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

Ben explains the fundamental differences between managing general partners and running a company, emphasizing that A16Z has an unusually high concentration of talent. Rather than giving direction, he helps GPs understand how conversation affects investment decisions and focuses on finding entrepreneurs who are literally the best in the world at something specific, rather than being good at many things.

• •Managing GPs requires less direction and more guidance on decision-making processes due to their exceptional expertise
• •Focus investment decisions on finding entrepreneurs who are the best in the world at one thing, not good at many things
• •The biggest mistake is getting too focused on weaknesses rather than exceptional strengths
• •GPs like Martin Casado (best networking architect in 20 years) require different management than typical executives

A16Z evaluates investors at the moment of investment decision rather than waiting 10-15 years for portfolio outcomes. Ben assesses how well investors find opportunities, win deals, and the quality of investments at the time they're made, recognizing that backing exceptional entrepreneurs like Mira or Ilya has inherent value regardless of eventual outcomes.

• •Waiting for portfolio outcomes (10-15 years) is too long to evaluate investor performance
• •Evaluate investors on three criteria: finding opportunities, winning deals, and investment quality at decision time
• •Backing exceptional entrepreneurs has value even if the company doesn't work out
• •This approach prevents missing good investments or making bad ones during long evaluation periods

Based on advice from Dave Swensen, A16Z keeps investing teams no larger than a basketball team (5 people) to maintain quality conversation around investments. Verticalization was necessary as software ate the world and the firm needed to scale while maintaining small team dynamics. Teams are designed around clusters of important entrepreneurs in specific markets.

• •Investing teams should not exceed basketball team size (5 people) to maintain quality decision-making conversations
• •Verticalization allows the firm to scale while keeping individual teams small and effective
• •Verticals are designed around where important entrepreneurs cluster, not arbitrary categories
• •Cross-team communication happens through shared meetings, GP off-sites, and people attending each other's team meetings

Ben stays informed by spending time with people doing the actual work—deal partners, IT teams, accountants, and entrepreneurs—because knowledge lives at the point of attack, not with managers. He encourages people to bring him problems immediately, as most issues require clarity rather than correctness and can be resolved in seconds.

• •Knowledge lives with people doing the work, not managers—spend time at the point of attack
• •Good decision-making combines intelligence with knowledge from those closest to the work
• •Organizations need clarity more than correctness to move forward effectively
• •Encourage people to escalate issues immediately—most can be resolved in 14 seconds with clear direction

Verticals are chosen based on clusters of important entrepreneurs creating multibillion-dollar companies, with different categories (crypto, bio, American Dynamism) having very different needs. A16Z rejected ESG/clean tech as a vertical because it introduced non-economic criteria into investment decisions, instead focusing on American Dynamism with clear technological and economic opportunities in defense, intelligence, and energy.

• •Verticals must match product to market—crypto, bio, and American Dynamism entrepreneurs have very different needs
• •Timing matters: don't be too early or too late to a market opportunity
• •Rejected ESG/clean tech because it introduces criteria beyond 'will this be a giant company and make money'
• •American Dynamism focuses on real tech change in defense modernization, public safety, energy, and rare earth minerals

Ben articulates that the best thing society can do is give people a chance to contribute to something larger than themselves. America's free market capitalist system has been crucial to humanity's rise in wealth, lifespan, and population. A16Z's job is helping America win technologically, which is essential for economic and military success, and this mission empowers team members to pursue ambitious initiatives like the Mexico alliance.

• •Giving people a shot at meaningful contribution has been better for humanity than utopian equality systems
• •America's rise coincides with free market capitalism and rule of law over the last 250 years
• •For America to maintain global importance, it must win technologically for both economic and military reasons
• •Believing you can change the world enables junior team members to pursue high-impact initiatives

AI is so disruptive that every incumbent company is under threat and needs to acquire the DNA of the future to survive. Contrary to expectations, large foundation models haven't become all-knowing brains—instead, application-specific models and orchestration matter more. Cursor uses 13 different AI models and released its own coding foundation model, showing that application complexity isn't subsumed by large models.

• •Every incumbent company faces existential AI threats and will acquire companies to gain future DNA
• •Large foundation models provide important infrastructure but don't solve all use cases as expected
• •Application behavior and model orchestration are more important than having the biggest model with most GPUs
• •Cursor's 13-model approach and custom coding model demonstrate the importance of specialized, orchestrated systems
• •The fat tail of human behavior requires specific modeling that general models don't capture

Different use cases require different AI models, contradicting the assumption from four years ago that one giant model would handle everything. Justine Moore's research shows there's no single god-level video model—different applications need specialized approaches, fundamentally changing expectations about AI development and deployment.

• •Different use cases require different specialized models, not one universal solution
• •Video generation demonstrates this clearly—no single model dominates all video use cases
• •This represents a fundamental shift from expectations four years ago about AI development
• •Benchmarks can be misleading when evaluating model capabilities across diverse applications

A16Z is achieving good ownership levels (20% or better) in recent investments. While some special companies command lower ownership, they become valuable so quickly that it works out. Building companies remains very hard, and most smart entrepreneurs recognize that having a partner who can help them succeed matters more than initial valuation. Few VCs can actually help companies succeed, making A16Z's position special.

• •Recent investments achieving 20% or better ownership in most cases
• •Special companies with exceptional founders may require lower ownership but grow value rapidly
• •Building companies is still very hard—financial partners who can help succeed are rare and valuable
• •Despite 3,000+ VC firms, very few can actually help companies succeed, maintaining A16Z's competitive position

AI should be viewed as a new computing platform comparable to the internet era, which produced multiple massive winners (Meta, Netflix, Amazon, Google). AI products are having even bigger economic impact with unprecedented customer adoption and revenue growth rates. While valuations have risen dramatically, the underlying demand is stronger than anything Ben has seen in his career, suggesting this isn't a bubble but a reflection of real market dynamics.

• •AI is a new computing platform—compare winner count to all businesses built on computers, not just internet companies
• •Internet era produced multiple huge winners (Meta, Netflix, Amazon, Google)—AI will likely produce more
• •Customer adoption and revenue growth rates are unprecedented in technology history
• •Valuations have risen fast, but underlying demand has risen equally fast—not typical bubble dynamics
• •Even Nvidia multiples aren't historically outrageous given growth rates and earnings size
• •Expect more companies worth $1B+ and $10B+ than in previous technology eras