A frontier AI lab building real-time world models - the kind where the camera is still rolling, the prompt has barely landed, and the picture has already changed.
It is May 2026, and inside a San Francisco office that smells faintly of GPU exhaust and cold brew, two ex-Israeli intelligence officers are watching a camera feed of themselves. The feed has been redrawn. They are knights. Behind them, the wall has become a cathedral. The whole transformation takes about as long as a single blink, and not even a slow one. This, give or take a frame, is what Decart sells.
Decart is a frontier AI lab. It builds real-time world models, which is the sort of phrase that sounds slightly improbable until you watch one work. The company runs three product lines - Oasis, which generates whole interactive worlds from a prompt; Lucy, which rewrites live video while you are still in it; and DOS, the deeply unglamorous but absolutely essential plumbing that lets these models run on whichever silicon happens to be cheapest this quarter. A week before this profile was written, the company closed a $300 million Series C at a roughly $4 billion valuation. Investors included Nvidia, Andrej Karpathy, and the family that founded Nintendo. It was, as understatements go, a competent week.
The dirty secret of generative video, until about ten minutes ago, was that none of it was real time. You typed a prompt. You waited. You got back a five-second clip, rendered well after you stopped caring. The models were beautiful. The pipeline was a film school in slow motion.
Meanwhile, the world had moved on. Twitch streamers wanted to swap costumes mid-broadcast. Game studios wanted worlds that responded to a player's eye-line. Advertisers wanted to drop a new product into a livestream the way a copywriter drops a comma. None of this was technically possible. The phrase 'real-time generative video' was the kind of thing investors smiled politely at, then changed the subject.
Decart's founders, Dean Leitersdorf and Moshe Shalev, looked at the bottleneck and concluded something inconvenient. The issue was not the model. It was the stack. Diffusion models were getting smarter by the month; the hardware around them was still being asked to do impressions of a 2018 graphics card. Somebody, they figured, had to rewrite the floor.
Leitersdorf, who finished a PhD in distributed computing in his mid-twenties and once did a postdoc in Singapore mostly out of curiosity, met Shalev in the Israeli signals intelligence corps. Shalev had taken an unusual route to that uniform - he grew up in Israel's ultra-Orthodox community and joined Unit 8200 the way other people join a band. They started Decart in 2023 with a bet most investors found slightly indecent. Inference, they argued, was about to become the most expensive line item in software. Whoever built the fastest, most hardware-agnostic inference stack would not just sell a tool. They would sell the floor.
Sequoia and Oren Zeev wrote the first $21 million check. Benchmark followed. Less than a year later, the company was valued at $3.1 billion. The phrase 'from stealth to billions' is a cliché in this industry and usually a lie. Here it was approximately true.
2023 → 2026 / Three years, four products, one cap table everybody wants to be on.
Decart's product line is unusually legible for an AI company. Most labs ship one model, then a confusingly similar second one, then a third with a Greek letter. Decart ships three things that do three different things and admits it.
The first real-time, interactive generative video model. Famous for generating playable Minecraft from a single prompt and a keyboard. The world is invented frame by frame as you walk through it.
Real-time video-to-video, 1080p at 30fps. Swap characters. Swap weather. Swap an entire room. The stream keeps going while you change its mind.
Live-Stream Diffusion. The first model to do infinite, real-time, low-latency restyling of a video feed. Sub-40ms is not a typo.
The optimization stack underneath everything. Lets Decart's models run on Nvidia, AMD, or Amazon's chips with roughly equal interest. The product investors actually pay for, even if they came for the demos.
It is a curious feature of frontier AI that the most persuasive proof is rarely a slide. It is a demo somebody films on their phone. Decart leans into that. Oasis 2.0 ships as a Minecraft mod with the command /oasis start. Type a prompt - 'Venice', 'candy land', 'cyberpunk Tokyo' - and the world you were just standing in is, frame by frame, replaced with the one you asked for. The game has not stopped. The model has just decided what comes next.
On the enterprise side, the proof is less photogenic and considerably more profitable. Streaming platforms, game studios and broadcast advertisers use Decart's API to do things that used to require an animation team and a quarter. CoreWeave demoed Lucy 2.0 live at Nvidia GTC. Amazon is running world models on its own chips through DOS. The open-source release of Lucy Edit was, within a week, being called 'nano banana for video' by the kind of people who name things like that.
Read Decart's about page and you'll see the phrase 'infrastructure layer for the next generation of low-latency AI.' Translated: the company believes that the cost of generating a frame, a token, or a world is about to fall off a cliff, and they intend to be holding the rope. Everything else - the models, the games, the cameras - is a downstream application of that one idea.
It is a mission with a delicious tension inside it. Decart sells consumer-grade magic - prompts that turn cameras into cathedrals - on top of an infrastructure layer that is, by design, indifferent to which GPU you bought. The magic is what gets investors in the room. The indifference is what keeps them there.
The knights and the cathedral on the office wall are still there. So is the camera. What has changed, in the two years since this whole thing was a slide deck and a thesis, is what 'real time' is allowed to mean. It used to mean 'fast enough.' Now, in Decart's hands, it means 'before you noticed.'
That distinction is small in description and large in consequence. If a video can be rewritten before the eye catches the original, then film, livestreaming, advertising, gaming and surveillance all have a new floor to stand on. The cost of generating a frame is no longer a budget line. It is a background process. The cost of inventing a world is no longer a production schedule. It is a sentence.
Whether that ends well is a separate question, and an interesting one, and not really Decart's to answer. What they have done, and what the $4 billion is mostly paying for, is the part that comes first: making it possible at all. The knights are still on the wall. The cathedral is still behind them. The model is already deciding what they will look like next.
Decart's demos are best experienced live. A few starting points -