Spatial AI is a new term to some, but students of convergence know that Agentic AI, Spatial AI, and Quantum will be the technologies that converge to lead us into the future. This is not a fringe idea anymore. It is where capital, talent, and platform strategy are all moving. Per Variety, Spatial AI reached roughly $130 billion in 2024, with most analysts projecting it to grow at 21-24% CAGR through 2030-2032.

Roblox just published something the AI industry should be studying closely. Per Respawn , Roblox has 381.8 million monthly active users and 144 million daily active users in 2026, with 35 billion hours of annual engagement and 44+ million games on the platform. This is not a small experiment. This is one of the largest real-time digital environments on the planet. And they have just announced Roblox Reality, and it answers a question that has stalled spatial computing for a decade.

The question is simple. What happens after the demo?

OpenAI's Sora can generate breathtaking video. Google's Veo can render cinematic worlds. Runway, Pika, and a dozen others can produce footage that fools the eye on first watch. Yet none of these systems can do the one thing that turns a generated world into a place, which is to share your spatial world with another person and trust that you are both seeing the same thing.

Roblox's own team gave this failure mode a name that travels. In the architecture announcement from SVP of Engineering Anupam Singh, the company described pure neural world models as "vivid dreams, spectacular to look at, but fleeting and incredibly lonely." They lack persistence and multiplayer coherence. They lack the rules that turn a moving image into a place you can return to.

That phrase is the most important framing I have seen in AI all year, because the vivid dreams problem is the unsolved problem at the heart of Spatial AI. It is also the problem Roblox just claimed it has solved.

Why Spatial AI is the Category That Matters

Spatial AI is the next layer of the agentic stack. This is the category where artificial intelligence stops producing images and starts producing places, environments that multiple people can enter, share, and trust.

Apple shipped Vision Pro and is building generative spatial experiences on top of visionOS. Meta has rebuilt itself around Reality Labs, with Horizon Worlds, Quest, and Ray-Ban Meta glasses as the consumer surface. Snap has spent more than a decade on Spectacles, Lens Studio, and AR-first commerce. Niantic, Nvidia Omniverse, and every major game engine vendor are positioning for the same prize.

The prize is persistence with trust at scale.

Nobody has shipped that at scale. Apple has the hardware and the cash, yet Vision Pro is still a single-user device looking for a multiplayer use case. Meta has the scale and the spend, yet Reality Labs is the largest sustained operating loss in tech history, and Horizon Worlds has not produced the persistent shared experiences the thesis requires. Snap has the cultural relevance and the AR install base, yet Spectacles remain a product looking for a category.

Roblox just walked past all of them with a different architecture.

The Vivid Dreams Problem Across Every Spatial AI Category

Look at where generative AI has stalled commercially, and you find the same shape every time. AI agents that can hold a conversation, yet cannot remember you across sessions. AI shopping assistants that can recommend products, yet cannot complete a transaction with settlement and recourse. AI copilots that can draft strategy, yet cannot anchor it to the system of record your finance team actually uses. AI companions that feel intimate in the moment and dissolve the second you close the tab.

Vivid dreams, every one of them.

The field assumed better models would close the gap. Bigger context windows, sharper reasoning, more parameters. None of that fixes the problem, because the problem is architectural, not a question of capability. This is the shift most teams are still missing. A model alone cannot hold shared truth across users, persist state across sessions, or guarantee that the rules of the world stay constant while you are inside it. That is simply outside what models do.

Spatial AI makes this gap impossible to ignore. A video model that hallucinates a different landscape every second is a curiosity. A spatial system that hallucinates a different landscape every second is unusable. You cannot ship a game, a commerce experience, an enterprise training environment, or a multi-user meeting on a foundation that drifts. That is why Spatial AI has been the category everyone wants and nobody has won.

The Spatial AI Division of Labor Roblox is Shipping

Here is the architecture. The Roblox Game Engine, running on cloud servers, handles authoritative game logic. Physics, collision, state synchronization, where every player is, who won the race, what the rules are. A new Video World Model, called Super Upsampler, runs on edge infrastructure and handles the visual output. Photorealistic textures, lighting, secondary motion, and fluid dynamics layered on top of the engine's source-of-truth geometry.

The engine owns the rules. The AI owns the look.

It is the design principle that will define the next 24 months of spatial and agentic AI.

Roblox did not build a better generative model. It built a system where the generative model knows what to delegate. The engine is the ground truth. The AI is the surface. That split is the reason Roblox can claim Spatial AI leadership when Snap and Apple cannot. Snap and Apple have been trying to make the model the experience. Roblox put the model on top of an experience that already worked.

This is the post-LLM architecture, and it generalizes well beyond gaming.

In agentic commerce, the language model negotiates and the blockchain rail settles. Diaz Nesamoney’s work, which is building autonomous purchasing infrastructure for AI agents, at DaVinci Commerce points exactly here. The model is the front of house. The settlement layer is the back of house, and that is where trust lives.

In AI identity, the model reasons about intent and a cryptographic wallet verifies who is acting. IronClaw and OpenClaw exist because a model that cannot prove who it is acting on behalf of cannot be trusted with anything that matters.

In enterprise AI, the agent acts and the system of record governs. The companies seeing real adoption are the ones whose agents are wired into Workday, SAP, Salesforce, and the ledger that actually decides what is true inside the business. Impressive demos alone do not move the needle.

The same shape, over and over. Generation on top. Determinism underneath. The model is brilliant at ambiguity, perception, and synthesis. The infrastructure underneath is brilliant at memory, consistency, and proof. Neither one wins alone.

What Roblox has not yet proven for Spatial AI

The architecture argument is sound. The execution is unfinished.

Singh himself has said that scaling the Super Upsampler to millions of concurrent players is still being worked out, with no firm public timeline. Edge inference at Roblox's user volume is an engineering problem nobody has solved at this scale, and the cost curve matters as much as the technical proof.

The Video World Model also has to clear the safety bar a platform with 144 million daily users, many of them under 16, will be held to. None of that invalidates the thesis. It does mean the window between architectural claim and shipped product is where Apple, Meta, and Snap still have room to respond.

What boards should be asking about Spatial AI this quarter

The next wave of category winners in AI will be the companies that solved the vivid dreams problem in their domain, the ones who figured out which parts of the experience must be real and built infrastructure that holds. The model layer is commoditizing faster than anyone predicted, so the best model alone will not be the moat.

For Spatial AI, Roblox just declared its answer. Apple, Snap, and Meta now have to respond to an architecture argument they did not see coming. For commerce, identity, enterprise, and the agentic stack underneath all of them, the answer is still being written.

Two questions belong on every board agenda this quarter. Which parts of your AI experience are vivid dreams, and which are persistent truth. And which side of that line is your capital actually funding.

Roblox just claimed Spatial AI by refusing to build the thing everyone else was building. Most companies will miss it because they are still optimizing the model, not the system. Pay attention to what comes after LLMs, because the future of Spatial AI is already here, and it does not look like a bigger model.