A Hangzhou company surged 185 percent on its first trading day by selling spatial data for robots. Its IPO marks the moment frontier AI labs stopped competing on the same axis and started choosing which world to inhabit.
Manycore Tech debuted on the Hong Kong Stock Exchange today and surged 185 percent. The company raised $156 million. Over the past decade, it accumulated 500 million 3D assets from its interior design platform, then pivoted to selling that data as training material for robot makers. Its entire IPO thesis rests on a single claim: general-purpose AI cannot enter the physical world without vertical data.
The market agreed. And the agreement tells us something about where the AI industry is heading.
After Convergence
Six weeks ago, seven frontier AI models launched from six organizations in twenty-nine days. The top four scored within three percentage points of each other on standard benchmarks. This journal called it The Convergence. The base layer of intelligence became a commodity.
What happens after convergence is specialization.
On April 16, OpenAI launched GPT-Rosalind, its first domain-specific model. Built for drug discovery, genomics, and protein engineering, it ships only to trusted-access customers: Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific. A Codex plugin connects it to more than fifty scientific tools. Bloomberg headlined the release as OpenAI taking on Google in drug discovery.
On the same day, Anthropic shipped its latest model with cybersecurity protections derived from Mythos, the system it withheld from public release because it could find and exploit thousands of zero-day vulnerabilities across major operating systems and browsers. Through Project Glasswing, Anthropic committed $100 million in usage credits to handpicked security partners. The capability that made the model dangerous became the product.
Two days earlier, Google released Gemini Robotics-ER 1.6, designed for physical instrument reading and real-world operation alongside Boston Dynamics hardware. Google had already upgraded Gemini 3 Deep Think in February, which solved four open mathematical conjectures autonomously and scored 48.4 percent on Humanity's Last Exam. Two verticals from one company: the physical world and abstract reasoning.
Meta maintained its open-source approach. General-purpose models, no domain restrictions, no access gates.
The Confession
The choice of vertical is a confession about moat.
OpenAI chose life sciences because it believes impact is measurable. Drug discovery has clear endpoints: molecules that work, trials that succeed, patients that improve. Partnerships with Amgen and Moderna create switching costs that no benchmark score can replicate. If GPT-Rosalind accelerates even one drug through Phase II, OpenAI owns the most defensible customer relationship in AI.
Anthropic chose cybersecurity because it believes capability is dangerous. The decision to withhold Mythos and gate access through Glasswing turns a safety concern into scarcity. Every competitor can build a general-purpose model. The one too powerful to release has a different value proposition entirely.
Google chose the physical world because it believes ubiquity survives commoditization. From digital reasoning to robotic manipulation to mathematical proof, Google is building across every substrate simultaneously. Manycore's IPO validates this bet from the supply side. Someone has to provide the 3D spatial data those robots train on. Google is building the demand.
Meta chose distribution. Open-source models become the default architecture for anyone who cannot afford proprietary access. In a market of vertical specialists, the horizontal layer captures the volume.
The Pattern
This is not unprecedented. After PC hardware commoditized in the 1990s, value migrated to operating systems, then applications, then data. After smartphone hardware converged, value migrated to app stores, then services. Convergence has always been the trigger for specialization.
What is new is the speed. Six weeks from convergence to vertical divergence. The base capability became undifferentiated in a single quarter. The labs responded within days.
Manycore crystallizes the logic. Five hundred million 3D assets from a decade of interior design work are valued at $156 million as robot training data. The moat is domain-specific accumulation over time, something no general-purpose competitor can shortcut.
The vertical bet is which world to inhabit. And the choice reveals what each lab believes is hardest to copy: partnerships, restrictions, physical presence, or distribution. The models converged. The strategies diverged. That divergence is where the next decade of value will be determined.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)