Manycore Tech IPO: The Spatial Intelligence Bet Behind China's First 'Hangzhou Six Dragons' Listing

In April of 2026, a company called Manycore Tech—known in China as Qunhe—rang the bell on the Hong Kong Stock Exchange. And it did something dramatic: the stock opened up more than 171 percent above its issue price, pushing the market cap past thirty billion Hong Kong dollars. What makes this interesting isn't just the pop. It's that Manycore became the very first of Hangzhou's so-called "Six Dragons" to go public. Those Six Dragons are a cluster of buzzy AI-adjacent startups—the group includes DeepSeek and the robotics maker Unitree, among others. So this was a milestone.

Now, if you don't know Manycore, you might know its main product. It's called Kujiale—a cloud-based 3D interior design platform that Chinese home-renovation designers absolutely love. On paper, that sounds like ordinary design software. But the story they told investors going into this IPO wasn't "design software." It was two words: spatial intelligence. Let's unpack what that actually means, and whether the technology holds up.

So, spatial intelligence. There's no single textbook definition, but in practice it bundles together four capabilities—all centered on perceiving, modeling, and reasoning about physical, three-dimensional space.

The first is scene understanding. Not just "there's a sofa in this room," but the geometric and semantic relationships: the sofa sits 1.2 meters to the left of the coffee table, and it's facing the TV. It understands how things are arranged.

The second is spatial reasoning. This is where you answer questions that require genuine 3D logic. For example: if I move this chair over here, is the walkway still wide enough for a wheelchair to pass? That's not a text question. That's a geometry question.

Third is 3D reconstruction—rebuilding a full three-dimensional model of a space from raw inputs like photos, LiDAR scans, or floor plans.

And fourth is navigation and path planning—figuring out a viable route through a 3D environment. That's the beating heart of what mobile robots and AR systems need.

Manycore's platform, which they call SpatialVerse, leans mostly on the first and third of those—scene understanding and 3D reconstruction. And in 2025 they wrapped that up into an API product called Aholo, so outside developers can tap into it.

Okay, but here's the real question. Why is 3D scene data suddenly worth so much? And to understand that, you have to understand a gap that anyone following embodied AI already knows about.

Think about the big language models—your GPT-4-class systems. They were trained on text and 2D images. They can talk about physical space in a linguistic way, sure. But they have no real spatial intuition. No grounding in the actual geometry of the world. Ask them to genuinely reason about how furniture fits in a room, and there's nothing underneath the words.

Now, embodied AI—robots, agents that have to physically move through the world and manipulate objects—those systems need something completely different. They need training data that encodes how real spaces are structured, how objects relate in three dimensions, how physical interactions actually play out. And that kind of data? It's scarce. Really scarce.

This is where Manycore's claim gets compelling. They say they hold more than 70 million structured indoor 3D scene models. And here's the key detail: each one was built by a real designer using Kujiale's tools, for a real renovation project. So you get accurate object positions, real dimensions, actual material properties. This isn't messy photogrammetry, and it isn't synthetic scenes churned out by an algorithm. These are real designs tied to real spaces. That combination—massive volume plus semantic richness—that's the moat.

Let me describe how the whole thing fits together, because there's a clean flow to it.

At the bottom, you have the data layer. That's the Kujiale platform with its 70 million-plus structured 3D scenes, plus all the user design-behavior data—how people actually build these rooms.

That data feeds up into a model layer with three functions. One model handles scene understanding—detecting objects and their spatial relations. Another does generative 3D—automatically suggesting layouts. And a third does spatial question-answering, translating natural language questions into spatial queries.

And then those models feed the application layer. The scene-understanding model powers the external SpatialVerse and Aholo API for outside developers, and it also generates training scenes for embodied AI, and it drives AR and VR scene generation. Meanwhile the generative model powers Kujiale's own AI design assistant. So the same core data flows out in several directions—internal product on one side, external AI infrastructure on the other.

That Aholo platform is really the front door for the outside world. It hands robotics teams, AR app builders, and architectural software vendors a set of APIs and SDKs, so they can plug into Manycore's spatial data without having to build their own scene-understanding stack from scratch.

Now, a quick word on the founder, because it explains a lot. His name is Huang Xiaohuang. He came out of Zhejiang University's honors college, got a master's in computer science from the University of Illinois at Urbana-Champaign, and then—here's the telling part—he spent 2010 to 2011 as a software engineer at NVIDIA, working mostly on CUDA. So this is someone deep in GPU computing. He founded Manycore in 2011, and the original goal was almost mundane: take desktop CAD tools—AutoCAD-style design software—and move them into the browser and the cloud.

And that's the fascinating twist. What looked like a boring "let's put CAD in the cloud" SaaS play was quietly building a data flywheel. More designers on the platform means more 3D scene data. Better scene data enables better AI-assisted design features. And better features pull in even more designers. Round and round. Fifteen years of that spinning is exactly how you end up with 70 million scenes.

Let's talk money for a second, because the valuation is doing something specific. Manycore booked revenue of about 755 million RMB in 2024, and 399 million in the first half of 2025—growing around 9 percent year over year. The company isn't profitable yet, though its losses have been narrowing. So if you value it purely on today's SaaS numbers, that huge IPO pop doesn't make sense. The valuation isn't really about the software business. It's a bet—a bet that structured spatial scene data becomes a critical training asset for the next generation of AI. That thesis is what flipped the market's framing from "unprofitable design software" to "scarce AI infrastructure." Same company, totally different story.

And it's worth being precise about how this differs from the computer vision AI you might already know. Classic CV—think YOLO for object detection, or ResNet for image classification—works at the pixel level, in 2D. "Does this image contain a cat?" Spatial intelligence operates a whole level up. The questions sound like: Can a wheelchair get from the entrance to the kitchen in this layout? Does this furniture arrangement make for an ergonomically sane work-from-home setup? Which objects would fall over if I remove this load-bearing element? To answer any of those, you need 3D geometry, semantic understanding of how objects relate, and physical common sense. And none of that just falls out of training on flat 2D images.

So let me pull together the takeaways.

First, the big pattern worth watching: vertical SaaS companies that spent years quietly accumulating large, structured, domain-specific datasets may end up with far harder-to-copy moats than flashy AI startups built from scratch. Manycore didn't set out to build AI infrastructure—it accidentally built the dataset for it, one designer at a time.

Second, the whole valuation rests on a reframing. Nothing about the revenue changed. What changed is the market deciding that scarce 3D spatial data is strategic AI infrastructure, not just a design-tool byproduct.

And third, whether that bet pays off is genuinely conditional. It depends on how fast humanoid robotics and immersive AR hardware actually mature. If those technologies arrive, the companies sitting on high-quality 3D scene data get a real head start. If they stall, this is just design software with an expensive story attached. That's the tension baked into this IPO—and it's why the first of the Six Dragons to list did it on the strength of data it's been gathering, almost by accident, for fifteen years.