Nine days ago in Taipei, Jensen Huang asked an AI agent to design a house. Not a concept sketch, not a mood board collage, not one of those vaguely aspirational renderings that real estate developers use to sell condos eighteen months before anyone pours a foundation, but an actual house, on a PC, in minutes.
The agent opened Rhino, modeled the terrain, calculated setbacks and building envelope constraints from the site parameters Huang fed it, proposed building forms optimized for cost and comfort and structural efficiency, then generated interior layouts with walls and circulation paths and room proportions that looked, at least on that massive keynote screen, like something a competent designer might produce after two weeks of iterative work. When it made a mistake, it caught the error and corrected itself. It exported the model to Blender, transferred materials, rendered photorealistic views from multiple angles under different lighting conditions. All of it running locally on NVIDIA's new RTX Spark, an ARM-based machine with a Blackwell GPU and 128GB of RAM that will ship from ASUS, Dell, HP, Lenovo, Microsoft, and MSI this fall.
It was, by any measure, an extraordinary demonstration. It was also, if you understand what actually happens between a client's first conversation with an architect and the moment a general contractor breaks ground, a demonstration of the easiest 30 percent of the job.
What the Demo Actually Showed
Schematic design. That's the industry term for what Huang's AI agent performed on stage, and it maps roughly to RIBA Work Stage 2 or AIA Phase 1 in the standard service agreements that govern how architects charge for their time. Schematic design is where ideas take initial shape: massing studies, room relationships, circulation diagrams, preliminary floor plans. It is creative, exploratory, and genuinely important to the success of a project. It is also, in the economics of a custom residential project, roughly 15 to 20 percent of an architect's total fee, which itself runs 8 to 15 percent of construction cost according to AIA median data.
Run the numbers on a $500,000 custom home: architect's total fee runs $40,000 to $75,000, of which schematic design accounts for $6,000 to $15,000. Design development, construction documents, bidding, and construction administration consume the rest, and those later phases are where the complexity lives, where the liability concentrates, and where no AI tool available today can produce output that a building department would accept for a permit application.
Patrick Murphy, CEO of Maket, one of the few platforms offering AI-generated residential floor plans with agentic editing capabilities, puts it plainly: "It really does 70 to 75 percent of the work. And then you get to a point where you meet an architect and you save a fair amount of time and money." That remaining 25 to 30 percent, the structural review, the code compliance verification, the engineering coordination, is not a rounding error. It is the part that determines whether your house stands up, passes inspection, and doesn't kill anyone.
Construction Documents Are a Different Universe
A photorealistic render of a house tells you nothing about whether it can be built legally. Nothing. It does not tell you whether the floor-to-ceiling heights comply with IRC Section R305, whether the egress windows meet the 5.7-square-foot minimum opening required by R310.1, whether the stair geometry satisfies the riser-tread relationship in R311.7, whether the structural headers over those beautiful floor-to-ceiling windows can actually carry the roof load above them, or whether the mechanical systems will fit in the spaces the AI allocated for them, because AI agents do not yet understand that a furnace requires combustion air intake and a condensate drain and six inches of clearance from combustible materials, and that the elegant utility closet the algorithm drew is four inches too narrow in one dimension that matters enormously to the HVAC installer who has to work inside it.
Construction documents, the drawings that a contractor actually builds from, contain thousands of individual decisions that must satisfy overlapping and sometimes contradictory requirements from the International Residential Code, local amendments, zoning ordinances, energy codes, accessibility standards, HOA covenants, and the engineering calculations that a licensed professional stamps with their name and their liability. A set of residential CDs for a custom home typically runs 20 to 40 sheets. Every dimension, every note, every detail section carries legal weight.
No AI tool produces these today. Not Maket. Not ARCHITEChTURES, which automatically generates 3D models and compares energy performance across alternatives. Not Autodesk Forma, which compresses early-stage building design into hours. And not NVIDIA's demo agent, which was never intended to.
Where AI Design Tools Actually Stand
Honest inventory. Maket generates residential floor plans from natural language input and lets users refine layouts conversationally, changing room sizes and configurations through dialogue rather than clicking and dragging on a screen. ARCHITEChTURES produces parametric multi-unit housing layouts optimized against performance criteria. Autodesk Forma handles site analysis and early massing. Zaha Hadid Architects has reportedly shortened design feedback cycles from days to hours using AI-based simulation tools for complex geometries and lighting studies.
All of these operate in the schematic and early design development phases. All of them require a licensed professional to take the output and develop it into something buildable. And all of them, including the most sophisticated among them, struggle with the spatial reasoning that experienced architects perform intuitively: understanding that a bathroom above a living room means plumbing in the ceiling of your most formal space, that a west-facing glass wall in Phoenix means a $400-per-month cooling penalty that no amount of aesthetic elegance justifies, that the stunning double-height entry your client fell in love with during the schematic presentation will require a structural steel moment frame that adds $22,000 to the construction budget and three weeks to the schedule.
RIBA's September 2025 white paper found that 88 percent of architects expect AI to become increasingly important for business success within the next decade. RIBA President Chris Williamson captured the professional consensus in two words: "excitement and trepidation." Four months later, the AIA survey showed only 8 percent of firms have moved beyond experimentation into implementation, with adoption driven significantly more by architects under 50, which suggests the profession sees the trajectory clearly but finds the current tools insufficient for production work.
What RTX Spark Changes
Local execution. That is the real story buried inside the keynote theatrics, and it matters far more than the house design itself.
Every AI design tool currently available runs in the cloud, which means your project data, your client's site information, your preliminary designs, and your fee structure (if the tool infers it from project parameters) all travel to someone else's servers. For firms handling high-net-worth residential clients who would prefer their Beverly Hills lot dimensions and security system layouts not reside on a startup's cloud infrastructure, local execution is not a feature but a prerequisite.
RTX Spark ships with a 20-core ARM CPU designed with MediaTek, a Blackwell-class GPU approaching desktop RTX 5070 performance, full CUDA support (a first for Windows on ARM), and 128GB of unified RAM. It runs x86 emulation for legacy applications, meaning existing architectural software works without recompilation. Jensen Huang's framing was deliberately grand: "For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work."
Strip away the marketing language and what remains is genuinely significant for small and mid-size architecture firms. A machine powerful enough to run inference on large models locally, capable of driving Rhino and Blender and rendering engines simultaneously, priced for professional desktops rather than data centers. If the agent frameworks mature, and if the architectural plugins follow, a two-person firm could theoretically compress its schematic design phase from two weeks to two days without sending a single byte of client data to an external server.
That "if" carries a lot of weight.
What This Means If You're Building a Home
Today, in June 2026, AI will not design your house. It will help your architect explore more options faster during the earliest phase of the project, generate massing studies and floor plan alternatives at a pace that was impossible when every iteration required hours of manual modeling, and produce compelling visualizations that help you understand what a space will feel like before anyone draws a construction detail. These are real efficiencies with real value, particularly for clients who struggle to read architectural drawings and need photorealistic context to make confident design decisions early rather than requesting expensive changes during construction.
What AI will not do, and what the GTC demo did not attempt, is produce the documentation your contractor needs to pull a permit and build the structure safely. Procore's April 2026 industry report found that over half of construction leaders expect automation to disrupt the industry within five years, but the same report emphasized that human expertise remains indispensable for the foreseeable future.
If you are hiring an architect for a custom residential project this year, ask whether they use AI tools in their schematic design process. If they do, you may see more design options explored more quickly, and you may save time in the early weeks of the project. Expect to pay full professional fees for design development and construction documents regardless, because no tool exists that can reduce the professional effort required for those phases, and any architect who claims otherwise is selling you a liability problem disguised as a discount.
What Architects Should Actually Worry About
Not replacement. Compression.
If AI compresses the schematic design phase from weeks to days, clients will eventually expect schematic fees to reflect that compression, the same way clients stopped paying for hand-drafted presentation drawings once CAD made them trivially reproducible. Architects who bill hourly for schematic work will feel this first. Those who charge fixed fees for the schematic phase will absorb the efficiency gains as profit margin until competitive pressure forces the savings downstream to clients.
Longer term, the question is whether AI advances from schematic design into design development, then into construction documentation, and finally into the engineering coordination that currently requires multiple licensed professionals working in parallel. Each of those transitions requires not just better algorithms but new legal and professional frameworks: who stamps an AI-generated structural calculation, who carries the liability when an AI-designed detail fails in the field, who insures a project where the construction documents were produced by software that no state licensing board has evaluated or approved.
Those are five-to-ten-year questions. Jensen Huang's demo did not answer them. It did not try to.
Limitations of This Analysis
NVIDIA's GTC Taipei demo was a controlled stage presentation, not a documented real-world project, and no independent evaluation of the AI agent's output quality, structural feasibility, or code compliance has been published. Maket's claim that its platform performs 70 to 75 percent of schematic design work is self-reported by the company's CEO and has not been verified by a third-party study comparing AI-generated and traditionally produced schematics across a meaningful sample of residential projects. AIA adoption data is from 2025, collected before RTX Spark's announcement, and may understate current experimentation rates. Fee calculations use AIA median data for the 8 to 15 percent range and apply a 40 percent schematic-phase allocation derived from standard AIA service agreement fee distributions, but actual fees vary significantly by market, project complexity, and firm. No home has been designed start-to-finish using AI tools and built to completion with a certificate of occupancy, which means every projection of AI's impact on residential design economics is, at this point, extrapolation from partial evidence.