Optimal conditions

Sony bought Gaikai in July 2012 for $380 million. OnLive had shown cloud gaming on stage at GDC three years earlier. The demos worked. In the lab, on a good line, close to the data center, they worked beautifully. Then they met real homes, real Wi-Fi, and real prices. Gaikai became PlayStation Now, and PlayStation Now was eventually folded into a PlayStation Plus tier. Google launched Stadia in 2019 and shut it down in January 2023. Cloud gaming is a real business today, but in 2026 the leading service, GeForce Now, caps paid plans at 100 hours a month. It took the better part of two decades to get from "we can do this" to a product with a price that works.

I was at Sony for part of that. The lesson I took away: just because something works in optimal conditions does not mean it works in the real world, and almost never at a price people will pay.

Generated worlds are at the 2009 stage now. Here is the case for patience, with numbers.

Where the race actually is

On 11 October, ByteDance showed a real-time world model on stage at a Douyin event: type a sentence, get a 3D world, walk through it with keyboard and mouse, build things in it, step through a door into outer space. It has no name and has not been released. The only performance figure in circulation, about 20 frames per second, comes from unofficial Chinese reporting that also calls it "not smooth." Bloomberg's anonymous sources tie it to PICO headsets. ByteDance has not confirmed any of it.

Sixteen days later Google released Genie 4, with native audio, live text steering, and, per TechCrunch, up to two minutes of consistency. Google published no resolution, frame rate, latency, or hardware figures. On public evidence, ByteDance's best case is parity with Genie 3, which shipped in August 2025. That is roughly 12 to 15 months behind.

But frame rate is not the race. Within a year of Genie 3's 720p at 24 fps, at least four other teams hit 720p at 20 to 40 fps, one of them open source. Then an independent benchmark, PlayWorld, put nine models through long-horizon play. The best, Genie 3, scored 2.12 out of 5. The fastest on paper came last. These models still forget the world they just made. On a monitor that is a curiosity. In a headset, where you turn around and expect the room behind you to still be there, it breaks the whole thing.

Every player is a new render farm

A game is built once and copied to millions of machines. A generated world is made again for every person in it. Demis Hassabis said it plainly: "a consumer of it is another instance of the creation of it."

Almost nobody publishes what that costs. ByteDance did, in its own research. Its Seaweed APT2 paper (June 2025) runs a real-time world at 1280x720 and 24 fps on eight H100 GPUs per stream. At November 2026 rental prices, $2.23 to $2.90 per GPU-hour, that is roughly $20 to $35 per user-hour once you allow for idle capacity. GeForce Now works out to $0.20 to $0.40 an hour. The gap is 50 to 175 times.

The market is telling you the same thing. Decart's public real-time API is two cents a second, which is $72 an hour. Google's Project Genie sits behind its most expensive AI subscription and caps each session at 60 seconds. When the richest company in the field rations a product to one minute at a time, the cost curve is not a rounding error.

What a headset adds

Everything above is for one flat 720p view. A headset is harder in three ways at once.

• Pixels. A Steam Frame or Quest 3 in stereo at 90 Hz needs about 38 times the pixels per second of 720p at 24 fps. Galaxy XR needs about 111 times. Generated by brute force, my estimate is 110 to 1,300 H100s per user, or $250 to $3,700 an hour.
• Latency. Head motion has to reach your eyes in about 20 milliseconds or people feel it. APT2's published latency is 160 to 170.
• Two eyes. The best published stereo world generator, StereoWorld, manages about 5 frames per second at 480x640.

None of those three closes by brute force. There is a way around them, and headsets already use it: let the cloud generate the world more slowly, and let the headset handle your head. Reprojection, the same trick behind Asynchronous TimeWarp and SpaceWarp, keeps the view stable at 90 Hz while the world itself updates at 24 to 45. That brings the cost down to an estimated $25 to $125 per user-hour. Better, and still a premium price for a niche product.

The road that is already open

There is a cheaper version, and it works today. Instead of generating every frame, generate the world once as real 3D, usually Gaussian splats or meshes, and let the headset render it like any other scene. World Labs' Marble has exported worlds you can view natively on Quest 3 and Vision Pro since November 2025. Tencent's HY-World 2.0 outputs editable meshes and splats for Unity, Unreal, and Isaac. Generation is a one-time cost. After that, stereo, 6DoF, 90 Hz, and latency are the headset's job, which it already does well.

The catch is that these worlds are static. Nothing in them responds to you the way a Genie world does. The credible headset product is a hybrid: a generated 3D scaffold rendered locally, with cloud models animating only the parts that need to be alive. Google calls hybrid world models a "promising path." Nobody has shipped one.

Where the cost makes sense

So who pays $25 to $125 an hour for a generated world, or a few hundred dollars to generate one well? Not a consumer looking for a reason to buy a headset. A business that already pays far more for the alternative.

An architect walking a client through a building that does not exist yet. A design team reviewing a product at full scale before tooling. A factory planner testing a new line layout, or a robotics team training a fleet in a thousand variations of one warehouse. Every one of those has a controlled network, managed devices, a short session with a clear purpose, and a budget line that already covers rendering farms, physical mockups, or site visits. That is the opposite of a living room on bad Wi-Fi, and it is exactly where cloud rendering has always worked first.

Niantic got there before the rest. It sold Pokémon Go and its games business to Scopely for $3.5 billion in 2025, and the company that remained, Niantic Spatial, now sells mapping and spatial AI to enterprises. In September it launched a Places Library of 100 simulation-ready real-world environments to train robots in. From Pokémon to the warehouse floor.

Who follows? Look at where the first paying customers already are:

• Waymo built its world model on Genie 3. A robotaxi fleet is a better customer than a gamer.
• Odyssey's newest model runs robot arms, humanoids, a car, and drones before it runs a game.
• World Labs' new Atlas model is available to enterprise customers first.
• NVIDIA's Cosmos is aimed squarely at robotics and autonomous vehicles.
• Tencent's 3D worlds export straight into the engines that factories and studios already use.

The first people who pay for generated worlds, at today's prices, will be enterprises and robots. Consumers come later, and through the cheap road first.

A realistic timeline

Cloud gaming is the best guide we have. It took 11 years to get from OnLive's first demo to GeForce Now and Xbox Cloud Gaming as mass-market products with a price, and another six for the economics to settle into hour caps and tiers. Generated worlds are moving faster, because AI compute gets cheaper faster than streaming video ever did. They are still on the same curve.

The cost line decides when. Below is my estimate of the hourly cost of streaming a generated world into a headset, starting from today and falling at three speeds: 10 times cheaper a year if new chips and model distillation both break right, 3 times if they don't, and 5 times in the middle. A consumer product needs to get close to what cloud gaming charges now.

My central case: 2030, before a consumer product with a price streams a cloud-generated world into a headset anywhere in the world. The fast case is late 2028. The slow case is 2031 or later. Generated 3D scenes that download and render on the headset itself will come sooner, probably in 2027, but that is a different product: a world you visit, not one that is alive around you.

Generated worlds will help consumer headsets. Just not soon, not first, and not the way the demos suggest. If cloud gaming is the guide, the headset that sells because of a generated world is a 2030 product. The enterprise headset that pays for itself because of one could arrive in 2027.

Disclosure: I spent eight years at PICO, which ByteDance owns. Nothing here comes from non-public PICO information; everything about ByteDance is from public reporting and published papers.