Decart AI

Decart's current product architecture has three connected layers. DOS is the systems layer: hardware-aware model design, inference optimization, kernel and compiler tooling, and productization intended to improve the economics and latency of AI workloads across NVIDIA GPUs, Google TPUs, and Amazon Trainium. Lucy is the immersive-experience layer, generating and editing video while it plays, with applications in live streaming, commerce, advertising, gaming, and virtual try-on. Oasis 3 is the physical-AI layer: an API-accessible interactive world model that generates controllable environments, responds to robot or vehicle actions, and produces synchronized multi-camera views. This is a more substantive proposition than a conventional text-to-video application because the target loop is continuous perception, action, and feedback.

The technical importance is systems-level as much as generative-model-level. Real-time video and world models must maintain temporal consistency, accept control signals, handle long-running sessions, and meet predictable latency and cost targets. Decart says Oasis 3 can operate below 200 milliseconds end to end at 22 frames per second in a stated configuration, while its May 2026 company announcement claims materially higher throughput for selected DOS workloads. These are company-reported performance claims, not independent benchmarks, but they identify a real bottleneck: impressive offline generation is not enough for a simulator, interactive product, or agent that must respond on every step. The potential edge is the co-design of models, compilers, kernels, and deployment infrastructure, which can be harder for an application-only competitor to reproduce.

Commercially, Decart appears to be moving from viral consumer demonstrations toward infrastructure and enterprise distribution. The company offers APIs and documentation, positions DOS for external workloads, and says it has revenue-generating contracts with cloud providers, AI labs, and hyperscalers. Its May 2026 release names Amazon as a strategic customer and describes a commercial and go-to-market collaboration around AWS Trainium; those statements should be validated in diligence through contract scope, recurring revenue, customer concentration, and independently verifiable production usage. Earlier reporting documented a $21M seed round and a $32M Series A, while Decart now reports a $300M round led by Radical Ventures and more than $450M raised in total. The scale of financing and named strategic backers are meaningful traction signals, but they also raise the execution bar: the company must convert expensive technical capability into durable software revenue rather than relying on capital-intensive model demonstrations.

The competitive field spans several categories. Runway, Luma, and Pika compete for generative-video users and developer attention; Google, OpenAI, and other frontier labs can fold video or world-model capabilities into broader platforms; NVIDIA, AWS, and cloud infrastructure vendors can absorb optimization features into their own stacks. In simulation, CARLA, NVIDIA Isaac Sim, Omniverse, and specialized robotics platforms offer more explicit physics, tooling, and validation workflows. Decart's differentiation is its attempt to combine photorealistic generative environments, action-conditioned interaction, synchronized views, and low-latency deployment. That can unlock rare-event generation and flexible scenario authoring, but a learned world model is not automatically a trustworthy physics simulator. Customers will need evidence of behavioral fidelity, reproducibility, calibration, safety, and transfer from synthetic training to real systems.

The national-security and defense relevance is credible at the capability level, especially for autonomous-vehicle testing, drone and maritime scenario generation, mission rehearsal, sensor and perception evaluation, and training environments where dangerous or rare conditions are difficult to collect in the real world. It is not yet a confirmed defense business in this record: public materials identify robotics, autonomous vehicles, enterprise, media, and cloud relationships, but do not establish government contracts, classified deployments, or validated military performance. Strategic value therefore comes from the possibility that DOS and Oasis become enabling infrastructure for physical AI across commercial and security users, not from an assumed defense revenue stream. Diligence should focus on deployment sovereignty, data rights, export controls, secure isolation, failure behavior, scenario provenance, and whether defense users can obtain deterministic and auditable results from a probabilistic generative system.