On August 3, capital dropped two nine-figure bombs on the embodied AI track.

PokeBot closed a hundred-million-dollar Pre-A round co-led by Shunwei Capital and Matrix Partners; the same day, Qiuzhi Tech announced a $100M+ angel+ round. Neither company has been around long; both raised hefty sums. But what's notable isn't the amount — it's their focus: neither builds robot hardware, both build robot "brains".

From Limbs to Brains

This directional shift is critical.

Looking back at the embodied AI funding boom of 2024-2025, money mainly flowed two ways: the "body camp" — building robot bodies, motors, reducers, dexterous hands, competing on whose hardware specs were higher, whose movements were smoother; and the "Demo camp" — a viral video of a robot walking / pouring water / opening doors, then a big funding round, with questions of mass production and real utility left for later.

But in H2 2026, the wind has clearly shifted. Capital is starting to ask a question it didn't ask much before: what exactly can your robot do?

Walking prettily or pouring water steadily means nothing if it only performs preset motions — that's an expensive toy, not a productivity tool. What makes a robot work autonomously in complex environments isn't how fast the legs move or how dexterous the hands are; it's how smart the brain is — can it understand natural language instructions, recognize unfamiliar objects, make autonomous decisions in anomalous situations, learn and improve from data?

That's why money is flowing to the "brain camp." PokeBot builds a general robot manipulation brain; Qiuzhi Tech builds embodied multimodal world models — both solving the "how robots think" problem. Hardware gaps are shrinking fast (the motors and reducers you can buy, your competitor can buy too), but the brain gap is the real moat.

The Math Behind ¥43.8B

Securities Times statistics tell the story: in the first half of 2026, domestic embodied AI funding totaled roughly ¥43.8 billion, with over half going to companies focused on robot brains, visual perception, and world models. Pure hardware companies' funding share continues to decline.

A QbitAI article yesterday put it bluntly in the headline: "The High-Valuation Demo Era Is Over, Embodied AI Starts Accounting by Productivity." Previously investors would cut checks after seeing a cool demo video; now they want your ROI: in real scenarios (factories, warehouses, restaurants), how much human labor does one robot replace? What's the payback period? What are O&M costs? Can't answer those questions, and no matter how pretty the video, it's useless.

We covered Ant Lingbo's ¥1.5B independent funding in Issue 027 — Lingbo also builds embodied brains. Add today's two rounds, and the "brain camp" is becoming the absolute mainstream of embodied AI. That's actually a sign of industry maturity — when a sector moves from "showing off" to "doing the math," real players start entering the game.

Of course, the brain camp has its own problems. How do world models generalize to unfamiliar scenes? How do you close the sim-to-real gap? Where does data come from? None of these questions have perfect answers yet. But at least capital is pointing the right direction: making robots smart matters more than making them look good. After all, you buy a robot to do work, not to watch a show.

See you tomorrow.