Two months ago, we introduced Atlas, an autoregressive world model with pixel-perfect camera control.
Today, we’re releasing Frameboy, a demo allowing you to explore worlds in real-time with a highly optimized version of Atlas, streaming from cloud GPUs at up to 15 fps.
World Labs posted the release at 19:25 UTC on 5 November with a demo video. It says Atlas was introduced two months ago and that Frameboy lets users explore worlds in real time with a highly optimized version of Atlas, streaming from cloud GPUs at up to 15 fps.
Fei-Fei Li, who leads World Labs, quoted the company’s post at 20:21 UTC: โReal time Atlas! Try it now!โ
World Labs, the world-model company led by Fei-Fei Li that AMD agreed in September to acquire in an all-stock deal valued at about $8.2 billion, released Frameboy on 5 November, a research preview that “streams frames from cloud GPUs to your browser at up to 15 fps” as users steer the camera through a generated 3D world, according to the company’s blog.
At up to 15 frames a second, that is up to 900 generated frames per minute of exploration, a quarter of the frame rate of the 4K stream at 60 fps that World Labs itself described last year as needing over 100K tokens a second, by our arithmetic. The company says Frameboy “is not cheap compared to local rendering methods”.
Business Pill · Physical AI
A short explainer of physical AI: AI systems that perceive and act in the physical world. It teaches the general idea only and does not discuss this story.
The key insight: As we read it, World Labs is moving its world model from a render-and-wait tool to a live stream, and every streamed frame is a fresh model call on a cloud GPU. At up to 15 fps that is up to 900 frames a minute in each session, by our arithmetic, and the company itself says the service is not yet cheap compared with local rendering.
What World Labs Released
Frameboy serves “a variant of Atlas that has been highly optimized to achieve the fastest, most efficient inference possible”, World Labs says. Atlas, introduced on 1 September, is the company’s world model; at launch it generated “up to 1 minute of video at 1440p” with camera control. That launch and Frameboy’s release are 65 days apart, by our count.
The demo lets users explore a curated list of scenes. World Labs says each was “seeded from a single input image”, and that real places captured from several viewpoints can be explored the same way. The company has also built the streaming mode into Viewfinder, a web app for building scenes frame by frame, which it says is “currently available to test in closed beta”.
The company’s argument is about how people use the model. “If users have to wait tens of seconds or even multiple minutes to see a video generated by Atlas, it forces them to be cautious and methodical”, the post says. With a real-time model, it adds, when the cost of a bad idea is near zero, “there are no bad ideas”.
The uses World Labs lists show who it expects to pay: users can “explore a famous landmark, scout out virtual viewpoints for a film shoot, inspect progress on a construction site, flesh out a concept for a 3D videogame level, or simulate an environment for a robot to navigate through”.
The Cost Side
Each frame is generated by the model on a server, not drawn by a graphics card in the user’s device. World Labs says Frameboy “currently must run on a powerful GPU server in the cloud” and adds: “if history is any indication, what starts slow and expensive becomes fast and cheap as both hardware and software evolve”. The post gives no price, GPU type or count of users a server can carry.
The company set out the scale of the problem in its RTFM post of October 2025: “generating an interactive 4K video stream at 60fps requires producing over 100K tokens per second”, which it called “neither feasible nor economically viable given today’s computing infrastructure”. RTFM, its earlier real-time model, ran “at interactive framerates using just a single H100 GPU”.
Hardware is where the company is heading. In its September post World Labs said it had “signed a definitive agreement to join AMD” and that the partnership began “starting with model training and inference optimization on AMD GPUs”. We covered AMD’s $8.2 billion all-stock deal, which AMD expects to close by the end of 2026, subject to regulatory approvals. The Frameboy post does not say which GPUs serve the demo.
How It Compares
Google DeepMind called Genie 4, announced on 27 October, “our first real-time world model to generate audio natively”; Google did not give a frame rate or a price. Google’s video model is priced per second of output: Gemini Omni 1.2 Flash costs $0.07 per second of 720p video, for clips of up to 15 seconds. World Labs has not priced Frameboy, so the two cannot be compared on cost.
As we read the post, the product bet is that speed changes behaviour: World Labs argues that waiting minutes for a video makes users cautious, while a live stream invites exploration.
As we read the list of uses, the paying customers World Labs has in mind are film crews scouting locations, construction teams, game designers and robot simulation, not only consumers.
As we read the AMD post, the hardware link is deliberate: World Labs says it is joining AMD to get closer to the hardware, after optimizing training and inference on AMD GPUs.
World Labs, Frameboy: Atlas in Real-time, 5 November 2026
“Once latency and cost drop low enough, users no longer even perceive a discrete act of โgenerationโ โ the model can simply stream frames back continuously as they navigate through the scene.”
Three Implications
FOR FILM, GAMES AND DESIGN World Labs lists scouting viewpoints for a film shoot and fleshing out a videogame level among the uses of Atlas.
FOR ROBOTICS The company also lists simulating an environment for a robot to navigate through.
FOR COMPUTE DEMAND World Labs says a 4K stream at 60 fps would need over 100K tokens per second; Frameboy runs at up to 15 fps on cloud GPU servers, at a cost the company calls not cheap compared with local rendering.
The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, a streamed world model is that idea in its purest form: each frame of exploration is a new inference call on a cloud GPU, so usage and cost rise together. The framework does not discuss World Labs or world models.
What Is Not Established
The post gives no resolution for Frameboy’s stream, no price, no GPU type and no limits on session length or users. The 900-frames-a-minute figure is the demo’s maximum rate, and the comparison with a 4K stream at 60 fps covers frame rate only. The AMD deal had not closed when the post was published.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
World Labs now streams its Atlas world model live from cloud GPUs at up to 15 fps, up to 900 frames a minute, a quarter of the 60 fps rate it used last year to describe the compute problem, by our arithmetic. It says the service is not yet cheap, and it is joining AMD in an $8.2 billion all-stock deal.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Facts and quotations are from World Labs’ blog posts of 5 November 2026, 1 September 2026, 28 September 2026 and 16 October 2025, and its and Fei-Fei Li’s posts on X; the deal value is from AMD, as reported in our earlier coverage. Frame counts are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
