Key Takeaways

  • Google Ventures, under Bill Maris and Rich Miner, ditched intuition for a "machine learning" (ML) approach to venture investing.
  • They built their strategy on a trove of historical venture data, using ML to run millions of simulations for ideal portfolio construction.
  • This data-driven method determined the optimal fund size for Google Ventures, estimating returns at 4.1x.
  • Maris's third lesson: “Don't bet against computer science.” Applying the right tech to the right problem delivers favorable outcomes.

The Method: Engineering a Venture Fund

When Bill Maris and Rich Miner set out to create Google Ventures, they faced the same challenge every venture capitalist does: how to build a portfolio that actually works. Instead of relying on gut feel or industry dogma, they approached it like a deep engineering problem. Maris explained, “Our plan was to obtain all the data of venture that we could find. And being Google, you can imagine it was a lot of data, historical data, you name it.” This wasn't just collecting; it was weaponizing information.

They didn't just analyze; they simulated. “[W]e used machine learning to do two things,” Maris clarified. First, to “design the ideal portfolio construction by running millions and millions of simulations... and back testing and all of the things you can imagine that data scientists would do.” Second, this ML also helped them determine the ideal fund size. The result? A strategically constructed fund with an estimated return of "about 4.1x." This rigorous, data-first approach became the bedrock for Maris's third lesson: "Don't bet against computer science. I've seen it happen many many times in many many fields. If you apply the right kind of computer science at the right time to the right problem, you will get to the right answers."

Where This Breaks Down

While Google Ventures' data-driven success is compelling, its context matters. Most founders and investors won't have Google's unique access to “all the data of venture” or the internal ML talent to build such sophisticated simulation engines. This method works best when historical data is plentiful, structured, and representative of future outcomes. For truly nascent markets, pre-product-market fit startups, or highly qualitative decisions, the data might be too sparse, noisy, or biased to provide a definitive "right answer."

Moreover, an over-reliance on historical data can blind you to paradigm shifts. The computer science that worked yesterday might not apply to tomorrow's problem. Human intuition, qualitative analysis, and a willingness to explore unquantifiable ideas still play a role, especially when you are building something genuinely new that has no historical precedent.

What to Do With This

Don't just admire Google's scale; steal the mindset. Identify one critical, repeatable problem in your business where you have (or can collect) data. Are you trying to optimize customer acquisition channels? Predict churn? Prioritize product features? Instead of guessing or going with the loudest voice, ask: how can you apply "computer science" to simulate outcomes and find a more optimal answer? Pull your customer data this week. Segment users by behavior. Can you run even simple A/B tests to inform your next marketing spend, or use basic regression to predict which customers are about to leave? Find your version of the "millions and millions of simulations" to get closer to a 4.1x return on your effort.