Microsoft invests $1 billion in OpenAI

Training a large model is mostly a bill. The research is cheap next to the computers, and the computers had grown more expensive with every increase in model size. OpenAI had been set up as a nonprofit, and by 2019 that structure had become its binding constraint. In March the laboratory answered it by creating OpenAI LP, a capped-profit company under the control of its nonprofit board, able to take investment in a way a charity cannot.

Four months later it took some. On 22 July 2019 Microsoft and OpenAI announced a multiyear partnership: $1 billion from Microsoft, paid in installments, Microsoft as OpenAI’s exclusive cloud provider, and a joint effort to build what the companies called AI supercomputing technology on Azure. Both said the purpose was to develop artificial general intelligence whose benefits are broadly distributed, and Microsoft was named the preferred partner for commercializing what the work produced.

The shape of the deal mattered as much as the number. The exclusivity ran in both directions. OpenAI would train and serve its systems on one company’s computers, and that company would have first claim on the products. The arrangement paired a laboratory that needed computing at a scale only a handful of firms own with one of the firms that owns it, and the money moved in a circle: capital in, cloud bills out.

For OpenAI it removed a ceiling. The models that followed were far larger than anything it had trained before, and the products built on them, including ChatGPT three years later, ran on the infrastructure this partnership paid for.