Meshy Says ARR Reached $100 Million in Under Two Years as AI 3D Race Intensifies

The QbitAI report, published on Oct. 4, 2026, said the debate followed the release of GPT-6 Astra, which can operate Blender, generate scripts and use spatial understanding to build geometry-based structures. It can construct a house in Blender and export it to Unreal Engine 5 as a walkable scene. Social media users also used Blender MCP to prompt Astra to build a hospital corridor from scratch.

But in a side-by-side test, QbitAI gave the same reference image to Astra and Meshy. The Meshy output preserved hair strands, facial features, cloak, belt, armor layers and a shield leaf pattern. Astra's output was more abstract: hair became a single helmet-like block, the face became cartoon-like and the body looked like geometric parts assembled in Blender. The report used this to argue that general models excel at tasks decomposable into geometric relationships, such as CAD, parametric modeling and procedural generation, while character and complex prop assets require learning shape and detail from large real-asset datasets.

According to QbitAI, Meshy disclosed that its ARR grew from $1 million to $100 million in under two years. ARR, or annual recurring revenue, measures sustainable, repeatable subscription or contract revenue. The report compared this with Bessemer Venture Partners' The State of AI 2025, which described "shooting star" AI companies as reaching $100 million ARR in about four years and "supernova" companies in about 1.5 years. ElevenLabs took about 20 months from first product launch to $100 million ARR, while HeyGen took 29 months from $1 million to $100 million ARR. Meshy did it in less than two years, faster than HeyGen. The report also noted that 3D is harder to commercialize than voice or video because assets must pass topology, texture and size checks before entering game engines or printers.

The report cited 3AGameFactory, an open-source framework by Peking University's OpenDCAI team, in which AI coding agents break down game requirements and decide whether each asset should be generated, programmed or sourced. Characters and key props were assigned to Meshy. After generation, assets still needed mesh cleanup and size validation; characters required rigging and animation. Bohan Zeng, a doctoral student and project lead, said the biggest difficulty remained the gap between "looking correct" and "being truly usable," with some problems appearing only after assets enter playable scenes. Meshy and Astra were also shown collaborating on a ruins scene: the general model decomposed the scene and prepared reference images, then called Meshy through an API to generate assets one by one, which were imported into Blender for placement, scaling and lighting.

Meshy's value in this workflow depends on asset quality: shape, detail, color and patterns. The report said a mechanical owl preserved layered armor plates, chest gears and a carved base as separate structures. In Meshy's internal 3D geometry alignment benchmark, Meshy 7 led tested models on single-image input across proportion, spatial distribution and surface detail; its surface-detail score was 59.8%, versus roughly 49% to 55% for other models and 52.7% for Meshy 6 Lite. Meshy 7.1 generated an orc with clear armguard knots and woven, worn skirt edges visible in untextured geometry. Meshy said it raised geometry generation resolution from 2048 cubed to 4096 cubed; its Detail Richness benchmark score under 4K rendering was 23.1%, above other tested models. In an end-to-end texture alignment test, Meshy 7 scored 66.2%, the report said. The team is also exploring direct mesh generation; Meshy T2, using a Flow Matching route, can generate vertices and connections and control mesh complexity by budget.

Commercial demand, according to the report, is clearest in games and 3D printing. Game art outsourcing firm Juego Studios lists stylized 3D props at about $300 to more than $2,000 and a 3A-quality character at more than $5,000. 37 Interactive Entertainment has integrated Meshy into production: for complex characters and monsters, teams generate heads and limbs separately, then assemble and refine them in 3D software. This gives a high-poly base more than 60% complete before manual sculpting and reduces sculpting work in the base-model stage by 30% to 40% in complex next-generation and ancient-style projects. For simple props and characters, complete models can be generated and cleaned for further work; in some casual and mobile game projects, modeling time per asset fell from five days to about 2.5 days.

In 3D printing, Bambu Lab said that by the end of 2025 MakerWorld's global and China sites had about 50 million registered users and tens of millions of monthly active users, while active creators numbered about 300,000 and quality models nearly 2 million. More people can buy printers than can create original models, the report said. Meshy is targeting that gap, letting people without modeling skills turn ideas into printable 3D models. As Meshy added multi-color 3D printing capabilities, Bambu Lab, Creality, Elegoo, FlashForge and xTool entered its customer list, according to the report.

Meshy's revenue is described in three layers: personal subscriptions from independent developers, teachers and printing enthusiasts; enterprise API use by companies including 37 Interactive Entertainment, NetEase Games and Nexon; and hardware ecosystem integration with 3D printer brands. QbitAI said half of the 10 companies by global market value or valuation have placed Meshy in key parts of their production chains. The report argued that enterprise APIs and hardware ecosystems are key to scaling revenue. Higher model quality reduces rework, allowing Meshy to enter more production processes; more production processes turn one-off trials into frequent API calls, creating a loop of model capability, usage and revenue. The report also said Meshy is moving from AI for Work toward AI for Fun and introduced a real-time interactive multimodal content architecture.