Flapping Airplanes Raises $1B for Single-Kernel AI Training

Roughly a year after its inception, AI research firm Flapping Airplanes has closed a $1 billion Series A financing round co-led by Index Ventures and Lux Capital, alongside Google Ventures, Sequoia Capital, and Menlo Ventures. The deal assigns the neolab a $6 billion post-money valuation, quadrupling its $1.5 billion evaluation from a $180 million seed round nine months earlier in January. Founded by Stanford doctorate Ben Spector, his brother Asher Spector, and former Neuralink engineer Aidan Smith, the company aims to collapse entire training runs into a single megakernel on a GPU.

Modern machine learning systems routinely trigger thousands of separate kernel launches across training steps for operations like attention or matrix multiplication. Flapping Airplanes builds on Ben Spector's previous work with Stanford's Hazy Research group on ThunderKittens—a domain-specific language for optimized GPU operations. The startup is developing a "Kittens" virtual machine that acts as a primary surface for scheduling workloads natively on hardware. While existing software tools like CUDA Graphs, PyTorch Inductor, Triton, and XLA offer selective operation fusion, Flapping Airplanes claims its low-level primitives keep full training passes inside one kernel while simultaneously serving inference demands or running external tasks.

Despite the immense valuation, Flapping Airplanes currently has no public product, open-source releases, named customers, or published benchmarks. The broader megakernel approach is already gaining traction elsewhere: the open-source Luce Megakernel project recently executed all 24 layers of Qwen 3.5-0.8B within a single CUDA launch, achieving 1.87 generated tokens per joule on a 2020 GPU. For machine learning engineers, proven megakernel frameworks could drastically improve hardware utilization and reduce energy costs by eliminating host-system handoffs. However, long-running kernels introduce notable friction, including increased memory pressure, slower compile times, and complicated cross-GPU fault recovery.