Author | Zhou Yixiao, WeChat | smiletalker
In 2024, the largest funding round in the AI 3D generation track was $15 million, invested in Kaedim, a startup originating from the University of Bristol in the UK, led by a16z Games. In the same year, Luma AI, the highest-valued player in the track, abandoned 3D to pivot to video generation and never looked back. At that time, investors considered the market too small, and the number of teams globally seriously commercializing AI 3D could be counted on one hand.
Two years later, Meshy closed a nearly $400 million Series B funding round, with a post-money valuation exceeding RMB 10 billion, setting a new record for the largest single funding round in the AI 3D field to date. This round was co-invested by top-tier institutions and strategic investors including IDG Capital, Matrix Partners China, and Monolith, with existing shareholders such as Granite Asia, Sequoia China, BAI Capital, and Source Capital making oversubscribed follow-on investments. This roster almost covers the top lineup of China's leading venture capital firms, and all existing shareholders chose to continuously increase their stakes.
This is already an unprecedented scale for the AI 3D track, and even more counter-intuitive numbers lie behind the funding amount. According to Meshy, the company's Annual Recurring Revenue (ARR) grew approximately 12 times within a year, with month-over-month growth consistently maintaining at 20% to 30%. Registered users have exceeded 12 million, and the cumulative number of generated models has surpassed 100 million. While most AI companies are still trapped in the cycle of trading growth for valuation and valuation for the next round, a 3D company may have stepped onto solid ground first.
In the first half of 2026, the directions of spatial intelligence and world models heated up rapidly. Heavyweight players such as World Labs and AMI Labs entered the field one after another, but most are still in the early stages of product development with no scaled revenue. Meshy is one of the few companies in this direction that already possesses real, continuous revenue.
Among the several mainlines of AI multimodal generation, Large Language Models (LLMs) still capture the most funding and attention, while the competition in image and video has escalated from battles among startups to a melee between tech giants and independent labs. In contrast, pure 3D generation remains off the main stage. Meshy has seized this long-underestimated demand. It may not have the loudest voice, but it has been the first to turn 3D generation into a business generating tens of millions of dollars in annual revenue and still growing rapidly.
Where Exactly Are the Barriers in 3D?
Painting a picture and sculpting a statue require completely different capabilities. During the Renaissance, the number of people who achieved perfection in both was extremely rare, with Michelangelo being one of them. AI multimodal generation faces a similar situation; teams capable of generating images are not necessarily able to do 3D, as the underlying problem structures they face are different.
The core task of image generation is to learn distributions on a 2D pixel grid, where the data structure is uniform and the training framework is mature. The first barrier in 3D is a step higher: 3D objects do not have a unique standard representation. Meshes, point clouds, voxels, and implicit functions—each representation has its own trade-offs and cannot be losslessly converted into one another. Choosing which representation to use and how to design the network architecture to adapt to it is a question with no standard answer right from the start.
The second barrier lies in the supervision signal. Image generation can calculate the loss function pixel by pixel, whereas a 3D model needs to be correct from all viewing angles, with closed geometry, reasonable topology, and UV unwrapping capable of bearing textures. These constraints often conflict with each other.
The third barrier is data. There are tens of billions of images and billions of videos on the internet, but high-quality 3D data is only in the tens of millions, and its distribution is highly uneven. Game assets, CAD engineering drawings, and scanned objects each have their own characteristics, and mixing them for training rarely yields ideal results.
Cumulatively, these factors dictate that AI 3D requires a team to possess both large model training and computer graphics capabilities simultaneously. The talent supply for the former has increased significantly in recent years, while the latter remains scarce.
Meshy's founder, Yuanming Hu, happens to stand exactly at this intersection. An undergraduate from the Yao Class at Tsinghua University, he led the development of the Taichi programming language during his Ph.D. at MIT (which has over 28,000 stars on GitHub), which is itself a product of graphics and high-performance computing. After obtaining his Ph.D., he founded Taichi Graphics, attempting to commercialize underlying graphics technologies. After two pivots within 18 months, the direction changed and products were built, but market feedback was never optimistic enough. The turning point came from a user's feedback. That person said, "I wouldn't pay for your software, but if you sell the 3D models inside it for 5 yuan each, I might actually buy them." This pointed to a simple insight: what users want to buy has never been the tool, but the output of the tool. In 2023, the team decided to direct their years of accumulated graphics capabilities toward this clearer outlet and launched Meshy, the world's first publicly accessible 3D generative AI product.
Meshy's leadership also has a structural reason. The over 100 million model generations from 12 million registered users have precipitated not just the 3D data itself, but also user behavior signals: which models are downloaded, printed, or discarded and regenerated, and which combinations of prompts and parameters lead to higher satisfaction. In a track where 3D data is naturally scarce, the accumulation of such preference data is extremely difficult for latecomers to replicate. Coupled with the speed of model iteration—from the first generation in early 2023 to Meshy 6 in early 2026, with a major version update on average less than every half year—each generation is widening the gap with latecomers in geometric accuracy, topological quality, and generation speed.
Why AI 3D Can Make Money First
Users of most AI generative products are paying for inspiration and convenience. Text-to-image allows designers to quickly produce concepts, and text-to-video saves content creators shooting costs. These needs are real, but they also mean users have a higher tolerance for flaws; if an AI image has issues, you can generate another one, and if an AI video has glitches, you can change the angle.
The downstream of 3D is different. Game developers need models that can be directly imported into Unity or Unreal, with controllable polygon counts, clean topology, and correct skeletal rigging. 3D printing users need closed solid files; any issue with thin walls, holes, or self-intersecting faces will lead to print failure, wasting not just time but also materials. For these users, "looking good" is merely the ticket to entry; "being usable" is the prerequisite for payment.
This difference creates an interesting result. Users of 3D generation are more like purchasing means of production, while users of image and video are more like purchasing inspiration services. The former has more rigid willingness to pay, more stable repurchase rates, and lower price sensitivity. The 3D track, which appears smaller, might actually be the direction in AI multimodal generation that first develops a healthy business model.
Meshy is the company that has gone the furthest on this path. It has pushed AI-generated 3D models from "looking like it on the screen" to a state of "being ready for production." To achieve this, it is necessary to solve a series of graphics-level engineering problems such as geometric closure, normal orientation, and topological density control during the generation process, which is also the most direct manifestation of the graphics barrier at the product end. The generated results support mainstream formats such as FBX, OBJ, GLB, and STL, and can be directly imported into Unity, Unreal, and Blender. According to Meshy's data, the slicing success rate for 3D printing reaches 97%. Work that once required a professional art team, expensive DCC software, and two weeks can now be completed in just a few minutes.
This product capability has opened up two growth engines: one is game development, and the other is consumer-grade 3D printing. Publicly disclosed partners include industry leaders such as Nexon, NetEase Games, 37 Interactive Entertainment, Bambu Lab, and Creality, but Meshy's client list goes far beyond this. According to its disclosure, half of the top ten global tech companies by market cap and valuation are already using its products and services. In other words, Meshy has sold its products to some of the most demanding top-tier clients globally.
Meshy's recent updates in model capabilities happen to target the most urgent needs of users on both ends.
For game developers, Smart Topology is the latest generation of self-developed models, generating 3D models with native parts in about 10 seconds. The polygon count is precisely controllable (from 100 to 15,000), requiring no manual decimation or retopology, and directly outputting assets ready for real-time rendering. 8K Texture pushes the resolution and quality of AI textures to a new height, with significant improvements in detail retention, color reproduction, and lighting removal cleanliness. For developers who need to put AI-generated models directly into game engines, these two updates directly reduce post-processing steps.
For 3D printing users, Auto Split compresses the previously most time-consuming part-splitting process into a one-click operation. Seams are cut along the natural structure of the model, and parts can be accurately assembled and restored. The splitting results have been verified on physical printers. From model generation to splitting to printing on the machine, the entire process is closed within Meshy, and the actual effect is at the industry's leading level.
During the printing phase, we used an astronaut model that retains more geometric details and enabled Auto Split. The system identified the astronaut as 19 independent parts. In assembly mode, the parts can be restored to the complete figure; switching to "On Plate," parts such as the helmet, backpack, limbs, wrench, and base are automatically laid out. Subsequently, clicking on Bambu Studio, Meshy will generate a 3MF file and invoke the local slicing software. The manual division, surface filling, and arrangement that previously needed to be done in Blender or Meshmixer are now consolidated into the same workspace.
The assembly view intuitively shows the splitting relationship of each part.
Switching to "On Plate," the 19 parts are automatically laid out and can be continuously sent to slicing software such as Bambu Studio.
For users on both ends, the more post-processing steps that can be eliminated, the more justified their reasons to stay within Meshy.
On the 3D printing side, there is another underestimated driver. Bambu Lab's revenue nearly quadrupled from 2023 to 2025, breaking through RMB 10 billion, with a global consumer-grade market share of about 29%. Desktop printers are bringing 3D printing into homes and personal studios, but the first question for the influx of new users is what to print. Most people cannot model. After Meshy integrated with Bambu Lab's MakerLab, users can upload an image and get a file ready for direct slicing and printing. An exploding hardware market and a just-matured AI generative tool collided in 2025 to 2026.
At this year's GDC, Yuanming Hu mentioned a figure: the number of 3D models generated by Meshy in three years has already approached the total number of 3D models manually created by humans in history. After the camera liberated image production from painters, humans took about 85 billion photos a year at the peak of the film era. Once a technology reduces the cost of creation low enough, the supply of content will shift from linear growth to exponential release. 3D might be standing right at this inflection point.
Modeling Through Chatting
Smart Topology, 8K Texture, and Auto Split solve the production efficiency problems of professional users. 3D Agent does something else: it opens the door and lets more people in.
3D Agent is the world's first conversational 3D creation agent, now open to all registered users. Users do not need any 3D experience; through conversation, text, photos, or sketches, they can complete the entire process from creative brainstorming to 3D finished products in a single conversation. The generated models can be directly sent to slicing software for printing, or imported into Unity, Unreal, and Blender with skeletal rigging and animations.
In the testing of 3D Agent, we input the prompt: "I have absolutely no 3D modeling experience and want to make a retro astronaut collectible ornament suitable for FDM printing. Please first provide three visually distinct directions that are easy to print and split, and do not generate the 3D model immediately." The Agent did not rush to generate randomly, but first sorted out the printing principles, providing three directions: hardcore aerospace, constructivist cartoon, and steampunk. It also listed splitting suggestions for the helmet, torso and backpack, arms and wrench, and base.
After confirming the "chibi proportions with steampunk mechanical details," we asked it to generate three front three-quarter view concept images at once. The Agent didn't just provide images; it further compared the color tones, helmet, chest structure, backpack, and FDM friendliness of the three sets of solutions. After choosing Solution A, we used one sentence to ask it to convert it into 3D. A few minutes later, a complete astronaut with a circular base, a square life-support backpack, and a wrench appeared in the preview area.
3D Agent generates three sets of concept images at once and helps users choose from two dimensions: styling and print friendliness.
In the 3D model generated after selecting the solution, design elements such as the character, backpack, wrench, and base are all retained.
Throughout the entire process, we did not adjust the mesh, normals, or UVs. After exporting the STL, scaling the model height to about 120 mm allowed it to be normally placed on the print bed of the Bambu Lab A1 and sliced. The value of 3D Agent lies not only in that final model but in breaking down the professional process of "unclear idea—determining direction—comparing solutions—generating the final product," which originally required creators to master themselves, into a conversation that ordinary people can complete. The astronaut used in the previous Auto Split test came from this very conversation.
After the generated STL is actually imported into Bambu Studio, it can be normally placed on the A1 print bed and continue slicing.
Over the past two years, AI has lowered the barriers to writing, painting, programming, and video production, but 3D has always been the last high ground. The learning curve for professional 3D modeling software is measured in months or even years, keeping the vast majority of people with creative impulses out. 3D Agent lowers the entry point from "learning a professional software" to "opening a dialog box."
On one end, Smart Topology, 8K Texture, and Auto Split keep professional users staying and paying more; on the other end, 3D Agent lets more people in. Pushing from both ends simultaneously, one increases the average order value and retention, while the other expands the top of the funnel.
Beyond 3D, Something Bigger
If looking only at the existing commercialization achievements, Meshy is a very good 3D tool company. But the nearly $400 million Series B is clearly not pricing for a tool market.
Meshy defines its next phase as "real-time interactive multimodal content experiences." Translated into more specific terms, 3D model generation is the foundation; upon this foundation, interactive mechanisms, AI real-time generation, and multimodal fusion are superimposed to build a content engine and platform that people can enter, participate in, and create within.
Meshy Game Studio officially debuted, bringing its first AI game demo, "Black Box: Infinite Construction." It allows players to use natural language to drive weapon generation in real time, with ballistics, physical effects, and element combinations created by AI at runtime, with no preset scripts. This demo is still very early, but it demonstrates a form of content that previously almost did not exist: content emerging in real time during user participation, rather than being produced in advance and waiting to be consumed.
The customers for 3D model generation are creators and developers, numbering in the tens of millions. The users of real-time interactive content are consumers, numbering in the billions. The camera gave birth to the film industry, synthesizers gave birth to the electronic music industry, and game engines gave birth to a real-time 3D application ecosystem far beyond games themselves. Every capability leap in creation tools ultimately spawns new content categories and new consumer markets. Meshy's judgment is that AI-native 3D generation capabilities are also standing at a similar starting line.
Meshy has a prerequisite for doing this that most competitors lack: it already possesses production-grade usable 3D generation capabilities and a validated commercialization path. Other heavyweight players in the spatial intelligence direction are still in the early stages of their products, whereas Meshy is growing upward from a foundation that is already making money.
In AI multimodal generation, text exploded first, followed closely by images, and video has now become the new main battlefield. In contrast, 3D has always been the quietest line. But quietness and stagnation are two different things; the barriers kept out most players, but also carved out a market that no one is fighting for and where the willingness to pay is the most rigid.
Humans live in a three-dimensional world. When AI-generated 3D content is good enough, fast enough, and cheap enough, every direction—gaming, education, design, manufacturing, and spatial computing—unfolds into an independent market. 100 million AI-generated 3D models have already been created. Nearly $400 million has just been bet.