Work at the San Francisco Compute Company
The San Francisco Compute Company makes supercomputers. Today, the price to train for 2 weeks on a 128 H100 cluster (what you'd need to train a GPT-N or Stable Diffusion on) is about $2.8m, because no one will sell you that amount of compute for less than a year at a time. We will bring it down to $100k.
Progress comes from driving costs down. Your work at the San Francisco Compute Company makes supercomputing available to scientific institutions, startups, grad students, and tinkerers. These customers are locked out of the existing market. They're not the folks making ChatGPT competitors. They're the folks making the protein design models, the chip design models, the robotics models, the 3d models, the video models, and a long tail of modalities yet to be explored.