Manycore Tech, a Hangzhou-based AI infrastructure company, debuted on the Hong Kong Stock Exchange on April 17, 2026, surging 187% from its IPO offer price in early trading. The company raised $156 million in its offering — the largest first-day percentage gain by an AI company on the Hong Kong exchange this year — after repositioning from robotics manufacturing to supplying AI training data to robot makers.
This is not a software company that added “AI” to its marketing deck. Manycore collects, labels, and structures the real-world motion, sensor, and visual data that physical AI models require to learn how to interact with the world. The pivot is surgical, and the market rewarded it immediately.
From Robots to Data: The Strategic Pivot That Changed Everything
Manycore’s original business was robotics hardware — a crowded space in China where margins are thin and competition from Unitree, UBTECH, and dozens of smaller manufacturers is intense. The company recognized early that the real bottleneck in the physical AI stack wasn’t the hardware. It was the data needed to train the intelligence inside it.
The company repositioned as a data infrastructure provider for the robotics industry. Rather than competing on cost for actuators and frames, Manycore now sells annotated datasets covering robotic motion, environmental sensing, and manipulation tasks — the kind of data that takes months to collect at scale and requires specialized facilities to produce reliably.
Physical AI models require between 10 million and 100 million labeled data points to reach production-grade performance, according to estimates from researchers at Carnegie Mellon’s Robotics Institute and MIT’s CSAIL. Unlike language models, which can scrape text from the open web, robot training requires real-world capture setups: controlled environments, precision sensors, and human annotators who understand physical mechanics. That infrastructure is expensive to build from scratch. Manycore sells access to it.
Manycore’s Hong Kong IPO: The 6M Raise and 187% First-Day Pop
Manycore priced its IPO targeting approximately $156 million in gross proceeds — modest by US tech standards but substantial for the current Hong Kong exchange environment, where IPO volume remains well below 2021 peaks. The 187% surge on debut day pushes the implied market capitalization significantly above initial projections and signals strong institutional appetite for AI data infrastructure plays.
The Hong Kong Stock Exchange has been working to attract more technology listings since Alibaba’s secondary listing in 2019. Manycore’s debut represents a test case: a pure-play AI data company with a clear B2B revenue model, not a consumer platform chasing growth at any cost. Nebius’s $10 billion data center commitment in Finland signals a parallel dynamic globally — infrastructure-layer AI companies are attracting outsized capital as the market matures past frontier model hype.
For context on how rare 187% first-day gains are: the average first-day return for US tech IPOs in 2024 was 23%, according to Renaissance Capital. Even accounting for the smaller float and thinner liquidity typical of Hong Kong debut trading, a near-tripling on day one represents extraordinary demand compression from investors who see the robotics data supply crunch coming.
Why Robotics Training Data Is Now Worth More Than the Robots
The physical world requires physical data. General-purpose web text cannot train a robotic arm to sort produce or navigate a warehouse floor under varying lighting conditions. This is the defining constraint of physical AI development in 2026, and it is not a problem that more compute solves.
Tesla’s Optimus program consumes data from its vehicle fleet and purpose-built simulation environments. Figure AI’s partnership with OpenAI, announced in 2024, was partly premised on improving humanoid robot training pipelines — OpenAI has been aggressive about acquiring AI infrastructure capabilities across the stack. NVIDIA’s Isaac Sim platform generates synthetic training data, but synthetic data still requires real-world grounding to prevent sim-to-real transfer failures at deployment.
Chinese robot manufacturers face an additional constraint: data sovereignty concerns mean they cannot freely use training datasets generated outside China. Manycore, operating domestically and building collection infrastructure within Chinese regulatory frameworks, occupies a structurally protected position in the world’s largest robotics manufacturing market.
| Data Type | Use Case | Collection Challenge |
|---|---|---|
| Motion capture | Manipulation tasks, grasping | Specialized rigs, expert annotators |
| Environmental sensing | Navigation, obstacle avoidance | Diverse real-world environments, edge cases |
| Human demonstration | Imitation learning | Human labor at scale, safety protocols |
| Failure case data | Robustness training | Controlled failure environments |
Who Is Buying Manycore’s Data
Manycore has not disclosed a detailed customer list in its prospectus, which is standard for Chinese companies listing in Hong Kong. The likely customer profile follows the structure of China’s robotics industry: Unitree Robotics, UBTECH, Agility Robotics’ China operations, and the growing cluster of EV-adjacent humanoid programs at BYD and SAIC, which have each announced humanoid robot initiatives within the past 18 months.
MegaOne AI tracks 139+ AI tools and infrastructure providers across 17 categories. The robotics data segment has seen the fastest acceleration in new entrants over the past 18 months — companies racing to position themselves before the training data supply crunch becomes acute at scale. Manycore’s head start in physical infrastructure is the core of its competitive moat.
Global customers are a longer-term opportunity. US regulatory concerns about Chinese data provenance may limit direct sales to Tesla or Boston Dynamics parent Hyundai. But companies in Europe, Japan, and South Korea face no such restrictions, and Japanese robotics manufacturers — Fanuc, Yaskawa, Kawasaki — are historically aggressive buyers of high-quality training datasets when domestic alternatives are scarce.
The AI Data Supply Chain Is Under Pressure Everywhere
Manycore’s IPO arrives as data scarcity is becoming a first-order problem across all of AI development. Language models have largely exhausted the high-quality web corpus; researchers at Epoch AI estimated in 2024 that frontier LLM training will hit data limits between 2026 and 2028 without synthetic generation breakthroughs. Physical AI faces the same wall, arriving sooner.
The data collection layer — historically unglamorous, often outsourced to contractors — is being recognized as a strategic asset. Autonomous data collection systems like Nomad represent one approach to scaling data acquisition for unstructured environments. Manycore represents another: purpose-built physical infrastructure with domain-specific expertise, aimed squarely at the manufacturing and logistics robotics market.
The supply chain for AI training now has four distinct pressure points: compute (still GPU-constrained), energy (data centers consuming grid capacity at unprecedented rates), human feedback (RLHF labor markets tightening globally), and raw domain data — the last of which Manycore has positioned itself to address for physical AI specifically.
What the 187% Pop Signals for AI Infrastructure in 2026
Public market investors are not rewarding AI hype in 2026 the way they did in 2023. The companies attracting premium valuations share a common trait: they sit inside the AI supply chain at a position that is difficult to replicate quickly. Compute is hard to replicate. Energy infrastructure is hard to replicate. Specialized physical training data, collected in controlled environments with domain expertise, is hard to replicate.
Manycore’s first-day performance suggests the market understands the structural position: a Chinese company with domestic data collection infrastructure, domain expertise in physical AI, and no obvious local competitor at comparable scale. The 187% gain is not irrational — it is pricing in moat. Large AI companies are paying premium prices to secure supply chain positions, and Manycore has made itself a logical acquisition target or long-term infrastructure partner for any global robotics program that needs Chinese-market data access.
The IPO proceeds of $156 million give Manycore approximately 12-18 months to establish international data collection partnerships before well-funded competitors replicate its model in other geographies. That window is the company’s real strategic asset, and how it deploys this capital will determine whether today’s 187% pop marks the beginning of a durable business or a well-timed market moment.
The takeaway: Manycore’s Hong Kong debut is the most instructive AI infrastructure signal of early 2026. Not because of the dollar size — $156 million is modest against the scale of the robotics market — but because a 187% first-day gain on a B2B AI data company tells you precisely where smart capital thinks the physical AI supply chain will tighten next. Companies still focused only on their hardware roadmap should audit whether their training data strategy is equally mature.
