Flapping Planets raises $1 billion: Founder hatches cursor, wants to rewrite AI training from GPU bottom Beating AI News: Flapping Planets, an AI research laboratory that has been established for about a year, has completed a $1 billion Series A funding round led by Index Ventures and Lux Capital. The post investment valuation reached $6 billion, four times that of January this year. The company has only over 20 employees and has not yet launched any commercial products, but investors include Sequoia Capital, Google GV, and NVIDIA. Flapping Planets was co founded by Ben Spector, brother Asher Spector, and Aidan Smith. During his time at MIT, CEO Ben co founded the startup incubator Prod and participated in the early incubation of Cursor, Mercor, Etched, and Decart. Later, he joined Chris R é's research team at Stanford University and became the first author of a paper on the open-source GPU programming framework ThunderKittens. Asher holds a Ph.D. in Statistics from Stanford, while Aidan is a Thiel Fellow who worked for three years at Neuralink, a subsidiary of Elon Musk. In 2025, Ben and the Stanford team further developed Megakernel, attempting to integrate the computation of the entire model into one GPU core. Traditional reasoning frameworks break down calculations into hundreds of cores, starting, executing, and waiting one by one, and GPUs cannot fully utilize the bandwidth of memory during these gaps. Megakernel merges the calculation steps to advance the next step of data reading. When running Llama-3.2-1B on a single H100, vLLM and SGLang can only utilize up to 50% of the video memory bandwidth each time a sequence is generated. The new scheme has increased the utilization to 78%, and the inference speed is more than 1.5 times faster than SGLang. A few months later, the team extended Megakernel to Llama-70B, and in specific tests, the end-to-end throughput after integrating Tokasaurus inference engine was over 22% higher than SGLang. Both studies are open source and belong to the Stanford research results that Ben participated in before founding Flapping Planets. Flapping Planets is now continuing to transform GPUs along this route. The company claims that the latest framework can now perform model training, inference, and video games simultaneously in the same kernel. The team also wants to explore new learning algorithms through this. Existing frameworks such as PyTorch excel at performing calculations organized by steps, but find it difficult to efficiently handle certain fine-grained tasks that require a large amount of interleaved execution. GPU itself may not be unable to do it, it's just that existing software is not easy to fully utilize these capabilities. Changing the underlying programming approach may make algorithms that were previously too expensive and difficult to run feasible. They hope to further address the data efficiency issues of AI. Humans do not need to read the entire Internet to learn language, but large models rely on massive training data. There are a large number of ready-made corpora for search and programming, and it is difficult to obtain samples of the same scale in fields such as robot operation and scientific research. Asher previously used the example of a 1000 fold increase in data efficiency to illustrate the economic value that such breakthroughs may bring. If training a model only requires one thousandth of the original data, many industries lacking training samples may also use AI