Decart Raises $300M for Real-Time AI World Models
Key insights
- Decart has raised over $450M total, with Nvidia and Amazon participating as both investors and paying customers.
- The company's three products target inference acceleration, immersive world models, and physical-AI simulation as distinct verticals.
- Andrej Karpathy's backing positions Decart alongside a small group of world-model startups with top-tier technical validation.
Why this matters
Nvidia investing in a company that accelerates inference on its own hardware creates a closed loop where Blackwell adoption and Decart's growth are structurally coupled, raising the competitive barrier for any rival inference optimization play. The investor-as-customer model adopted by both Nvidia and Amazon signals that hyperscalers are no longer content to wait for market leaders to emerge in world models and are instead buying early positions to shape the stack. For founders in generative video or physical AI simulation, Decart's raise sets a new valuation anchor and signals that the real-time constraint, not just output quality, is now the primary technical differentiator investors are funding.
Summary
Decart, the Tel Aviv-based AI infrastructure company, has closed a $300M funding round at a $4B valuation, pushing its total raised past $450M. The round was led by Radical Ventures, with Nvidia, Amazon, Sequoia, Benchmark, and Andrej Karpathy all participating as both customers and investors.
The company operates across three product lines: DOS, its training and inference acceleration software; Lucy, aimed at immersive world models; and Oasis, targeting physical-AI and robotics simulation. The dual investor-customer structure with Nvidia and Amazon signals that Decart isn't just building for the market, it's being pulled in by the largest infrastructure players who want direct stakes in what they're deploying.
Essentially: (Decart, Nvidia, Amazon) are converging on real-time generative video and interactive AI as the next infrastructure layer.
- Nvidia's participation places Decart inside the Blackwell-era compute stack, not just adjacent to it.
- Andrej Karpathy's involvement adds technical credibility at a moment when world models are moving from research concept to product.
- The three-product architecture suggests Decart is hedging across gaming, simulation, and physical AI rather than committing to a single vertical.
Real-time generative world models are one of the few AI categories where compute speed, model architecture, and application layer must all be solved simultaneously, which explains why the infrastructure giants are investing rather than building internally.
Potential risks and opportunities
Risks
- If Nvidia internalizes world-model acceleration into future Blackwell driver stacks or NIM microservices, DOS faces direct product displacement from its own investor.
- Amazon's dual investor-customer role creates a scenario where AWS could replicate Lucy or Oasis capabilities inside Bedrock, leaving Decart exposed if the commercial agreement lapses.
- A $4B valuation on a company with three early-stage products creates significant down-round risk if any one vertical, particularly physical-AI simulation, fails to reach production deployment within 18-24 months.
Opportunities
- Competing inference optimization vendors (Together AI, Groq, Cerebras) face accelerated customer pressure to demonstrate Blackwell-era performance benchmarks before Decart's DOS establishes a reference standard.
- Robotics simulation platforms (Wayve, Physical Intelligence, 1X Technologies) could partner with or position against Oasis as physical-AI world models move from research to procurement conversations.
- Israeli deep-tech funds and government innovation bodies gain a credible new anchor company for attracting follow-on AI infrastructure investment into the Tel Aviv ecosystem.
What we don't know yet
- Revenue and customer concentration: Decart has not disclosed how much of its commercial traction comes from Nvidia and Amazon versus independent third parties.
- Whether Oasis (physical-AI world models) is shipping to robotics customers or remains in early access as of the announcement date.
- How DOS competes with or complements existing inference optimization layers like TensorRT and vLLM given Nvidia's direct investment in both.
Originally reported by decart.ai
Read the original article →Original headline: Decart Raises $300M at $4B Valuation for Real-Time AI World Models — Nvidia, Andrej Karpathy, Amazon, and Sequoia Among Backers