We’ve developed the world’s most complete megakernel framework. It can perform entire training runs, serve inference, play video games, and even do all three simultaneously in a single kernel.

We’ve spent the past year developing new primitives for GPUs. We believe this enables new algorithms to change how machines learn, open new markets, and dramatically reshape AI’s cost structure. We have raised a $1 billion Series A, co-led by Index and Lux.

Our systems work makes training & inference cheaper—but it also makes GPUs more expressive. The same framework that allows us to overlap memory-bound inference with compute-bound training also enables odd architectures, objectives, and optimizers.

Our mission is to unlock new markets in AI, including robotics, scientific discovery, and more, by pioneering radical new techniques to improve efficiency. We imagine a world where models can be intelligent without ingesting half the internet.

We have barely scratched the surface of GPUs, and we think GPUs are more flexible and valuable than most believe. Internally, we already see evidence of this at trillion-parameter pretraining scale. If we are right, this will drive substantially more demand for AI.

If you are interested in helping us abuse GPUs in new and exciting ways, please reach out.

We are tremendously grateful for institutional support from Nova, Sequoia, Valor, Atreides, Kleiner Perkins, Fundomo, BCV, GV, Hummingbird, Hof, Absolute Ventures, A*, CRV, Definition, Seven Stars, Genesis, Jump Trading, XTX Ventures, Snowy Leopard, NVIDIA, and more.

We are also delighted to welcome individual investors including @StevenACohen2, @standuquesne, Henry Kravis, @mickymalka, and others.

in the demo overlay, training gets 116 sms, inference 24 and doom 8. spatial partitioning like that is already possible with mps or green contexts, so the number worth publishing is the gain over those, not over separate stock launches. and a pure kernel win has a ceiling: llama 3 405b trained at 38 to 43% bf16 mfu on h100s per the paper, so even 100% of peak is only about 2.5x. anything beyond that has to come from the algorithm side. what is training mfu on the 116 sms?

Woah

Megakernels, Megakernels, Megakernels!

@MainframeAgent

wat

What?

huge

🚀

✈️

based

CV

Will this have ROCm and MPS support?

training, inference and doom in a single kernel is a flex