The missing piece in physical AI has never been the robot hardware. It's been the environments robots learn in, synthetic, scalable, and realistic enough that behaviors trained in simulation transfer cleanly to the physical world. Decart's Oasis 3, released June 10, 2026, is the clearest attempt yet to solve that problem at commercial scale. The company generates photorealistic driving and physical environments on demand, in real time, via API, at a price point that makes large-scale simulation economically feasible for the first time. At $0.02 per second of generated environment, the math of synthetic training data has permanently changed.
What Actually Happened
Decart launched Oasis 3 on June 10, 2026, positioning it as an interactive world model capable of generating photorealistic multi-camera driving environments in real time for autonomous vehicle development and physical AI training. The system produces synchronized front-facing and two side-camera outputs simultaneously, covering the multi-view geometry that self-driving and robotics systems require for spatial reasoning. According to TechCrunch, Oasis 3 operates at approximately 8,000 tokens per frame at tens of frames per second using an auto-regressive architecture that generates environments one frame at a time, a design choice that prioritizes temporal consistency over raw generation speed and produces fewer visual artifacts during fast scene changes than previous approaches.
The commercial availability is the key development. Oasis 3 is accessible via API at $0.02 per second of generated environment, with enterprise pricing negotiated separately for large-scale deployments. That pricing is made possible by the company's DOS (Decart Optimization Stack), a vertically integrated inference software layer that Decart claims runs its models "more than an order of magnitude cheaper than anyone else in the industry." Dataconomy reported that DOS 2.0, released alongside Oasis 3, delivers the throughput necessary to make real-time interactive environment generation financially viable for commercial AV programs, a benchmark that earlier world models from competitors failed to clear because inference costs rendered them too expensive for large-scale training use.
The launch comes weeks after Decart closed a $300 million funding round at a nearly $4 billion valuation, with strategic investors that include Nvidia, Toyota Ventures, Adobe Ventures, and eBay Ventures, alongside personal participation from Andrej Karpathy, formerly of Tesla and OpenAI. CEO Dean Leitersdorf told Phandroid that the company has "burned through drastically less than $100 million" in lifetime spending, a capital efficiency claim that, if accurate, would make Decart one of the most efficient frontier-tier AI companies operating today. Toyota's participation as a strategic investor signals that at least one major automotive manufacturer believes Oasis 3's simulation fidelity is close enough to production quality to matter for their own AV programs.
Why This Matters More Than People Think
The simulation gap has been one of the persistent problems in physical AI development. Models trained exclusively on real-world data run into fundamental coverage issues: rare scenarios, edge cases, and adverse weather conditions simply don't appear in collected datasets with sufficient frequency to train robust behavior. Synthetic data can generate those scenarios at scale, but synthetic data quality has historically degraded model performance when real-world transfer is attempted, the notorious "simulation-to-real gap" that has haunted robotics research for decades. Oasis 3's photorealism claim targets that specific problem. If a world model generates environments indistinguishable from real sensor data, the transfer gap narrows, and the entire economics of physical AI training shift.
The implications extend well beyond autonomous vehicles. Decart has stated that robotics is the next planned expansion of the Oasis architecture, the same world model that simulates photorealistic driving environments can, in principle, simulate warehouse logistics, surgical robotics environments, and the physical household settings that domestic robot developers are attempting to crack. The autonomous driving market gave world models their first commercial proving ground because AV programs generate large budgets and have clear, measurable performance requirements. But the underlying technology is domain-agnostic. A model that can generate photorealistic sensor data for a car navigating an intersection can, with appropriate training data, generate photorealistic sensor data for a humanoid robot navigating a kitchen counter.
However, skeptics point out that the "photorealistic" benchmark is not the same as the "training-useful" benchmark. TechCrunch's coverage noted several caveats in Oasis 3's current output, temporal consistency issues during fast camera movements, artifacts under high-brightness conditions, and a tendency toward mean-regression in rare-event scenarios, exactly the scenarios synthetic data is most needed to generate. The company's response is that these are engineering problems actively being addressed rather than architectural limitations. But the risk is that enterprise AV customers pay for photorealistic synthetic data that doesn't transfer cleanly to production training pipelines, an outcome that would slow the technology's commercial adoption by years.
The Competitive Landscape
Decart is not operating in an uncrowded market. Google's Genie 3 world model, World Labs' Marble platform backed by Fei-Fei Li, and a cohort of generative video companies including Runway and Kling are all competing for the physical AI simulation market. The differentiation Decart claims rests on two factors: real-time interactivity and inference cost. Genie 3 can generate physically plausible environments, but the interaction latency at scale makes it better suited for offline training data generation than real-time agent evaluation. Runway and Kling produce high-quality video but lack the physics consistency needed for robotics and AV applications where spatial reasoning properties must be preserved across frames. Decart's architecture choices, the auto-regressive frame-by-frame generation and the vertically integrated inference stack, are explicitly designed to win on those two dimensions.
The investor roster gives a structural read on where Decart is positioned in the value chain. Nvidia's participation is the most strategically pivotal: Decart's DOS stack runs on Nvidia hardware, and Nvidia's interest in making physical AI simulation economically accessible aligns directly with the company's broader strategy of expanding AI compute demand into physical world applications. Toyota's stake signals AV customer validation. Andrej Karpathy's personal investment is a technical endorsement from the person most responsible for Tesla's synthetic data strategy, a strategy that made Tesla's Autopilot system one of the most data-efficient large-scale robotics programs ever built. Together, these investors represent the infrastructure layer, the deployment layer, and the technical credibility layer that a simulation platform needs to reach commercial scale.
The historical parallel is NVIDIA's acquisition of Mellanox in 2020 to control the networking stack for AI compute, a vertical integration move that seemed niche at the time and turned out to control a critical bottleneck. Decart's DOS stack is an analogous bet: if inference cost is the binding constraint on physical AI simulation adoption, and if DOS is genuinely an order-of-magnitude more efficient than competitive inference solutions, then Decart has built control over a bottleneck that every AV program and robotics company will eventually need to pass through. The competitive risk is that Nvidia, having seen this play before, builds an equivalent stack internally and bundles it with the GPU infrastructure at no separate charge.
Hidden Insight: The Trillion-Token Problem in Physical AI
The large language model revolution was driven, in substantial part, by the availability of near-unlimited text training data on the internet. Physical AI faces a fundamentally different data problem: the physical world generates sensory data continuously, but that data is expensive to collect, difficult to annotate, and concentrated in the specific environments where robots and vehicles are deployed, not in the tail distribution of rare events where model failures actually occur. Oasis 3's commercial significance is that it offers, for the first time at scale, a market solution to the physical world data problem that doesn't require sending fleets of robots or vehicles into dangerous conditions to collect edge-case data.
The numbers make this concrete. An autonomous vehicle program running safety testing on real roads generates roughly 30 to 50 terabytes of sensor data per vehicle per week. A fully photorealistic world model that can generate equivalent data synthetically at $0.02 per second of simulation could produce the same data volume at a fraction of the cost, with complete control over scenario distribution. If an AV program needs 10,000 hours of intersection-crossing data in heavy rain at night, a scenario that real-world collection takes months or years to accumulate, a simulation system can generate it in hours. The training implications compound: models trained on synthetically augmented datasets that specifically oversample rare scenarios should perform better in those scenarios than models trained purely on collected data.
The broader physical AI infrastructure picture positions Decart as a potential infrastructure layer rather than just a simulation product. The company's 100,000+ developer community on its Lucy real-time video model has established an ecosystem of developers building on Decart's APIs before Oasis 3's commercial launch. The transition from developer experimentation with video generation to commercial AV simulation is not guaranteed, but the installed base of developers who already know the Decart API represents a distribution advantage that competing world model platforms don't currently match. Decart's founders, Dean Leitersdorf and Moshe Shalev, were previously researchers at Google Brain, which means the company's architecture choices reflect awareness of Google's own world model research limitations, a useful competitive intelligence source for knowing exactly where to differentiate.
What Oasis 3 ultimately represents is a bet on the proposition that the physical world's complexity is learnable from video at sufficient scale. The auto-regressive generation approach Decart uses is the same architectural principle that made GPT-4 possible for language, sequential prediction of the next token. Applied to video frames rather than text tokens, the same principle should, in theory, produce world models capable of learning the physical constraints of the real world from observation alone. Whether that theory translates into training-useful simulation at the quality bar enterprise AV programs require is the central open question. The $300 million Decart raised, and the Toyota and Nvidia endorsements attached to it, suggest the market has decided to find out at commercial scale.
What to Watch Next
The 30-day signal is customer disclosure. Decart has stated it is targeting autonomous vehicle companies as the initial commercial segment. Watch for any named AV program announcements, a formal partnership with a Toyota subsidiary, a Waymo API integration, or a disclosure from a Chinese AV startup, that would validate enterprise adoption rather than developer experimentation. The absence of named commercial customers in the launch announcement is the most glaring caveat in the otherwise bullish market reception; world model technology has a history of impressive demos that don't survive contact with the specific requirements of production training pipelines.
The 90-day question is whether the robotics expansion Decart has telegraphed arrives on schedule. The company has stated that new versions of Oasis targeting physical AI beyond autonomous driving are planned for the coming weeks. If Oasis 3's robotics version can simulate the manipulation tasks that humanoid robot developers at Figure AI, Agility Robotics, and Boston Dynamics are training against, grasping, placement, and bimanual coordination in diverse physical environments, it would represent a genuine expansion of the addressable market and a direct challenge to NVIDIA's own Cosmos physical AI simulation platform, which was launched June 9 with overlapping ambitions but different architectural choices.
At 180 days, the question is whether DOS 2.0's cost advantage survives the infrastructure buildout it will attract. Decart's claim of order-of-magnitude inference efficiency relative to competitors is a target painted on the company for every major inference provider, AWS, Google Cloud, and Nvidia's own inference platform, to shoot at. If cloud providers match or approach Decart's cost efficiency through hardware-level optimizations for auto-regressive video generation, the DOS advantage erodes and Decart competes primarily on model quality rather than infrastructure economics. The company's capital efficiency record suggests it can run lean in a commodity market, but a pricing war with three of the world's largest cloud operators is a different kind of lean than a startup that has spent under $100 million lifetime.
The physical world's training data problem is not a hardware problem or a robot design problem, it's a simulation quality problem, and Decart has just priced it at two cents per second.
Key Takeaways
- $0.02 per second, API-accessible — Oasis 3 is the first commercial world model for physical AI simulation priced for large-scale AV and robotics training use rather than research experimentation
- Order-of-magnitude cost advantage claimed — Decart's DOS 2.0 inference stack underpins a cost efficiency claim that, if verified at enterprise scale, would restructure how AV programs budget for synthetic training data
- $300 million at $4 billion valuation, Nvidia and Toyota backing — Strategic investors from the GPU infrastructure layer and the automotive deployment layer signal genuine commercial validation rather than speculative interest
- Auto-regressive architecture trades speed for consistency — The frame-by-frame generation approach produces fewer temporal artifacts than diffusion-based competitors but faces open questions about rare-event fidelity in production training pipelines
- Robotics expansion is the next gate — Decart's stated plan to extend Oasis beyond driving simulation into humanoid robotics training would put it in direct competition with NVIDIA Cosmos and determine whether the platform is AV-specific or physically general
Questions Worth Asking
- If synthetic training data generated by world models can match real-world sensor data in quality, does physical AI development become primarily a software and simulation problem rather than a hardware and data collection problem?
- What happens to the competitive moat of AV companies that have invested billions in proprietary real-world data collection fleets if that data can be replicated synthetically at $0.02 per second?
- Should the physical AI simulation market be treated as critical infrastructure requiring open standards, or will allowing one or two companies to control photorealistic simulation create a new platform dependency for the entire physical AI industry?
