Do institutional investors invest efficiently? To study this question I combine a novel dataset of over 16,000 startups (representing over $9 billion in investments) with machine learning methods to evaluate the decisions of early-stage investors. There is an inference problem because I only observe outcomes for the subset of startups that are funded; I address this one-sided inference problem by quantifying the returns forgone by investing in a given startup (observed) relative to an investment available on a public market (counterfactual). By comparing investor choices to an algorithm’s predictions, I show that up to half of the investments were predictably bad—based on information known at the time of investment, the predicted return of the investment was less than readily available outside options. The cost of these poor investments is 1,000 basis points, totaling over $900 million in my data. I provide suggestive evidence that over-reliance on the founders’ background is one mechanism underlying these choices. Together the results suggest that high stakes and firm sophistication are not sufficient for efficient use of information in capital allocation decisions.