Luma AI raises $900M Series C at $4B+ valuation and partners with HUMAIN to build 2-gigawatt compute cluster in Saudi Arabia

Luma AI closed a $900M Series C at a $4B+ valuation and announced a partnership with HUMAIN, the Saudi Arabia-based AI company, to build a 2-gigawatt compute cluster. President Trump and Crown Prince Mohammed bin Salman announced the deal at the US-Saudi Investment Forum.

The 2GW cluster is predominantly built for inference, not just training. Luma's current inference-to-training compute ratio is already 2-to-1, and the company expects that gap to widen as AI workloads shift from text to video. Video generation and understanding are computationally far more intensive than language. Luma's CEO argues that as multimodal AI matures, the majority of all tokens and compute cycles will move toward video.

“Luma raised a 900 million Series C. Along with Humain, which is an AI company being built in Saudi Arabia, we are building a 2 gigawatt compute cluster that we're going to use to train multimodal AGI... Luma's inference to training compute ratio is 2 to 1 already and we're seeing that ramp growing further. By end of 27 or early 28 we will have majority of the capacity at hand.”

Luma's multimodal thesis for AGI

Luma's cluster strategy rests on a specific view of what AGI requires. Text should not be abandoned because it provides the interpretive layer that makes reasoning coherent. But models need to unify audio, video, image, and language together. Physics and causality come from video. Logic and reasoning come from text. Dropping either breaks the system.

The more ambitious claim concerns simulation. LLMs are strong at tasks fully representable in text such as code, analysis, and structured reasoning. But physical-world problems like designing jet engines or rocket engines require models that can run physical simulations and derive real answers, not just generate plausible-looking outputs. If models can accelerate that kind of complex systems design, the timeline for building physical infrastructure compresses significantly. Generative world models, in this framing, become the simulation engine that makes that possible.

Deployment timeline

Some capacity is already live and Luma is an active customer today. The cluster is located in Saudi Arabia. Full deployment of the 2GW cluster is a multi-year build starting in 2026, with the majority of capacity expected online by end of 2027 or early 2028.