AI
Topology
For most of computing history, 3D was the most expensive kind of content. Someone had to build it by hand, one polygon at a time. Since 2022, AI has cut the time it takes to generate a rough 3D draft by about four orders of magnitude, from 90 minutes to under a second, and a draft now costs well under a dollar.
One thing has not gotten cheap: structure. In 3D, topology is how a shape is put together, and it decides whether the shape can be animated, edited, simulated or manufactured. AI made the rough draft nearly free. Getting the structure right is the part it is only now learning, and it is where the value is moving.
These are the curves behind that shift, plotted on a log scale so the rate of change reads off the slope.
Generation went from hours to under a second in two years. Progress now goes into detail rather than speed.
II · CodeGeneral models now build 3D you can edit.Frontier models write CAD scripts and operate 3D tools, so they produce structure rather than loose meshes. The leaders are close together.
III · CADEngineering still demands exact structure.AI helps engineers finish the parts they draw, but no major vendor generates finished parts yet. The value sits in the CAD kernel and the design history.
IV · GamesGames adopted AI faster than they trust it.AI-made content is spreading through new releases while developer sentiment sours. Clean topology, rigs and consistent style are still open problems.
V · SimulateGenerated worlds moved from demos to products.Interactive world models became paid products and drew a multibillion-dollar acquisition. Whether they output pixels or lasting 3D is still open.
Generation compounds fast
How fast each curve has moved; longer bars moved faster. Generation sped up by orders of magnitude, then stalled in 2024, and detail is now what improves. Open training data grew at roughly the pace of chips, then stopped.
Linear scale is on. Curves that fall by orders of magnitude flatten into an L, so read rates with Log.
From a prompt to an object
Making 3D used to mean modeling it by hand or measuring something real. Generation needs neither. The first text-to-3D methods in 2022 had no 3D training data to speak of, so they borrowed it: DreamFusion asked a 2D image model to grade thousands of renders of a blob until the blob looked like the prompt. It worked, slowly.
Once Objaverse put close to a million 3D objects in the open, researchers could train models that output a shape directly. Generation time fell about four orders of magnitude in eighteen months. The constraint moved from compute to data, and from making a draft to making something you can ship. Since 2024, speed has held still while detail has climbed.
General models learn to build in 3D
Chapter I ends at a constraint: drafts are cheap, structure is not. The route that attacks it most directly comes from general models, not 3D specialists. Neither GPT-6 Astra (OpenAI, 3 Sep 2026) nor Claude Opus 5.5 (Anthropic, 22 Sep 2026) generates a mesh. Both write programs that produce geometry, such as CAD scripts, Blender Python and engine code, or operate 3D software the way a person would. Then they render the result, look at it and fix it. Code is the only route into 3D that produces structure by default: an exact, editable object.
Three findings hold across every test so far. Checking their own work is what makes these models good at 3D. The top of the field is crowded: five models from three labs are statistically tied on the one independent benchmark. And near-top quality costs anywhere from $3.79 to $31.50 a scene, depending on the model.
Rebuilding parts from pictures
The one CAD benchmark both labs report is BenchCAD. A model sees multi-view renders of a mechanical part and has to write the CadQuery program that rebuilds it. The benchmark has about 17,900 verified programs across 106 industrial part families, roughly half anchored to ISO, DIN, ASME and IEC standard tables. Scores are voxel IoU: how much the rebuilt solid overlaps the real one.
With a Python tool to run its code and inspect the result, each lab reports its new model at about 0.96. Without the tool, Opus 5.5 scores 0.73. That 0.23 gap is the main finding: these models are good at CAD mainly because they can check their own work.
Two cautions. Voxel overlap rewards the right overall shape; it does not check tolerances, threads or assemblies. And the 0.96 figures come from each lab's own tool setup. The BenchCAD authors, testing models their own way, found that current systems "often recover coarse outer geometry but fail to produce faithful parametric CAD programs."
The same loop is reaching real CAD software. Onshape Labs ships a FeatureScript MCP server for Claude, and community plugins let models sketch, extrude, fillet and mate parts in Onshape. One builder had an earlier Claude model assemble a 4-part monitor arm from a sketch, and found the bottleneck was reading images, not executing CAD.
No tool-free score has been published for GPT-6 Astra.
All scores are self-reported by the labs. Opus 5.5: Anthropic system card §8. GPT-6 Astra, GPT-5.6 Sol and the Fable 5.1 figure: OpenAI launch materials. No-tools scores via the llm-stats.com BenchCAD leaderboard. Benchmark: BenchCAD paper (May 2026).
From grey box to playable
In games, both models act as directors of a toolchain, not as asset generators. They write concept prompts, model and rig in Blender, import into Unity, Unreal or Godot, write the gameplay code and play-test the result. Neither lab has published a game-development benchmark, so the evidence below is demos and early-user reports, and it should be read that way.
Sources: Crypto Briefing, VoxelMatters, MindStudio, 36Kr, Anthropic, yuv.ai, GameDev Academy, Medium survey, AA-3D. Vendor and tester figures are as reported.
AA-3D: who builds the better 3D scene?
The lab benchmarks above are self-reported. AA-3D, launched by Artificial Analysis on 12 October 2026, is the first independent, human-judged test of the skill this chapter is about. Each model works as an agent in Artificial Analysis's Stirrup harness and turns a written brief into an interactive three.js scene. People then compare the scenes two at a time without knowing which model made which, and the votes become an Elo rating.
- Briefs
- 55 tasks in 7 domains, from the Eiffel Tower to a playable pinball table
- Judging
- Blind pairwise votes from a recruited panel, 11K+ votes on the live board
- Score
- Bradley–Terry Elo with 95% intervals. GPT-6 Luna is fixed at 1,000
An interval that reaches the shaded band overlaps the leader's interval, which is Artificial Analysis's test for a statistical tie. On the live board that covers the top five; at launch it covered the top four: Sonnet 5.5, Opus 5.5, Astra and Fable 5.1. Hover a row for its rank range and cost. Source: AA-3D leaderboard, read 16 Oct 2026.
Opponent-balanced win rate with 95% intervals; n is the number of scenes in voting (Opus/Astra). Source: AA-3D launch article, data as at 12 Oct 2026.
Games show the widest gap in the benchmark. Opus 5.5 wins 88% of its matchups, after balancing for opponents, against 63% for Astra, and it is the only domain where their intervals don't overlap. Two caveats apply. The cells rest on 6 and 8 scenes, and Opus has no voted scene for the two vehicle simulations, Mountain-bike trail and Glider soaring, while Astra does. These are also browser mini-games such as pinball, solitaire and dodgeball, not engine projects.
Machines, the closest thing to CAD, is level: 82% against 78%. The briefs ask for an exploded F1 car, a watch movement with labelled parts, a working trebuchet, judged on overall preference rather than dimensional accuracy. AA-3D and BenchCAD measure different things, and both find the two models even on mechanical work.
Beyond Artificial Analysis's 12-point noise band, Opus leads games (+24) and natural phenomena (+17), and Astra leads space (+15). The other four domains are within noise.
Elo shown at the end of each bar. "Tied for first" means the model's 95% interval overlaps the leader's. Cost is Artificial Analysis's estimate from list prices and a modeled cache-hit rate, not billed spend.
Cost. The five models tied for first cost between $3.79 and $31.50 a scene, an 8× spread for the same tier of quality. Gemini 4 Argon is the cheapest way in, and Claude Sonnet 5.5 leads the board for $14.63. Opus costs $20.50 and Astra $18.11, close despite very different working styles: Astra takes 60 turns and 153k output tokens per scene, Opus 137 turns and 393k, but Astra's list price is 2.5 times higher.
Measured against general intelligence, Astra is the overachiever. Artificial Analysis reports that it beats the Elo its Intelligence Index predicts by 46 points.
Two scenes, one brief. Which would you keep?
These are the real scenes Opus 5.5 and GPT-6 Astra built for AA-3D, in random order. Open both, pick one, then see who made it and how the panel voted.
Benchmark notes: sandbox, build process and caveats
| How they build (launch data) | Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Cost per scene | $20.50 | $18.11 |
| List price, per 1M tokens in / out | $4 / $20 | $10 / $50 |
| Median agent turns per scene | 137 | 60 |
| Output tokens per scene | 393k | 153k |
| Tool time per scene | 45.5 min | 19.7 min |
| Looked at its own render by turn 20 | 93% of scenes | 96% of scenes |
| Median buttons and controls in a scene | 17 | 19 |
| Blender calls per scene | 1.9 | 1.5 |
| Scenes that entered voting | 50 of 55 | 54 of 55 |
| Pinball build: time, turns, tokens | 136 min · 141 · 377k | 115 min · 69 · 156k |
| Trebuchet build: time, turns, tokens | 129 min · 207 · 379k | 143 min · 76 · 203k |
Sandbox. Each run is offline, with three.js 0.186, Rapier physics, Blender 4.5, Python, Node and a headless browser for self-testing. It gets 4 CPU cores, 8 GB of memory and no GPU, and stops at 250 turns or 3 hours. Each model gets one attempt per task, and panelists must spend at least 30 seconds with each scene.
What separates the best scenes. Models that look at a render of their own scene early score higher, and so do scenes with more interactive controls and more detail. Blender is barely used, about twice per scene for Opus and 1.5 times for Astra. Nearly everything is hand-written three.js.
- AA-3D scores web scenes in three.js with software rendering and no GPU. It says nothing direct about Unity or Unreal pipelines, rigging, or exact geometry.
- Panelists vote on overall preference. There is no published rubric for accuracy, aesthetics or interactivity.
- Each domain cell rests on 2 to 10 scenes. Artificial Analysis treats gaps under about 12 points as noise.
- Cost is estimated from list prices and a modeled cache-hit rate, not billed spend.
- A separate community arena, threejseval, ranks Opus 5.5 (High) first and Astra (Medium) fourth as of 16 Oct 2026. It uses different agent setups: Cursor for Anthropic models, OpenCode for OpenAI.
Opus 5.5 and Astra, side by side
| Measure | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|
| Released | 3 Sep 2026, enterprise first | 22 Sep 2026, API and paid plans |
| How it makes 3D | Writes code; operates Blender and other apps directly | Writes code; drives Blender, engines and CAD through computer use and MCP |
| Native mesh output | No | No |
| BenchCAD IoU, with tools (0–1) | 0.959 | 0.962 |
| AA-3D Elo (live, 16 Oct) | 1166 ±33 · #3, rank range 1–5 | 1191 ±34 · #2, rank range 1–3 |
| AA-3D games win rate | 63% (8 scenes) | 88% (6 scenes) |
| AA-3D machines win rate | 78% (9 scenes) | 82% (9 scenes) |
| AA-3D cost per scene | $18.11 · 60 turns | $20.50 · 137 turns |
| Computer use | 72.6% on OSWorld 2.0 | 81.8% on OSWorld 2.1 (partial credit) |
| Game-dev evidence | Vendor demo (Blender to UE5); Playco and MindStudio tests; AA-3D games domain | Vendor polish test; WebDev leaderboard; community games; AA-3D games domain |
| Main gap | Fine detail and production polish | Binary engine assets and custom pipelines |
OSWorld versions differ, so those two numbers are not directly comparable. BenchCAD and OSWorld scores are vendor-reported. AA-3D is independent; its domain win rates are opponent-balanced launch figures.
Code is the second way to make 3D, after generation, and the only one that produces structure by default. On CAD the two labs are level near the top of the only shared benchmark, and an independent human panel agrees. In games, Opus 5.5 has the edge on small playable scenes, and neither model is a production pipeline yet. The openings this leaves are around the models rather than in them: agent connectors for CAD tools and game engines, verification of agent-built geometry, code-CAD kernels that turn scripts into exact solids, and review tools for the step between a prototype and a shippable asset.
In CAD, probably right is wrong
A game can ship a shape that looks right. A machined part cannot. CAD models are boundary representations: exact solids whose faces and edges come from a feature history of sketches, constraints and operations, tied to tolerances and standards. In CAD, topology is not a quality of the mesh; it is the definition of the part. That is why the generation boom of Chapter I barely touched mechanical design, and why the AI that has shipped in CAD tools mostly adds structure to geometry that people make.
The market is large and growing without AI's help. CIMdata puts product lifecycle software, which includes CAD, at $88.3 billion in 2025, up 9.9%. Engineers use AI widely but lightly: 98% of leaders in Autodesk's 2026 AI Pulse survey use at least one AI tool, yet only 33% of engineers in a Digital Engineering reader survey use AI in their work, and use of generative design fell to 12%.
- 01Sketch and constraintsShipping
- Autodesk's AutoConstrain, rolled out from January 2025
- Predicts the dimensions and constraints that turn a loose sketch into a parametric one
- 02Parametric partEarly
- No major vendor has shipped text-to-part
- Autodesk's neural CAD: announced September 2025, unreleased in October 2026
- What works is agents driving the CAD kernel:
- Fusion connector for Claude (April 2026)
- Onshape's FeatureScript MCP server (August 2026, early access)
- Zoo's Zookeeper agent (February 2026)
- 03Scan or mesh to CADEarly
- Backflip: mesh to parametric CAD, generally available August 2026
- Estimates turning a part into CAD falls from about $1,500 to $10 (company figure)
- Autodesk's AutoTimeline rebuilds a feature history for imported geometry, due 2027
- 04AssembliesManual
- Generators work one part at a time
- Spectral Labs' SGS-1 can't do one-shot assemblies
- SOLIDWORKS 2026 automates only fastener placement
- 05DrawingsShipping
- SOLIDWORKS 2026 and Fusion generate 2D drawings automatically
- Including section views and callouts
- 06Simulation and reviewShipping
- AI physics surrogates and automated design review ship today
- Where the larger rounds went: PhysicsX, Neural Concept, CoLab
- 07Standards and tolerancesManual
- BenchCAD's DIN 2095 spring: the model drew the helix but missed the closed, ground ends
- No released training set has helical sweeps or involute gears, per the authors
Best single-sample IoU (overlap between rebuilt and true solid) from the comparison table in cadrille (arXiv 2505.22914); only new records shown.
Easy parts are nearly solved; real parts are not. On DeepCAD, a benchmark of simple sketch-and-extrude parts, AI went from 47% to about 90% overlap in four years, mostly by training on synthetic CAD programs. Harder tests drop sharply: about 51–55% on mechanical parts (MCB), about 68% on real scans, and on BenchCAD's industrial part families, frontier models scored 0.21–0.28 under the paper's own protocol. The labs report about 0.96 with their own tool setups (Chapter II); the gap is the cost of structure.
The data is synthetic. Human-made datasets with construction history stopped at DeepCAD's 178,000 models in 2021. Every newer million-scale set is generated, and building Autodesk's Zero-to-CAD took 60 billion tokens, with only 22% of designs valid on the first try.
Round sizes as announced by the companies or reported by the press, 2023–2026. Hover a dot for the lead investor.
Shape generation is early; engineering data has drawn the larger checks. Companies that generate CAD geometry, including Zoo, Adam, Leo AI and Hestus, have raised early rounds of $1.5M to $10M, and Backflip's $30M Series A is the largest so far. Larger rounds have gone to companies that sell engineering structure and data: CADDi, which turns drawings into searchable manufacturing data, reached a $1.2B valuation in September 2026.
Incumbents are investing in adjacent layers. Cadence bought Hexagon's design and engineering unit for about €2.7B (February 2026), and Autodesk agreed to buy MaintainX for about $3.6B (May 2026) and put $200M into World Labs. No AI text-to-CAD startup has been acquired yet, so the category is still open.
| Company | What has shipped | Announced or in preview |
|---|---|---|
| Autodesk (Fusion) | AutoConstrain (2025); AI drawings; Fusion connector for Claude that builds geometry in Fusion's kernel (April 2026) | Neural CAD text-to-part (announced September 2025, unreleased); AutoTimeline feature-history recovery (2027) |
| PTC (Onshape, Creo) | Onshape AI Advisor, a help assistant | FeatureScript MCP server through Onshape Labs (early access, August 2026) |
| Siemens (Designcenter) | Design Copilot guidance; 500,000+ uses a month of AI features such as command prediction (vendor figure, June 2026) | No text-to-geometry feature found |
| Dassault (SOLIDWORKS) | AI drawings and fastener auto-assembly in SOLIDWORKS 2026; LEO virtual companion (July 2026) | No generative part geometry found |
| Zoo | Text-to-CAD (2023), its own geometry engine, and the Zookeeper agent that plans constraints before modeling (February 2026) | Reports its parametric code error rate fell from 50% to 16% (vendor figure) |
| Backflip | Mesh and scan to parametric CAD, as a Fusion add-in (August 2026) | $30M Series A (NEA, a16z), December 2024 |
| Spectral Labs | SGS-1: images, drawings or meshes to parametric B-rep parts (September 2025) | No one-shot assemblies, by its own account |
| Adam | Onshape copilot that cleans up feature trees rather than making new shapes (open beta, March 2026) | $4.1M seed (TQ Ventures), October 2025 |
CAD is where this essay's argument is sharpest. The value sits in the kernel and the feature history, which incumbents own, so the practical route for AI in 2026 is to drive their kernels rather than replace them. The openings are around that core: recovering structure from scans and meshes, checking AI-built parts against standards and tolerances, and generating the structure-rich training data no one has.
Sources: CIMdata (Jun 2026); Autodesk AI Pulse 2026; Digital Engineering 24/7 2026 Technology Outlook; Autodesk, PTC, Siemens and Dassault announcements; Engineering.com; cadrille; BenchCAD; Zero-to-CAD (Autodesk, Apr 2026); company funding announcements (TechCrunch, SiliconANGLE, Fortune, tech.eu). Company figures are as reported.
In games, volume is cheap and trust is not
Games tolerate approximation that CAD cannot, and they need a lot of 3D, so they became the first big market for AI asset generation. But a game-ready asset still needs clean topology to deform, UVs to texture, a rig to animate, LODs to run fast and an art director to approve it. Generation got cheap first. The rest of the pipeline is where the cost and the arguments are.
Use has stalled while opinion has turned. In GDC's annual surveys, the share of developers using generative AI rose from 31% to 36% and stayed there in 2026, while the share saying it is harming the industry rose from 18% to 52%. Visual and technical artists are the most negative (64%), and 68% of 3,779 artists in Poliigon's 2025 survey never use AI 3D model generators.
- 01ConceptShipping
- The most common use of generative AI in games
- Mostly brainstorming and concept work (GDC surveys)
- 02Base mesh and sculptEarly
- Meshy with 37 Interactive: 50% less modeling time on simpler titles
- Only 30–40% less sculpting on complex ones
- Sculpting is about a third of a character's time in one outsourcer's breakdown
- 03RetopologyEarly
- Vendors now compete on engine-ready topology:
- Tripo Smart Mesh P1.0 (March 2026)
- Tencent Hunyuan3D Studio (September 2025)
- Neither publishes topology metrics
- Tripo's own launch: AI output usually needs its topology rebuilt by hand
- 04UVs and texturesShipping
- Texture and material generation ship in Substance Sampler and Unity
- EA partnered with Stability AI on PBR materials (October 2025)
- AI-placed UV seams beat a standard tool in a 20-artist study
- 05RiggingShipping
- Auto-rigging works for standard humanoids and quadrupeds:
- AccuRIG 2.0 and Meshy
- Epic's Mesh to MetaHuman, which rigs any human mesh, including AI output
- Props and faces remain manual
- 06AnimationShipping
- Cascadeur generates up to 89 frames of motion between two key poses (2025.1)
- 07LODs and engine setupEarly
- Unity 6.2 builds LODs on import but ignores skin weights and blend shapes
- Unreal's MCP plugin, tested July 2026: an agent built a six-room greybox
- It could not author procedural graphs
- 08Art directionManual
- No tool owns style consistency across hundreds of assets
- This is where backlash lands
- Clair Obscur lost two Indie Game Awards over AI placeholder textures
GDC State of the Game Industry, 2024–2026 reports (2,300–3,000+ respondents a year). Question wording changed slightly between years.
Sulka Haro's census of 53,597 releases (July 2026). Valve narrowed its disclosure rules in January 2026. AI-flagged games are about a third of launches but an estimated 10–27% of sales, and reach 100 reviews about 55% as often as other games.
Round sizes as announced or reported, 2023–2026. VAST's yuan rounds converted by DealStreetAsia.
A funded category with two leaders. Meshy raised nearly $400M at a $1.5B valuation in July 2026 and says it passed $100M in annual recurring revenue in September. VAST, which makes Tripo, raised four rounds in 2026, the largest about $446M in September, with game publishers Perfect World and 37 Interactive among its strategic investors. Both companies' traction figures are self-reported.
Others raised far less: Kaedim $15M (2024), Sloyd €3M (2024) and Scenario $6M (2023). Luma raised $43M in 2024 alongside its Genie 3D model, then put its $900M Series C (November 2025) toward world models (Chapter V). Google folded CSM into DeepMind in January 2026.
| Platform | What has shipped for 3D | Where it stops |
|---|---|---|
| Roblox | Open-source Cube 3D (March 2025); 4D generation in open beta (February 2026), with 160,000+ objects made in one game's early access; part-level CubePart (May 2026); a prompt-to-game Build tab used to publish about 9,000 games | Generated objects must fit fixed part schemas: a car is five named parts |
| Epic (Unreal) | Experimental in-editor assistant (UE 5.7) and MCP plugin (UE 5.8); Mesh to MetaHuman rigs any human mesh through Epic's cloud | An agent could not author procedural content graphs in a July 2026 test |
| Unity | AI tools in open beta (May 2026), with an assistant, an MCP server and a 3D Object Generator | The generator makes static, single-part props that Unity frames as placeholders |
| Genie 3 and Project Genie worlds (Chapter V); acquired CSM, which turned images into engine-ready meshes | Genie outputs pixels with 60-second generations, not editable assets |
Labor and players are pushing back. SAG-AFTRA's video game strike ran about 320 days before a July 2025 agreement added consent and disclosure rules for AI replicas, and games from Call of Duty: Black Ops 7 to Clair Obscur drew backlash over AI art.
Games show the thesis from the demand side. Generation is abundant, but AI-flagged games earn a smaller share of sales than of launches. The platforms are moving to own the structure steps: Roblox fits generated objects to schemas, Epic rigs any mesh in its cloud, Unity calls its output placeholders. The openings are in engine-ready structure, meaning topology, rigs and LODs that survive animation, and in style control that art directors trust.
Sources: GDC State of the Game Industry 2024–2026; Sulka Haro's Steam census and analysis (July 2026); Poliigon State of 3D 2025; Meshy and VAST announcements, DealStreetAsia, Bloomberg; Roblox, Epic, Unity and Google announcements; Puget Systems (Jul 2026); SAG-AFTRA. Company figures are as reported.
Two bets on how machines will hold a world
Generation gives you objects, and games need whole places. Simulation asks for a whole world that responds as you move through it. Two approaches compete. Video world models predict the next frame from the last ones and your input, with no explicit geometry. They are general and fluid, and they forget what was behind you. Persistent 3D worlds generate geometry first and render it, which costs flexibility but gives you a place that stays put and can be exported to a game engine or a robot simulator.
Which one wins decides whether explicit 3D is the substrate of future simulation or an export format. In 2026 the two bets began to merge: World Labs' Atlas outputs video and 3D Gaussian splats together, and Tencent's HY-World 2.0 turns text into navigable splat or mesh worlds. The money followed: AMD agreed in September to buy World Labs for about $8.2 billion.
Hover or tap a mark for what it does, its specs and who made it. Sources: company announcements and papers; fps and memory figures are as reported by each lab.
Seeing in 3D got cheap too
This essay is about making 3D, but the input side moved as well. In 2020 a Neural Radiance Field needed a day or two of GPU time to turn photos into a scene; since late 2024, feed-forward models rebuild a 32-photo scene in about a second. Gaussian splatting made those scenes run in real time, and automotive lidar has held at about $200 a unit since 2024. Capture is no longer the scarce input, so this essay treats it as background.
The 3D AI stack, layer by layer
When a cost curve falls as fast as 3D generation's did, the layer that is falling usually stops being the business. Value moves to the scarce input next to it. In 3D, the scarce inputs are clean training data, production-ready output, and distribution into the places 3D gets used: games, commerce, construction, robots.
| Layer | What is happening | Pricing power | Where margin may hold |
|---|---|---|---|
| Generation models | Open-weight models (TRELLIS, Hunyuan3D, Stable Fast 3D, TripoSR) track closed APIs closely. | Contested | Models trained on licensed, high-quality data, and tuned for one domain such as characters, CAD or interiors. |
| Frontier LLM agents | General models (GPT-6 Astra, Claude Opus 5.5) build 3D by writing CAD scripts, Blender Python and game code, and by operating 3D tools. Both score about 0.96 IoU on BenchCAD with tools. | Concentrated | Agent connectors for CAD tools and engines, verification and QA of agent-built geometry, and code-CAD kernels that turn scripts into exact solids. |
| CAD kernels & engineering data | Incumbents own the kernels and feature histories. No major vendor has shipped text-to-part; the AI that ships adds structure to geometry people make. | Holding | Scan- and mesh-to-CAD, checking AI-built parts against standards and tolerances, and engineering data and review, where the larger rounds went. |
| Production last mile | Agents now texture, rig and export to engines in demos, but clean topology, art direction and studio pipelines still take people. | Holding | Art direction and studio pipelines hold; UVs and basic rigging are falling. Watch the share of agent-built assets that reach an engine with no manual fix. |
| World models | Compute-heavy, led by frontier labs and well-funded startups. AMD agreed to buy World Labs for about $8.2B in September 2026. | Concentrated | Simulation and synthetic training data for robotics and autonomy, where a crash in sim is cheap. |
| 3D data | Open 3D data is about 540× smaller than open image data, and no larger open set has appeared since 2023. | Scarce | Scan archives, asset marketplaces and anyone with rights to large 3D libraries. |
| Formats & delivery | glTF and OpenUSD are standard for meshes; splats, which generated worlds such as Marble and HY-World 2.0 export, are still being standardized. | Forming | Compression, streaming and viewers that make generated worlds and large scenes load on phones, headsets and the web. |
| Capture | Sensors are cheap and reconstruction is open source; scenes rebuild in about a second. | Commodity | Vertical workflows: real estate, construction progress, insurance, digital twins. |
Less attractive on this evidence: thin wrappers around image-to-3D APIs, where the core model is open and nearly free, and standalone capture apps.
The shape of the map
Four years of 3D AI, since DreamFusion, fit in one sentence: everything that produces pixels or rough shapes got cheap, and everything that produces structure is still expensive. That gap explains where value is likely to move.
-
Got cheapDrafts are close to free.
A rough 3D object takes under a second and well under a dollar on open models, and capturing a real scene is just as cheap. In these layers the business is the workflow around the model, not the model.
Chart in I · CreateTime to one rough 3D object<1sfrom about 90 min in 2022Turbo3D · 2024DreamFusion · 20221 s1 min1 hLog scale; each dot set a speed record. About four orders of magnitude faster, at well under $1 a draft. -
Still expensiveStructure is the bottleneck.
Clean topology, exact dimensions, rigs and materials still take people. General models that write CAD and Blender code are the first route that produces editable structure by default, and they now score about 0.96 IoU, out of 1.0, on the main CAD benchmark.
Charts in II · CodeBenchCAD IoU · 1.0 = exact match0.96with a Python toolRunning code to check its work is worth about 0.23 for Opus 5.5. Each lab reports its own setup. -
Still scarceData is the scarce input.
Open 3D data is hundreds of times smaller than open image data, and structure-rich data such as CAD histories, rigs and part labels is rarer still.
Chart in I · CreateLargest open sets, areas to scale3DImages≈540×fewer open 3D objects than open images- Objaverse-XL: about 10.2M objects (2023), still the largest open 3D set
- Re-LAION-5B: about 5.53B images
-
Two lanesSimulation is splitting in two.
Pixel-based world models are already used for driving simulation; Waymo's is built on Genie 3. Persistent geometry keeps the jobs where the output must be exact or editable: manufacturing, buildings and game assets.
Timeline in V · Simulate29 world models tracked, Feb 2024 to Oct 2026Pixel-based worldsDriving simulation: the Waymo World Model (Feb 2026) is built on Genie 3Persistent geometryManufacturing, buildings and game assets, where output must be exact or editable
Where we would look first
- Verification of agent-built 3D. Self-checking is worth 0.23 IoU on BenchCAD. Tools that check geometry, tolerances and engine readiness sit where the gains come from.
- Structure-rich 3D data. CAD feature histories, rigs, part labels and licensed libraries are the scarcest input in every chapter.
- Connectors between general models and 3D tools. Whichever layer general models call by default decides where 3D work happens.
What would show the map is shifting
| Signal | Where it stands | Would change the view at | Why it matters |
|---|---|---|---|
| BenchCAD without tools | Best published 0.747 IoU (Claude Sonnet 5.5); none published for GPT-6 Astra | About 0.90 | When models get CAD right without running code to check, structure gets cheap too. |
| Open-weights gap on AA-3D | Best open model 216 Elo behind the leader (Kimi K3, 981 vs 1197) | Under about 50 Elo | A small gap would commoditize agentic 3D building the way open models commoditized generation. |
| Largest open 3D dataset | About 10.2M objects (Objaverse-XL, 2023), still the largest | 100M+ real objects | A jump of 10× or more would ease the data constraint behind every chapter. |
| World-model memory | About 1 minute verified for Genie 3; Decart and Ant Group claim hour-plus sessions (vendor claims) | Independently verified multi-hour consistency | Hours of consistency would let pixel-based worlds replace geometry for more uses. |
| Text-to-part from a CAD incumbent | Autodesk's neural CAD announced September 2025, still unreleased in October 2026 | General availability with independent accuracy data | Would show incumbents can generate structure, not only check it. |
| AI-flagged games' share of Steam sales | An estimated 10–27% of sales against about 33% of launches | Sales share matching launch share | Would show players accept AI-made content at parity. |
| Agent connectors in engines and CAD tools | Onshape Labs publishes an official MCP server; we found mostly community-built connectors for Blender, Godot and Unity | Official connectors from the major engines and CAD vendors | Official connectors decide which tools general models operate by default. |