Table of Contents

In April 2026, Manycore Tech (群核科技) listed on the Hong Kong Stock Exchange, becoming the first company among Hangzhou’s so-called “Six Dragons”—a cohort of AI-adjacent startups including DeepSeek, Unitree Robotics, and others—to complete an IPO. The stock opened 171.65% above its issue price, putting the market cap above HK$30 billion. For engineers unfamiliar with the company, its core product is Kujiale (酷家乐), a cloud-based 3D interior design platform popular with Chinese home renovation designers. The narrative it brought to the IPO, however, was “spatial intelligence.” This piece unpacks the technology behind that claim.

TL;DR

Spatial intelligence refers to a cluster of AI capabilities centered on perceiving, modeling, and reasoning about three-dimensional physical space. Manycore’s angle: after 15 years of operating Kujiale, it holds a database of over 70 million structured indoor 3D scenes—data that is increasingly scarce and valuable for training embodied AI systems.

What Spatial Intelligence Actually Means

The term lacks a single authoritative definition, but in practice it covers four capability areas:

Scene understanding: Identifying objects in a space and modeling their geometric and semantic relationships—not just “there is a sofa” but “the sofa is 1.2 meters to the left of the coffee table, facing the TV.”

Spatial reasoning: Answering questions that require 3D logic—“if I move the chair here, is the walkway still passable for a wheelchair?”

3D reconstruction: Rebuilding a complete 3D scene model from photos, LiDAR scans, or floor plans.

Navigation and path planning: Computing viable paths through known or unknown 3D environments, the core requirement for mobile robots and AR overlays.

Manycore’s SpatialVerse platform focuses primarily on scene understanding and 3D reconstruction, and packages that capability as an API for external developers through its Aholo platform (released in 2025).

Why 3D Scene Data Is Suddenly Valuable

The gap here is well understood by anyone following embodied AI research. GPT-4-class LLMs were trained primarily on text and 2D images. They can reason about physical space in a linguistic sense but have no genuine spatial intuition—no grounding in the geometry of the actual world.

Embodied AI systems (robots, autonomous agents that must navigate and manipulate physical objects) need something different: training data that encodes how real spaces are structured, how objects relate to each other in three dimensions, and how physical interactions play out. That training data is scarce.

Manycore claims more than 70 million structured indoor 3D scene models, each built by a real designer using Kujiale’s tools, with accurate object positions, dimensions, and material properties. Unlike photogrammetry captures or synthetic procedural scenes, these are designs that correspond to real renovation projects. That combination of volume and semantic richness is the moat.

How the Platform Works

graph TB
  subgraph "Data Layer"
    D1["Kujiale Platform\n70M+ Structured 3D Scenes"]
    D2["User design behavior data"]
  end

  subgraph "Model Layer"
    M1["Scene Understanding\nObject detection + spatial relations"]
    M2["Generative 3D\nAuto layout suggestions"]
    M3["Spatial QA\nNatural language → spatial queries"]
  end

  subgraph "Application Layer"
    A1["SpatialVerse / Aholo API\nExternal developers"]
    A2["Embodied AI training scenes"]
    A3["AR/VR scene generation"]
    A4["Kujiale AI design assistant"]
  end

  D1 --> M1
  D1 --> M2
  D2 --> M3
  M1 --> A1
  M2 --> A4
  M1 --> A2
  M1 --> A3

The Aholo platform (2025) is the external-facing interface, providing APIs and SDKs so that third-party developers—robotics teams, AR application builders, architectural software vendors—can integrate Manycore’s spatial scene data without building their own scene understanding stack.

Founder Background

Huang Xiaohuang (黃曉煌), founder and chairman, holds a BS from Zhejiang University’s Chu Kochen Honors College and an MS in computer science from the University of Illinois Urbana-Champaign. From 2010 to 2011 he worked as a software engineer at NVIDIA, primarily on CUDA development. He founded Manycore in 2011 with the initial goal of moving desktop CAD tools (AutoCAD-class design software) to the browser and cloud.

That SaaS trajectory—which initially looked like a standard cloud software story—turned out to be building a data flywheel: more designers on the platform means more 3D scene data; better scene data enables better AI-assisted design features; better features attract more designers.

Financials and Valuation Logic

Manycore reported revenue of RMB 755 million in 2024 and RMB 399 million in H1 2025 (9.4% year-over-year growth). The company is not yet profitable, but its losses have been narrowing. The IPO valuation reflects less the current SaaS metrics and more the bet that structured spatial scene data will become a critical training asset for the next AI generation—a thesis that shifted the market’s framing from “unprofitable design software” to “scarce AI infrastructure.”

Difference from Traditional Computer Vision AI

Classic computer vision AI (YOLO for object detection, ResNet for image classification) operates on 2D pixel-level questions: “Does this image contain a cat?” Spatial intelligence operates at a higher semantic level: “Can a wheelchair pass from the entrance to the kitchen in this layout?” “Does the placement of this furniture create an ergonomically viable work-from-home setup?” “Which objects would fall if I remove this load-bearing element?”

Answering these questions requires 3D geometry reconstruction, semantic understanding of object relationships, and physical common-sense reasoning—none of which emerge naturally from 2D image training data alone.

Summary

Manycore’s IPO illustrates a pattern worth watching across the AI era: vertical SaaS companies that accumulated large, structured, domain-specific datasets over years of real-world use may have harder-to-replicate moats than pure AI startups built from scratch. Whether spatial intelligence becomes a major AI battleground depends on how quickly humanoid robotics and immersive AR hardware mature. If they do, the companies holding high-quality 3D scene data will have a meaningful head start.

References

🇺🇸 English

Here is a company that spent fifteen years quietly building what might be the most valuable AI training dataset you've never heard of — and in April 2026, the market noticed.

Manycore Tech, the company behind a cloud-based 3D interior design tool called Kujiale, went public on the Hong Kong Stock Exchange and opened one hundred seventy-one percent above its issue price. That's not a typo. Day one, the stock more than doubled, pushing the market cap past thirty billion Hong Kong dollars. And Manycore isn't just any startup — it's the first company from what China's tech press calls the "Hangzhou Six Dragons" to actually IPO. That cohort includes names like DeepSeek and Unitree Robotics. So this listing is being watched carefully.

But here's what's interesting for engineers: this isn't a story about design software. It's a story about data strategy and a concept called spatial intelligence.

**What spatial intelligence actually means**

Spatial intelligence is a cluster of AI capabilities built around perceiving, modeling, and reasoning about three-dimensional physical space. Think of it as four distinct problems stacked together.

First, scene understanding — not just detecting that a sofa exists in a room, but knowing it's 1.2 meters to the left of the coffee table, angled toward the TV, with specific dimensions and material properties.

Second, spatial reasoning — answering questions that require 3D logic. Can a wheelchair navigate from the front door to the kitchen in this floor plan? If I move that chair, does it block the walkway?

Third, 3D reconstruction — taking photos, floor plans, or sensor scans and rebuilding a complete geometric model of a space.

And fourth, navigation and path planning — computing viable routes through environments, which is the core engine behind mobile robots and augmented reality overlays.

Manycore's platform focuses mainly on the first two: scene understanding and reconstruction. They've packaged this into an API called Aholo, released in 2025, so robotics teams, AR developers, and architectural software companies can integrate their spatial intelligence stack without building it from scratch.

**Why the data is the real story**

Here's the gap that makes this interesting. The large language models we've all been using — the GPT-class, the Claude-class — were trained primarily on text and two-dimensional images. They can discuss physical space in a linguistic sense, but they have no genuine spatial intuition. No grounding in real geometry.

Embodied AI systems — robots that have to navigate a room, manipulate physical objects, understand how furniture relates to doorways — need something fundamentally different. They need training data that captures how real spaces are actually structured in three dimensions, with accurate object positions, sizes, and relationships. And that data is genuinely scarce.

Manycore claims over seventy million structured indoor 3D scene models. Every single one was built by a real designer using Kujiale's tools for an actual renovation project. These aren't synthetic scenes generated by a computer, and they aren't raw photogrammetry captures. They're semantically rich, professionally authored designs that correspond to real-world spaces — with accurate object dimensions, material properties, and spatial relationships all baked in.

That combination of volume and semantic richness is the argument for why this data is hard to replicate.

**How the flywheel was built**

The founder, Huang Xiaohuang, studied computer science at the University of Illinois and spent time as a CUDA developer at NVIDIA before founding Manycore in 2011. His original goal was straightforward: move desktop CAD tools to the browser. Take what AutoCAD does and make it cloud-native.

What he actually built, almost by accident, was a data flywheel. More designers on the platform meant more 3D scene data accumulating. Better scene data enabled better AI-assisted design features. Better features attracted more designers. Fifteen years of that loop produced seventy million structured scenes.

The platform architecture flows naturally from this. At the base, you have the Kujiale design tool collecting scene data and user behavior. That feeds into models for scene understanding, generative layout suggestions, and spatial question-answering. Those capabilities then surface through the Aholo API to external developers, through AI design assistants inside Kujiale itself, and — increasingly — as training data infrastructure for embodied AI systems.

**This isn't your standard computer vision AI**

It's worth being precise about what makes this technically distinct from classic computer vision work. Traditional CV systems — object detection, image classification — operate on two-dimensional pixel-level questions. Does this image contain a cat? Where in the frame is it?

Spatial intelligence operates at a higher semantic level entirely. Does this furniture arrangement create a viable work-from-home setup? Which objects would shift if you remove that structural element? These questions require reconstructing 3D geometry, understanding semantic relationships between objects, and applying physical common-sense reasoning. None of that emerges naturally from training on 2D images alone.

**The financial picture**

Manycore reported revenue of around 755 million renminbi in 2024, and about 400 million in the first half of 2025. Not yet profitable, but losses are narrowing. The IPO valuation isn't really priced on those SaaS metrics — it's priced on the thesis that structured spatial scene data will become a critical training asset for the next generation of AI, particularly as humanoid robotics and immersive AR hardware mature. The market reframed the company from "unprofitable design software" to "scarce AI infrastructure." That framing is doing a lot of work in that 171% opening gain.

**Three things worth taking away from this**

One: Vertical SaaS companies that accumulated large, structured, domain-specific datasets over years of real-world use may have moats that are genuinely harder to replicate than pure AI startups built from scratch. The data asset compounds with time in a way that compute spend doesn't.

Two: Spatial intelligence is the missing layer between language models and embodied AI. The companies that hold high-quality 3D scene data have a meaningful head start if robotics and AR hardware hit their growth curves.

Three: The Hangzhou Six Dragons framing is worth watching. Manycore went first. DeepSeek and Unitree are presumably next. The pattern of deep technical specialization — pick a hard problem, build data infrastructure around it for years — seems to be the thesis the market is rewarding right now.

🇹🇼 中文

2026年4月，一家你可能沒聽過的公司在香港交易所上市，首日漲幅171%，市值衝上300億港元。這家公司叫群核科技，旗下有個你可能用過的產品——酷家乐，就是那個讓室內設計師做3D裝潢方案的工具。

但這次它說的故事，不是設計軟體，而是「空間智能」。

它也是「杭州六小龍」裡面第一個完成IPO的——那批被拿來和北京五道口AI創業圈比較的杭州AI新創群。今天我們來拆解兩件事：空間智能到底是什麼，以及為什麼一家做家裝SaaS的公司，能在AI時代說出這個故事。

---

**空間智能是什麼**

簡單說，就是讓AI能夠理解、建模、並推理三維物理空間的能力集合。這涵蓋幾個維度。

第一是場景理解——從圖片或3D掃描裡認出物件，理解它們的空間關係。「這個沙發在桌子左邊，距離大約1.2米」，這就是場景理解。

第二是空間推理——在三維空間裡做邏輯推理。「把椅子移到這裡，走道還夠寬嗎？」這種問題，傳統AI答不好，因為它沒有空間感。

再往下還有3D重建、導航與路徑規劃——機器人、AR應用需要的，在3D空間裡找可行路徑的能力。

群核科技的SpatialVerse平台主要聚焦在前兩個，並透過API把這些能力開放給外部開發者。

---

**為什麼現在突然重要**

背景要先說清楚：GPT-4這類大語言模型，訓練資料幾乎都是文字和2D圖片。它們在語言和圖片上很強，但對三維空間的物理關係基本上是盲的——你問它「桌子傾斜後杯子會怎樣」，它可以推理，但沒有真正的空間直覺。

具身AI就不一樣了。那種能在物理世界移動、操作物件的AI Agent，比如機器人，需要的是三維空間的感知與推理。訓練這種能力，你需要大量高品質的3D空間資料——而這類資料極度稀缺。

群核科技說：我有7000萬個結構化室內3D場景，都是真實設計師用酷家乐建立的，有精確的物件位置、尺寸、材質資訊。這不是爬蟲抓的，是人工生成的結構化資料。對具身AI訓練來說，含金量完全不同。

---

**技術怎麼跑**

群核科技的技術路線分三層。

最底層是資料層：7000萬個室內3D場景，加上用戶設計行為資料，這是整個架構的地基。

中間是模型層，有三個主要模型——一個做場景理解和空間關係分析，一個是生成式3D模型、能自動提供佈局建議，還有一個是空間問答模型、讓你用自然語言查詢空間資訊。

最頂層是應用層：對外有SpatialVerse API，對內有酷家乐的AI設計助手，同時這些場景資料也輸出給具身AI和AR/VR應用做訓練素材。2025年發布的Aholo平台，就是他們把這套能力正式對外開放的介面。

---

**跟傳統AI的本質差別**

傳統電腦視覺問的是：「這張圖裡有沒有貓？」

空間智能問的是：「這個房間的沙發擺法符合人體工學嗎？從門口到廚房的最短路徑是什麼？這個空間能讓輪椅通行嗎？」

要回答這些問題，AI不只要識別像素，還要重建3D幾何、理解物件的語意關係、具備物理常識推理。難度量級完全不在同一層次。

---

**創辦人和資料飛輪**

創辦人黃曉煌是浙大竺可楨學院出身，UIUC電腦科學碩士，在NVIDIA做過CUDA工程師。2011年創立群核科技，切入點其實很務實——把AutoCAD這類桌面軟體搬到雲端，讓更多設計師用得起。

十五年後回頭看，這個路線無意間建立了一個資料飛輪：設計師越用，3D資料越多；資料越多，AI輔助越精準；設計越精準，設計師越願意用。這個飛輪轉了十五年，最後轉出了一個具身AI時代需要的稀缺資源。

財務上，2024年營收約7.5億人民幣，還沒盈利，但虧損在縮窄。市場給它高估值的邏輯，不是SaaS的盈利能力，而是「空間智能資料稀缺性」這個框架。

---

好，總結三個核心要點。

第一，空間智能的真正門檻不是模型，是結構化3D資料——模型可以複製，但十五年積累的高品質場景資料很難快速追上。

第二，具身AI開闢了一個新的資料需求。傳統互聯網資料對訓練機器人幫助有限，垂直領域的結構化3D資料是新的稀缺資源，而且這個稀缺性還在早期。

第三，群核科技這個賭注和具身AI整個賽道的命運是綁在一起的。空間智能能不能成為AI的下一個主戰場，關鍵在於機器人和AR/VR硬體的普及速度——這是它估值故事裡最大的變數。

Tags

#ai#spatial-intelligence#ipo#startup#computer-vision

Related Articles