Decart is an AI research lab building real-time world models and the inference and training infrastructure that runs them. The company ships three connected products: the Decart Optimization Stack (DOS), an engine for low-latency inference and training across GPUs, TPUs and AWS Trainium; Lucy, a "world model for immersive experiences" that responds to user input in under 30 milliseconds and is now deployed across virtual try-on, live streaming, and dynamic in-video advertising ; and Oasis, a world model for physical AI that generates interactive, physically accurate, real-time simulations . With DOS 2.0, the team says Lucy and Oasis run at full HD at 100 frames per second, which is what makes the live try-on and on-the-fly scene transformation demos feel instant rather than rendered.
The launch marks Decart's Series B, a $300 million round led by Radical Ventures that brings total funding to more than $450 million. NVIDIA, Adobe Ventures, Toyota Ventures and eBay Ventures joined, alongside angels including Andrej Karpathy and Michael Eisner. Several of those names are also customers, which is the more interesting signal here: Decart is positioning itself as both a model lab and an infrastructure vendor to the companies it sells into, with stated traction in e-commerce, live streaming, gaming, and early work toward robotics and manufacturing.
The company was founded in 2023 by Dean Leitersdorf, who serves as CEO, and Moshe Shalev, who serves as CPO , both with roots in the Israeli technical ecosystem. For founders and operators watching the world-model space, this round is worth noting because it funds a vertically integrated bet (chips, compilers, models, applications) rather than a pure research effort, and because the customer list suggests the real-time video models are already being used in production rather than just shown in demos.
Comments (8)
Sign in to join the discussion.
Y
Yusra OkaforMay 22, 2026
Real-time world models is a phrase that does a lot of heavy lifting. I'll believe the 'continuously' part when someone leaves a demo running for 8 hours without it melting into vaseline.
T
Tomasz KwiatkowskiMay 22, 2026
The launch graphic has way too much going on for a Series B post. You raised a massive round, let the number breathe, you don't need three gradients fighting it.
M
Marisol VegaMay 22, 2026
Two core pillars and the tweet cuts off right before telling us what they are. Cliffhanger copywriting on a Series B announcement is a bold move.
P
Priya MenonMay 22, 2026
Any plans to publish on the diffusion-vs-autoregressive tradeoffs for your frame generation? Curious if you're closer to the Genie line of work or doing something stranger.
Z
Ziv HaleviMay 22, 2026
If it's frame-level autoregressive with a learned latent, the interesting question is how they avoid drift past ~30s. The Genie-style work hit a wall there.
K
Kenji AboagyeMay 22, 2026
Genuine question: any of this getting open-weighted or is it all behind an API? Asking for the dozens of indie game devs already trying to hack on world models.
R
Raffi T.May 22, 2026
Reminds me of when our portfolio team at Hedra was scaling video gen, same kind of compute curve. Decart's been heads down for a while, this round tracks.
T
Tomi AdesanyaMay 22, 2026
Every 'path to AGI' tweet adds six months to my AGI timeline. The actual product looks cool, the framing is doing the opposite of selling me.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Decart
Designing real-time world models that run instantly, continuously, and efficiently
Designing real-time world models that run instantly, continuously, and efficiently
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Decart is an AI research lab building real-time world models and the inference and training infrastructure that runs them. The company ships three connected products: the Decart Optimization Stack (DOS), an engine for low-latency inference and training across GPUs, TPUs and AWS Trainium; Lucy, a "world model for immersive experiences" that responds to user input in under 30 milliseconds and is now deployed across virtual try-on, live streaming, and dynamic in-video advertising ; and Oasis, a world model for physical AI that generates interactive, physically accurate, real-time simulations . With DOS 2.0, the team says Lucy and Oasis run at full HD at 100 frames per second, which is what makes the live try-on and on-the-fly scene transformation demos feel instant rather than rendered. The launch marks Decart's Series B, a $300 million round led by Radical Ventures that brings total funding to more than $450 million. NVIDIA, Adobe Ventures, Toyota Ventures and eBay Ventures joined, alongside angels including Andrej Karpathy and Michael Eisner. Several of those names are also customers, which is the more interesting signal here: Decart is positioning itself as both a model lab and an infrastructure vendor to the companies it sells into, with stated traction in e-commerce, live streaming, gaming, and early work toward robotics and manufacturing. The company was founded in 2023 by Dean Leitersdorf, who serves as CEO, and Moshe Shalev, who serves as CPO , both with roots in the Israeli technical ecosystem. For founders and operators watching the world-model space, this round is worth noting because it funds a vertically integrated bet (chips, compilers, models, applications) rather than a pure research effort, and because the customer list suggests the real-time video models are already being used in production rather than just shown in demos.
Comments (8)
Real-time world models is a phrase that does a lot of heavy lifting. I'll believe the 'continuously' part when someone leaves a demo running for 8 hours without it melting into vaseline.
The launch graphic has way too much going on for a Series B post. You raised a massive round, let the number breathe, you don't need three gradients fighting it.
Two core pillars and the tweet cuts off right before telling us what they are. Cliffhanger copywriting on a Series B announcement is a bold move.
Any plans to publish on the diffusion-vs-autoregressive tradeoffs for your frame generation? Curious if you're closer to the Genie line of work or doing something stranger.
If it's frame-level autoregressive with a learned latent, the interesting question is how they avoid drift past ~30s. The Genie-style work hit a wall there.
Genuine question: any of this getting open-weighted or is it all behind an API? Asking for the dozens of indie game devs already trying to hack on world models.
Reminds me of when our portfolio team at Hedra was scaling video gen, same kind of compute curve. Decart's been heads down for a while, this round tracks.
Every 'path to AGI' tweet adds six months to my AGI timeline. The actual product looks cool, the framing is doing the opposite of selling me.