Sara Hooker examines the historical interplay between hardware and software research and development, concluding that research ideas often succeed by winning the “hardware lottery” – research directions are promoted because they are well-suited to existing software and hardware and not because the ideas underlying these research approaches are superior. Drawing on examples from the “AI winter” in the 1980s and specialized chip development, she posits that the current success of deep learning may be a result of such a hardware jackpot. Due to the high cost and length of time associated with developing new hardware, she speculates that gains from computing may become increasingly uneven and could entrench mainstream research on deep neural networks. She posits that to avoid future hardware lotteries, we need techniques to quantify the opportunity cost of pursuing research ideas that are most compatible with existing hardware and software. She also suggests that software is well-positioned to cut down its own inefficiencies instead of relying on Moore’s law.