What It Takes to Win as a Founder | Ben Horowitz

Andrew Chen opens by recalling his first visit to Ben Horowitz’s LoudCloud/Opsware offices, where every conference room was named after a hip-hop artist. Horowitz had told Chen that, after selling the company to HP, he planned to retire. Instead, he started another company and later co-founded Andreessen Horowitz. Horowitz explains that he had another idea he needed to get out into the world. For him, that compulsion is central to entrepreneurship: a great entrepreneur feels they have no choice but to build the company, because someone who can simply walk away will probably eventually quit. That motivation becomes the episode’s frame. Horowitz says entrepreneurship is not mainly about choosing an attractive career or following a predictable process. It is about believing strongly enough in an idea that building it becomes necessary, even though the work will expose the founder to uncertainty, failure, and self-doubt.

Chen asks what makes AI different from earlier waves such as desktop software, the web, mobile, and cloud computing. Horowitz says AI resembles the early internet in one respect: every day brings a new and surprising capability. But he argues that AI is technologically different because it is a new kind of computer—a probabilistic computer—whereas the internet was primarily a network. That distinction changes adoption. During the 1990s internet boom, investors could see the eventual opportunity before the network was large enough to support most businesses. Horowitz recalls that Netscape had 90% browser share and 55 million users in 1996, while roughly half of those users were still on dial-up. Internet companies had to spend heavily on Sun Microsystems servers, EMC storage, and other infrastructure, but could reach only a limited audience; the resulting mismatch helped produce the dot-com and telecom crashes. AI applications can instead be put online and immediately reach the already-existing internet audience. Horowitz says this is why AI companies have grown faster than companies in previous waves: they do not need to wait for more people to join the network.

Chen asks whether AI-generated code will make software creation as accessible as online content creation, producing billion-dollar companies run by solo founders or very small teams. Horowitz agrees that a small group can reach a giant audience and make substantial money. He points to Instagram, which had only 13 employees when it sold for $1 billion, while noting that Meta later grew the operation substantially. The unresolved question is whether a tiny team can sustain the business or will eventually be overtaken by a company doing more. Horowitz argues that demand tends to expand rather than stabilize. He cites John Maynard Keynes’s prediction that people would work only 15 hours a week once they had food, clothing, and shelter, then contrasts it with the later proliferation of cars, televisions, smartphones, air conditioning, and other wants. As software becomes easier to build, companies may multiply like fast-fashion businesses, with copycats and perhaps personalized applications generated at runtime. He is unsure how much of today’s application model will survive. One possible future is that users organize data and ask agents to act on it directly, rather than opening conventional applications. Even if companies start small, rising expectations may require larger organizations to deliver more products, channels, and services.

Chen raises the criticism that many AI products are merely “GPT wrappers” without network effects. Horowitz calls the phrase misleading and compares it with the early reaction to relational databases. When relational databases appeared, people similarly asked why companies such as SAP, Siebel, or Salesforce were needed if customers could simply tell the database what to do. Those businesses proved that valuable applications could be built on top of a foundational technology. Horowitz says modern AI applications may contain many models themselves. He cites Cursor as having 14 models, and argues that products serving a specialized audience accumulate valuable data and understanding—for example, a programming tool that constantly observes how elite programmers design software. That knowledge can differentiate the product from a general foundation model. He also says that merely having a somewhat better model is not enough. ChatGPT and Grok may be only marginally different for a user who is already accustomed to one interface, so switching requires a product that is dramatically—effectively ten times—better. Winning the market then creates more data about the specific users, reinforcing the advantage. He expects other forms of network effects to emerge, possibly through crypto-based systems that let AI agents exchange money with one another.

Chen compares today’s AI builders with earlier generations of internet companies. Early web products such as Yahoo required Stanford PhDs to build basic infrastructure, web servers, and systems capable of handling unprecedented scale. A decade later, tools such as the LAMP stack made it possible for ordinary programmers to build online services, and companies such as Airbnb emphasized design and user experience rather than purely technical novelty. Horowitz agrees that AI may follow a similar progression. The current phase has a large concentration of PhD and computer-science talent, while mainstream developers are increasingly able to build AI applications. However, the timing of later phases depends on technical breakthroughs. He says pure language-model capabilities have plateaued somewhat, while reinforcement learning—particularly for coding—continues to improve. Each major architectural breakthrough could rearrange the application and infrastructure layers, so founders and investors must watch for where those breakthroughs occur rather than assume the current design of the stack is permanent.

Chen asks how smaller teams and cheaper company formation could affect venture capital. Horowitz says the VC model changed little from 1970 until 2008 or 2009, even as the internet made companies much easier to build. Earlier companies such as Sun Microsystems or Tandem Computers needed manufacturing, professional services, direct sales, and other expensive capabilities before selling their first product. By contrast, Twitter could begin with five people, laptops, and AWS; AI is another step toward lower-cost formation. He expects this to change the VC landscape because individual investments may become cheaper and the pool of promising companies may expand. Still, he says the core VC function will remain helping a company gain momentum and become more attractive to engineers, customers, follow-on investors, and potential partners. Two companies may have the same idea and equally strong teams, but the one that gains momentum first can hire better people, build a better product, win customers, and raise money more easily. Andreessen Horowitz evolved from offering mainly a smart board member and a personal network into a firm providing operational help, including customer introductions, executive recruiting, and support for international expansion, channels, and multiple products.

For Speedrun founders preparing to present to roughly 1,000 investors, Chen asks how they should evaluate seed-stage partners. Horowitz says new founders almost never know how to perform the CEO job. They must hire executives for functions they have never practiced themselves—he gives the example of hiring a CFO without knowing what a control structure is or how to communicate with Wall Street. They also have to manage ambitious employees, crises, and competitors offering enormous compensation. The most common failure he sees is a crisis of confidence. A founder knows a decision is needed but hesitates because they cannot tell whether it is correct. That hesitation makes the company nervous, slows execution, and creates politics as others try to fill the leadership vacuum. Horowitz recommends finding an investor who can create momentum and help with hiring and fundraising, but also one who can provide context-specific advice rather than generic slogans such as “hire A players.” He illustrates this with a CEO whose CTO was an excellent product leader but had made a finance employee cry. Horowitz advised the CEO not simply to fire him, but to explain that being an effective CTO required working across functions and not making colleagues afraid. A useful mentor, he says, helps a founder have that difficult conversation.

Horowitz describes the decision to take LoudCloud public with only $2 million in trailing-12-month revenue, 18 months after founding. The forecast was to grow from $2 million to $75 million, which he acknowledges was clearly a bad-looking idea. Yet the alternative was bankruptcy: the company had spoken with essentially every private-market investor and found no viable path. Going public was the less bad choice, but making it without hesitation was the difficult part. He says many companies went bankrupt because they could not make a similarly uncomfortable decision. He applies the same logic to re-architecting a product even when it will damage near-term fundraising and quarterly numbers, or firing a head of sales even when doing so leaves the company without an obvious replacement. Founders should focus on the best action available, not on every frightening consequence that could follow. He also rejects universal formulas for being a CEO. His book emphasizes situations because the right action depends on the company, product, people, and crisis. A mentor’s value is not a five-step method but a conversation with someone who has faced a comparable problem and can translate experience into a usable decision.

Horowitz says leaders and investors must adapt when a company enters a new era. At Andreessen Horowitz, expertise in mobile consumer internet or traditional SaaS may no longer be enough if the firm’s opportunities move elsewhere. The firm’s original marketing methods also stopped working as it became too prominent to rely on flattering press coverage, so it developed new capabilities around direct communication, podcasts, and a broader media presence. He admits that asking people who performed well in the past to develop new skills is one of the parts of leadership he enjoys least, but says it is necessary. He warns against judging technologies by whether they are better than their predecessors in every category. Early 3D printers looked revolutionary, but they did not improve according to Moore’s law because progress depended on difficult materials science. Early reviews of the iPhone complained that it was a bad phone, a worse computer, and had an inferior keyboard compared with a BlackBerry. Its breakthrough was putting a camera and GPS in a pocket, enabling new products such as Uber. Horowitz therefore looks for magnitude of strength rather than lack of weakness: a founder or technology should be world-class at something, possess a secret others lack, or make a fundamentally new capability possible.

Asked what he is most confident about in AI, Horowitz points to video and entertainment. He compares the transition from plays to film with what he expects from film to AI-generated film. Early movies simply recorded stage plays with a fixed camera, but the medium eventually developed its own language. Likewise, AI may initially generate isolated scenes or imitate podcasts and television, but Horowitz expects new creators to use it for a different kind of storytelling. Social media often makes the user the main character, which can exclude people who do not want that role; AI can create characters who are not the user and support a broader storytelling medium. In his closing advice to Speedrun founders, Horowitz congratulates them for betting on themselves and says their job is to make that bet pay off. Doubts will inevitably appear, including doubts that founders create themselves, but worrying about them produces nothing. He invokes Mark Zuckerberg, whom he remembers doubting Facebook’s prospects at age 22, to emphasize that feelings are not the measure of success; actions are. Founders should focus on what they can do. If their “guts aren’t boiling,” he says, they may not be trying hard enough—but fear is not a reason to stop.