Why Big AI Labs Won't Build Floor Plan to 3D | VirtualSpaces

Why the Big AI Labs Won't Build Spec-Accurate Floor Plan to 3D

Every time I explain what we build at VirtualSpaces, someone asks the same question. If AI is getting this good this fast, won't one of the big model labs just do this too, and do it better?

It is a fair question. It is also the wrong worry. The honest answer is not that the large labs can't build spec-accurate floor plan to 3D. It is that it makes no sense for them to. Those are very different claims, and the difference is the whole story.

Let me give away the weak version of the argument first, because I do not want to sell you something that falls apart on a second look.

The version that is wrong

The weak version goes like this: general AI models can't read a drawing or build a room, so we are safe. That is not true, and anyone technical knows it. Today's multimodal models can look at a floor plan and produce a nice-looking room, and they get better at exactly that, reading documents and generating images, every few months. If I told you the big labs simply cannot do this, you would be right to close the tab.

So I won't tell you that. The capability gap is real today, but it narrows over time, and a moat that depends on other people staying bad at something is not a moat. The real argument has nothing to do with what a frontier model can do. It has to do with what a frontier lab will spend its attention on. Once you see the problem that way, the worry flips into something closer to a guarantee.

Focus is the moat

A horizontal AI lab is playing an enormous game. It is trying to move general intelligence forward: reasoning, code, agents, the kind of capability that reshapes every industry at once. Measured against that, the entire market for AI interior design is a rounding error. Independent analysts size it at roughly $3.3 billion in 2025, growing toward $15 billion by 2033. That is a real market, and it is growing fast. It is also a fraction of a fraction of the prize a frontier lab is chasing.

Here's the part people miss: The reason a large lab will not build this is not that the problem is too hard. It is that the problem is too small for them and too specific to matter to their mission. Every engineer they could point at floor plans is an engineer they are not pointing at the frontier. That trade is brutal, and it runs in our favor. Niche is not a weakness here. Niche is the reason the giants stay away.

It helps to be concrete about the asymmetry. A frontier lab measures progress in capabilities that compound across the whole economy, and it staffs against that horizon. A single vertical, however good the business, pulls in the opposite direction. It asks a team to go narrow and deep on one industry's quirks for years, learning how a stairwell is drawn in Mumbai versus Manchester, how a kitchen island is dimensioned, where a firm hides the ceiling height. That is not a rounding error a lab overlooks by accident. It is a choice its entire mission is built to make.

Think about the shape of it this way.

Nothing on the left is wrong. It is exactly what a lab should chase. It is also the reason the thing on the right is left open.

A model is an ingredient, not the meal

The second reason is more technical, and it is the one I would stake the company on.

Converting a 2D floor plan into a room you can trust is not one model doing one thing. It is a pipeline of specialized parts working in order:

• A stage that reads dimensional notation, the tiny numbers running along the walls, into clean structured data, not a vague impression of a room.

• A stage that rebuilds the geometry to exact spec, with wall thickness, ceiling height, and door swings placed to the drawing's own numbers.

• A stage that furnishes and renders while staying locked to that geometry and to a live catalog of real, purchasable products, so the output never drifts into invented walls or furniture nobody can buy.

And the tolerance is unforgiving. A general model that renders a living room an inch off looks fine on a screen and becomes useless the moment a real sofa has to fit against a real wall. The gap between a picture that impresses and a model you can order furniture against is measured in inches, and closing those inches is not a matter of a larger model. It is a matter of building the whole pipeline to refuse to guess.

A general model can sit inside that pipeline as a component. It is not the pipeline. The pipeline is the work, and most of the work is unglamorous: geometry, constraints, catalog integration, the plumbing that turns a clever demo into something a developer prices a building on. Large labs ship models and APIs. They do not ship browser-native vertical products wired into furniture catalogs and property workflows. That is not a criticism. It is a different business with a different customer.

This is the same pattern you see everywhere the frontier has already passed through. General models made speech recognition and translation cheap for everyone, and yet the companies that win in law, in medicine, in logistics are still the ones that wrapped those capabilities in a specialized system aimed at one industry's mess. The base model is the electricity. Somebody still has to build the appliance. I wrote earlier about why floor plan to 3D is genuinely hard; the short version is that a floor plan is a document, not a picture, and reading it correctly is only the first of several hard steps. A bigger general model helps with that first step. It does not hand you the rest.

The data nobody else has

The third reason is the quietest, and maybe the most durable.

The system that reads floor plans was trained on a set of residential drawing conventions we built by hand over roughly a year. That corpus does not exist on the open web at the scale and quality you would need. You cannot scrape it. You cannot buy it off a shelf. It is the accumulated messiness of how real architects and real firms annotate real drawings, across US and international formats, cleaned and labeled by people who understood what they were looking at.

Frontier labs train on broad, general, publicly available data. They are not going to commission a niche architectural annotation corpus for a market they have already decided is too small to chase. This is where the two moats reinforce each other. The people with the resources to build that data have no reason to, and the people with a reason to build it have to earn it the slow way. We earned it the slow way. That year does not come back for whoever tries to start now, which is a big part of why converting a blueprint to 3D at spec is not a weekend project on top of a general model.

What erodes, and what does not

I promised honesty, so here is the caveat that survives a skeptic.

Not every part of this is equally safe. As general multimodal models improve, the first step, reading a plan, gets easier for everyone. The advantage there narrows over time. If our only edge were "we read drawings a little better," you would be right to worry, and so would I.

That is not the edge. The durable parts sit downstream. The spec-accuracy layer that guarantees the 3D room matches the drawing to the inch. The catalog integration that keeps every object in the scene real and purchasable and sized to fit. The constraint machinery that stops a generative model from quietly hallucinating a wall or a product to make a prettier picture. None of that arrives for free on the back of a rising tide of general capability. Each piece is an engineering choice aimed at one problem. So the claim is not "the big labs can't read a plan." The claim is sharper and holds up better: even when anyone can read a plan, turning it into a spec-accurate room you can actually buy from is a separate, specialized problem that a horizontal lab has no reason to own.

There is a tell here worth naming. If a large player ever did want this market, the rational move would not be to build it from scratch against a team that already has the data and the pipeline. It would be to partner or to buy. When the cheapest path into a problem is to acquire the people who already solved it, that says something about where the real work lives.

Why this is the good news

Step back and this stops sounding defensive and starts sounding like a map.

The frontier moves horizontally. It makes every general capability cheaper and better for everyone, us included. Some problems, though, are won vertically, by whoever is willing to sit inside one industry's mess and solve it end to end. Spec-accurate 2D to 3D is one of those. It is beneath a frontier lab's attention and squarely inside the roadmap of the companies that live in this space: the design platforms, the property portals, the furniture and fixture businesses that need every listing and every unit visualized before a single one is built.

That framing also answers the quieter question sitting underneath the loud one. If the giants stay away, is the prize big enough to matter? For a horizontal lab, no. For the vertical it sits inside, it is foundational: the visual layer that a design platform, a property portal, or a furniture business would rather own than rebuild from nothing. Value like that is not decided by how the market looks to the frontier. It is decided by how badly the right specialist wants the ground it stands on, and specialists pay for ground they cannot easily recreate.

That is what makes this look less like a feature and more like infrastructure. The same engine powers Virtual Staging for an empty listing, AI virtual staging for a developer's pre-sales, interior design renders for a designer's pitch, and a renovation preview for a nervous homeowner. Different products on the surface, one specialized machine underneath. General AI makes the surrounding world smarter and the demand larger. It does not, on its own, build the machine that sits at the center of it.

Where this shows up

Two products run on this today. Foursite turns 2D floor plans and blueprints into photoreal interior design renders and virtual staging, without an outsourced render studio. Remodroom takes a single room photo and returns a photoreal redesign you can change in minutes. Both point at the same idea from different doors: the room you see should match the room that exists, or the room that will.

So the next time someone asks whether the big labs will just do this, the answer is short. They could try. They won't, because it is not their game, not their data, and not their business. The frontier is theirs, and it is a magnificent thing to own. This corner is ours. The whole point of a corner worth owning is that the people who could take it have far better places to be.