’s going to happen yet”—but frames AI as a computing platform, not one product cycle. As the internet produced Meta, Netflix, Amazon, and Google, AI’s larger economic impact and “enormous design space” could support more companies above $1 billion and $10 billion.

• Jen surfaces the anxiety around valuations; Ben’s rebuttal is that customer adoption and revenue growth are part of demand he says is unlike anything previously seen. Even NVIDIA’s multiples are “not historically nuts” relative to its growth and earnings. His conclusion remains hedged: this is a “brave new world,” demand is “very intense,” and “we’ll see how it plays out.”

• On leaner companies retaining more equity, Ben says recent a16z investments often still secure 20% ownership or better. Some special companies with very special founders are exceptions, but those businesses can become valuable so quickly that lower ownership “has been fine.”

• More than 3,000 VC firms do not eliminate differentiation: “There are a lot of VCs, but very few who can actually help you succeed as a company.” a16z is also leaning harder into Speedrun because AI tools let entrepreneurs convert ideas into products earlier, extending the firm’s reach before a startup qualifies for venture funding.