DeepMind

In two thousand and ten, Demis Hassabis rented a small office in London and hired a handful of researchers to do something with no obvious commercial value: teach computers to play Atari games. No product. No revenue model. No customers. The machine taught itself to play Breakout and within hours found a strategy no human had tried — tunnelling the ball behind the wall to let it bounce endlessly. Google paid five hundred million in two thousand and fourteen for fewer than seventy-five employees and zero revenue. The price baffled everyone. But Hassabis had shown machines could learn without being told the rules. Google ran the largest advertising system on earth — four billion daily users. A one percent improvement in ad relevance was worth two to three billion annually. DeepMind burned over one billion a year on compute and salaries. Google funded every dollar because the research fed directly into Search, YouTube, Cloud, and Android. Better recommendations kept users watching.