Grasshopper has been in production use for about fifteen years. The number of architectural practices with a generative design capability is significantly larger than it was in 2010. The number of buildings that meaningfully owe their form to a generative process is — generously — a few hundred. Out of millions built.
The discipline produced careers. It did not produce a category shift.
This essay is about why, and why the next five years are different.
The four reasons generative design stayed niche
1. The tools stayed in specialist hands
Grasshopper as a UI is hostile to non-specialists. The entry cost is not a learning curve; it is a learning cliff. You need to understand data trees, graft and flatten, the difference between implicit and explicit iteration, and why your geometry explodes when you change the order of operations. That is a meaningful skill set that takes months to acquire.
The economic consequence inside firms was predictable. The work got fenced off. A "design tech team" formed — sometimes two people, sometimes ten — and became the internal consultancy everyone resented. The rest of the practice learned to submit requests and wait.
This is not unique to Grasshopper. It happened with BIM, with parametric facades, with computational engineering. The tool became a specialism rather than a multiplier, and the specialism carried the headcount cost of a permanent team.
The tools that change practices are the ones that make the current practitioner more capable, not the ones that create a new category of practitioner alongside them.
2. The fabrication gap never closed
Generative design produces a specific kind of geometry. It tends toward the branching, the fibrous, the self-similar, the topology-optimised — forms that emerge from system logic rather than from a hand moving through space. That language is compelling as theory. It is expensive to build.
CNC milling, folding sheet steel, and casting concrete are not hostile to complex geometry, but they are expensive at it. Robotics lowered the floor somewhat — the work I did at RMIT with Roland Snooks on the NGV Floe Pavilion used robotic polymer deposition to fabricate forms that would have been impossible a decade earlier. But robotic fabrication at that level of craft is research-grade work, not production infrastructure.
The result: parametric tools produced geometries that either got domesticated (simplified to fit a standard mill or formwork) or lived permanently on the render farm. The gap between what the software could compute and what a contractor could price and build never closed to the point where generative design was a sensible economic choice for a standard commission.
3. Optimisation problems are rarely architectural
The most common application of generative design in practice was not form-finding. It was optimisation: minimise solar gain on this facade, maximise floor plate efficiency given this footprint, find the structural form that minimises material use.
These are real problems. They were also mostly solved by specific, narrow tools decades ago. EnergyPlus, Karamba, TTToolbox — the engineering domain had point solutions long before Grasshopper arrived. What Grasshopper brought was a way to connect these tools in a visual programming environment, which was genuinely useful. But connecting tools is plumbing, not design.
The interesting architectural problems — how to sequence a building's programme so that chance encounters happen, how to design public space that works for the widest range of uses, how to make a building that a particular community will actually inhabit and maintain — do not fit the optimisation frame. They are not problems you can write a fitness function for.
Generative design occupied the intersection of computationally tractable and geometrically novel. That intersection is smaller than the discipline assumed.
4. The firm's incentives were against it
Generative tools save time. Firms bill time. The economics were adversarial from the start.
A practice that builds a parametric residential facade toolkit can now deliver a facade design in four hours that previously took forty. If that practice bills hourly, they have just destroyed 90% of their fee for that work category. If they switch to fixed fees, they capture the efficiency. But switching to fixed fees requires confidence in scope, which requires sophisticated client management, which requires a practice that has reorganised itself around value rather than process.
Most firms never made that reorganisation. The computational team built tools that saved time. The partners billed those hours anyway or quietly declined to use the tools on real jobs, reserving them for competitions. The business model suppressed the technology.
What changes when AI lowers the floor
The four failure modes above were structural. AI does not fix all of them — but it does break the first one completely, which has second-order effects on the rest.
Natural language collapses the specialist bottleneck. The current generation of tools — Speckle with LLM integration, Hypar, various GPT-to-Grasshopper translation layers, custom tools built on Claude or GPT-4o with code execution — allows a practitioner without parametric training to describe what they want in plain language and receive working geometry. The translation layer that required months to learn is now available to anyone who can articulate a spatial problem clearly.
This does not mean the specialist disappears. Someone still needs to know when the generated result is wrong, what the failure mode looks like, and how to constrain the problem so the model produces useful output. But the entry cost drops from a two-year skill acquisition to something closer to good prompt engineering and domain knowledge. Every experienced architect already has the domain knowledge.
Multi-modal models break the geometric language prison. The forms that generative tools produce are now being generated directly from image models, not just from mathematical operators. The visual language of spatial AI is not constrained to the parametric aesthetic. You can describe a building in words, sketch it, photograph a reference, and get geometry that reflects the intent rather than the operator's grammar.
Fabrication-aware generation closes the build gap from the design side. Structural AI — topology optimisation coupled to fabrication constraint modelling — is maturing. The gap between what is computationally beautiful and what is physically buildable is being closed not by making fabrication more flexible (though that continues) but by making generation smarter about what fabrication can accept. The constraint is entering the generation loop rather than being applied as a filter afterward.
The firm-incentive problem persists. This is the one AI does not fix directly. A firm that bills hourly will suppress AI tools for the same reason it suppressed Grasshopper: the efficiency belongs to the client, not the firm. The practices that will benefit from the next wave are the ones that shift pricing before they shift tools — fixed fees, value-based pricing, productised design services. The technology is not the constraint. The business model is.
The next five years
The "computational designer" role as it currently exists will disaggregate. The narrow practitioner who writes Grasshopper scripts and understands data trees but does not do design leadership or product thinking will be automated away — not suddenly, but steadily, as LLM-to-parametric tooling matures and the entry cost continues to drop.
What survives is the role that always had the most value and was least well-named: the person who can frame the right computational question, evaluate whether the machine's answer is useful, and translate between what software can calculate and what a practice actually needs. That is not a different kind of software skill. It is design judgment applied to systems thinking.
Generative design will become a feature, not a discipline. It will be a capability that lives inside standard architectural software — the way BIM is now a capability of Revit rather than a specialism separate from Revit. The firms that currently see "computational design" as a team will eventually see it the way they see "3D modelling": something everyone does at some level.
The practices that capture the value first are the ones that reorganise pricing, not tools. The most dangerous bet in the next five years is to invest heavily in building a generative design capability on an hourly billing model. The second most dangerous is to assume the current specialist headcount is permanent.
The category shift is coming. It is just arriving through the business model, not the software.
Ven Iyer is a platform builder and computational design practitioner with ten years across AEC and spatial AI. He led the design technology function at Hayball Architects and computational product work at Bollinger Grohmann, and co-founded Imersian. See his work or get in touch if this is the problem your firm is trying to solve.