Flapping Airplanes raises $1 billion: Founder incubated Cursor, aims to rewrite AI training from the GPU ground up

Beating AI News Flash: AI research lab Flapping Airplanes, founded about a year ago, has completed a $1 billion Series A funding round, co-led by Index Ventures and Lux Capital. The post-money valuation reached $6 billion, four times that of January this year. The company has only 20-plus employees and has not yet launched a commercial product, yet its investors include Sequoia Capital, Google GV, and Nvidia.

Flapping Airplanes was co-founded by Ben Spector, his older brother Asher Spector, and Aidan Smith. CEO Ben co-founded the startup incubator Prod during his time at MIT and participated in the early incubation of Cursor, Mercor, Etched, and Decart. Later, he joined Chris Ré's research team at Stanford University and became the first author of the paper on the open-source GPU programming framework ThunderKittens. Asher holds a PhD in Statistics from Stanford, and Aidan is a Thiel Fellow who worked at Elon Musk's Neuralink for three years.

In 2025, Ben and the Stanford team further developed Megakernel, attempting to fuse the entire model's computation into a single GPU kernel. Traditional inference frameworks break computation into hundreds of kernels, launching, executing, and waiting on each one individually, and the GPU cannot fully utilize memory bandwidth during these gaps. Megakernel merges computation steps, allowing the next step's data reads to proceed in advance. When running Llama-3.2-1B on a single H100 and generating one sequence at a time, vLLM and SGLang can utilize at most 50% of memory bandwidth, while the new approach raised this to 78%, with inference speed more than 1.5 times faster than SGLang.

A few months later, the team extended Megakernel to Llama-70B. In specific tests, end-to-end throughput after integrating the Tokasaurus inference engine was more than 22% higher than SGLang. Both studies have been open-sourced and are Stanford research results Ben participated in before founding Flapping Airplanes.

Flapping Airplanes is now continuing to transform GPUs along this path. The company says its latest framework can already execute model training, inference, and video games simultaneously within the same kernel. The team also wants to use this to explore new learning algorithms. Existing frameworks such as PyTorch are good at executing step-organized computation but struggle to efficiently handle certain fine-grained tasks that require extensive interleaved execution. The GPU itself may not be incapable; it is just that existing software cannot easily bring out these capabilities. Changing the underlying programming approach could make algorithms that were previously too costly and difficult to run feasible.

They aim to further address the data efficiency issue in AI. Humans don't need to read the entire internet to learn a language, yet large models rely on massive training data. While there is abundant ready-made corpus for search and programming, fields like robotic manipulation and scientific research struggle to obtain samples of comparable scale. Asher previously cited a 1,000-fold improvement in data efficiency as an example to illustrate the economic value such breakthroughs could bring. If training a model requires only one-thousandth of the original data, many industries lacking training samples could also adopt AI.