China's Beijing expos put robot bodies and AI brains on the same show floor

Two parallel Beijing expos this week showcase the physical humanoid stack alongside the foundation-model stack that runs it—the vertical integration is now the marketing message, not the roadmap.

In plain terms. Beijing's back-to-back robotics and AI expos this week frame the two halves of the Chinese embodied stack—hardware manufacturers and foundation-model vendors—as one industry, not two markets. Nikkei's framing is deliberate: robot bodies alongside AI brains, on the same show floor.

Context

The embodied-foundation-race thread has been the summer's most active. Xiaomi open-sourced its embodied foundation model ([#1813](ENCADRE titre="Xiaomi open-sources embodied foundation model" contenu="Xiaomi releases its open-source embodied AI model")). X Square's Wall-B beat Figure 03 on parcel-sorting throughput at 70% lower hardware cost ([#1909](ENCADRE titre="X Square's Wall-B outperforms Figure 03 in parcel sorting" contenu="X Square's Wall-B achieves higher parcel-sorting throughput than Figure 03 at 70% lower hardware cost")). Unitree priced its IPO at $9B with DeepSeek and Tencent backing ([#1841](ENCADRE titre="Unitree prices $9B IPO with DeepSeek and Tencent backing" contenu="Unitree secures $9B valuation in IPO with investments from DeepSeek and Tencent")). Alibaba and JD.com's robot maker filed for HK IPO ([#1981](ENCADRE titre="Alibaba and JD.com's robotics unit files for HK IPO" contenu="Alibaba and JD.com's robotics manufacturing subsidiary submits IPO application in Hong Kong")). Xiaomi rolled out a dedicated robotics AI model ([#21706146](ENCADRE titre="Xiaomi launches dedicated robotics AI model" contenu="Xiaomi introduces a new AI model specifically designed for robotics applications")). LG announced humanoid training on Nvidia software ([#40379698](ENCADRE titre="LG to train humanoids using Nvidia software" contenu="LG partners with Nvidia to train humanoid robots using its software")). The stack is now polymarket, polycapital, and cross-border.

What the article says

Per Nikkei Asia, this week's back-to-back Beijing expos showcase "robot bodies and AI brains" as the co-billed subject of the event. The framing itself is the story: Beijing is presenting the physical and foundation-model layers as a unified industry rather than as separate categories.

Analysis

The Chinese playbook here mirrors the 2010s electric-vehicle playbook: subsidise the vertical (batteries + motors + software + assembly) rather than the champion. That produced BYD and CATL. It produced 200 also-rans that mostly failed. Applied to robotics, the same dispersion is the feature: multiple humanoid makers competing in an integrated foundation-model supply market is what drives the cost curve down. X Square's 70% cost undercut on Figure 03 ([#1909](ENCADRE titre="X Square's Wall-B outperforms Figure 03 in parcel sorting" contenu="X Square's Wall-B achieves higher parcel-sorting throughput than Figure 03 at 70% lower hardware cost")) is what that dispersion looks like at unit level.

Scenarios

• Base case (55%): 3-5 Chinese humanoid makers scale to > 10,000 units/year by 2028, with domestic industrial deployments dominating.
• Consolidation case (25%): Unitree or Alibaba-backed player absorbs two mid-tier makers; the market repriced upward.
• Export case (20%): Chinese humanoids enter Southeast Asian factory retrofit contracts before US ones—export-control-adjusted route.

Risks

Physical AI has a demo-vs-deployment gap that shows up brutally at scale (safety, uptime, MTBF). The Beijing showcase is optimised for the demo. Watch the 12-month deployment reports, not the show-floor throughput.

So what

If you're a Western integrator, don't wait for Optimus. The unit economics of the Chinese stack are already crossing the "cheaper than automation contractor" line for pick-and-place and simple assembly. The right question is whether you can qualify a Chinese humanoid on your line before your local regulator decides to make that hard.