There's a phrase doing a lot of unearned work in the architecture-AI conversation right now: purpose-built. Every vendor roundup gestures at it. The major platforms, we're told, are training specialized models on architectural data — building codes, construction sequences, material libraries — so their AI doesn't just make pretty pictures but understands the discipline. Layer in "agentic" systems that chain steps together, and the implication is clear: the era of architecture AI being a general image generator with an arch-viz prompt is over.
Maybe. But the most important counter-example is sitting in plain sight. Veras 4, the tool every comparison crowns for image quality, runs on Google's Nano Banana Pro — a general foundation model Chaos did not train. So before we accept the "everyone's training domain models" narrative, it's worth doing what we always do here: read the incentives, separate the claim from the architecture, and ask what's actually true under the marketing.
Two very different things called "AI"
The whole confusion comes from one word covering two completely different engineering decisions. When a vendor says it has "AI for architecture," it is doing one of these:
Training a model means assembling a dataset and teaching a neural network from it. The model's knowledge is that data. To genuinely train an architectural model, a vendor needs an enormous, well-labelled corpus — drawings, codes, built outcomes, materials — plus the compute and the talent to train on it and keep training. It is slow, expensive, and the moat, if you can build it, is deep.
Renting a model means calling someone else's foundation model — Google's, OpenAI's, a strong open-weights model — through an API, and shaping the result with prompts, reference images, fine-tuning on a smaller dataset, and post-processing. It is fast, comparatively cheap, and it inherits every improvement the provider ships. The moat here isn't the model; it's everything you wrap around it.
"AI for architecture" hides two opposite bets: own the model and carry the cost, or rent the model and compete on everything else.
Neither is cheating. But they produce different companies, different risks, and different reasons to trust — or distrust — the result. And a roundup that lumps them together as "vendors building architectural AI" is flattening the single most important distinction in the category.
Reading the incentives
Why does almost everyone, when you look closely, turn out to be renting? Because the incentives point that way with brutal consistency.
A general foundation model from Google or OpenAI has been trained on a scale no architecture vendor can match — budgets in the billions, data in the petabytes. A purpose-built model trained on a comparatively tiny architectural corpus starts from behind on raw capability and has to claw back the gap with domain specificity. Worse, the rented model gets better on someone else's dime: when Google ships the next engine, the tool renting it inherits the upgrade for free, while the vendor who trained its own model has to fund the next training run itself.
That's exactly the logic visible in Veras. Rather than try to out-train Google at image generation, Chaos rented Google's best image model and spent its effort on the things Google will never build: reference-image input from a plan or sketch, no-prompt scene generation, and — the real moat — integration into BIM and CAD across multiple platforms. Veras's advantage is its product, not its model. Read cynically that sounds like a knock. Read honestly it's the smarter bet, and the roundups have the story backwards when they file Veras under "purpose-built architectural AI."
So who, if anyone, is actually training?
The vendors with a credible path to a genuinely owned domain model are the ones who already own the data — and that's a much shorter list than the roundups imply.
- Autodesk sits on an enormous corpus of real project data through Revit, AutoCAD and Forma, plus the context and site-analysis intelligence already shipping inside Forma. If anyone can train models on construction logic and code, it's the company that hosts the construction logic and code. The capability claim here is the most believable.
- Chaos has deep rendering and physical-light expertise and a clear AEC-specific positioning — but its flagship visual tool rents Google's model. Its domain training, where it exists, is more plausibly in the physics-and-pipeline layer than in the image model itself.
- Adobe has the model-training muscle and the creative install base, but "trained on architectural data" specifically is the weakest-supported of the three claims; Adobe's strength is general creative AI that architects use, not an architecture model per se.
Notice the pattern: the believable "training" stories belong to whoever controls a proprietary dataset, not to whoever markets the loudest. Domain data is the actual moat. Everything else is a wrapper around a rented engine — and there is nothing wrong with a good wrapper, as long as you know that's what you're buying.
Owns the model and the data (rare, data-rich incumbents) · Rents a general model but owns deep integration and a workflow moat (the strongest tools, e.g. the Veras pattern) · Rents a general model with a thin prompt-and-skin layer (most new entrants) · Markets "purpose-built" with nothing verifiable underneath (the category to discount). Where a tool sits tells you how durable its advantage is — and how much of its quality will quietly come from Google's next upgrade rather than its own work.
Why this matters to your stack
This isn't a philosophy-of-AI exercise. Where a tool sits on the owned-versus-rented spectrum changes three things you actually care about.
Durability
A tool whose quality is mostly its rented model is hostage to that relationship. If the provider changes pricing, terms or availability — or if a competitor rents the same engine — the advantage can evaporate overnight. A tool whose advantage is deep integration or proprietary data is far harder to dislodge. When you standardize a studio on a tool, you're betting on the durability of its moat, not this quarter's image quality.
Trajectory
Renting has a delightful property: the tool gets better when Google or OpenAI ships, with no effort from the vendor. That's why a thin wrapper can leap forward overnight. But it cuts both ways — improvement isn't in the vendor's control, and "we're on the latest foundation model" is a promise about someone else's roadmap, not their own.
Data exposure
Renting a model means your inputs may travel to the foundation provider as well as the vendor. "Trained on architectural data" should also make you ask whose data, and whether yours becomes part of the next training set. The owned-versus-rented question and the confidentiality question are the same question wearing two hats — which is exactly why we keep pulling that thread.
Our take: discount "purpose-built," judge the moat
The "everyone is training architectural AI" narrative is mostly marketing collapsing two different things into one flattering word. The truth in 2026 is that almost everyone is renting a general model from Google or OpenAI, and the best tools — Veras among them — win not by owning the model but by owning the integration and the workflow around it. The genuinely owned-model stories belong to the handful of incumbents who already control proprietary project data, with Autodesk the most credible of them.
For a practitioner, the lesson is liberating rather than cynical: stop treating "purpose-built" as a feature and start asking the only questions that survive contact with reality. Does the tool actually hold my geometry and understand my workflow? What is its moat — proprietary data, deep integration, or just a prompt and a logo? And where does my model go when I use it? Answer those and the owned-versus-rented label sorts itself out. Ignore them, and you'll keep buying the loudest claim instead of the most durable tool.
We read the incentives behind the tools so you can choose a stack that's still standing next year. Join the studio newsletter for the analysis, or read our companion piece on AI-native versus plugin renderers.
Analysis based on publicly described vendor capabilities and product architectures as surfaced in 2026 industry coverage. Claims about which vendors train versus rent models reflect public positioning and should be confirmed against vendor documentation. No affiliate relationship with any tool or platform named.