Flapping Airplanes raises $1 billion to train AI on less data

Today’s leading models swallow much of the public internet in training. Flapping Airplanes wants models that need a small fraction of that, which would make them cheaper both to build and to use.

So far the team has worked on the graphics chips (GPUs) themselves. Its megakernel framework packs a chip’s work into one program, so a single GPU can train a model, serve users and run a video game at once. The trick is mixing work types: while one task stalls waiting on memory, another keeps the chip’s maths busy, TechCrunch reports.

In January a $180 million seed round valued it at $1.5 billion; nine months on, the price is four times higher. Nvidia, Kleiner Perkins, GV and Sequoia also back the lab. Brothers Ben and Asher Spector, both Stanford doctoral researchers, founded it with Aidan Smith, a Thiel Fellow who worked at Neuralink.

A lab with no product now carries a $6 billion price tag on the promise of cheaper AI.