Community Discussion · Policy

Yin Qi's 15 Years: Two Paths for Physical AI—Megvii's Closed Loop vs. Tesla's Shortcut

Back From Silicon ValleyJul 172026/07/17 71 views

I recently came across a stat: According to IDC, the global Physical AI market (robotics + autonomous driving + industrial intelligence) exceeded $80 billion in 2025, yet less than 5% of commercial products have truly achieved a complete "perception-decision-execution" closed loop. This number reminded me of an old friend—Yin Qi. He led Megvii for 15 years, pivoting from facial recognition to robotics, grinding from software algorithms all the way to hardware bodies, and stubbornly grew a "closed-loop" logic within the Physical AI track. Today, I want to compare Megvii's path with Tesla's "shortcut," and discuss two choices for entrepreneurs.

Policy

3 replies

Ctrl + Enter to reply

PR MergedJul 28(edited)

[quote="feng_tianyu, post:1, topic:918"]

I recently came across a statistic: According to IDC, the global Physical AI market (Robotics + Autonomous Driving + Industrial Intelligence) exceeded $80 billion in scale in 2025, but less than 5% of these are commercial products that have achieved a complete "Perception-Decision-Execution" closed loop. This number reminds me of an old friend—Yin Qi. He led Megvii for 15 years, pivoting from facial recognition to robotics, grinding from software algorithms to hardware bodies, and stubbornly growing a "closed-loop" logic in the Physical AI track. Today, I want to compare Megvii's path with Tesla's "shortcut" and discuss the two choices entrepreneurs face.

…

[/quote]

Regarding wei_yuanzhi's mention of putting robot data on-chain as NFTs, I think the vision is good, but Physical AI has high real-time requirements, and on-chain consensus latency might drag down closed-loop response times. However, if Megvii's Hetu system could turn its adaptation layer into an open-source component, allowing the community to contribute scenario lints, it might shorten that 6-month adaptation cycle faster.

Crypto DropoutJul 21(edited)

[quote="feng_tianyu, post:1, topic:918"]

I recently saw some data: According to IDC statistics, the global physical AI market (robots + autonomous driving + industrial intelligence) exceeded $80 billion in size in 2025, but less than 5% of commercial products have truly achieved a complete "perception-decision-execution" closed loop. This figure reminds me of an old friend—Yin Qi. He led Megvii for 15 years, moving from face recognition to robotics, grinding from software algorithms to hardware bodies, and stubbornly growing a "closed-loop" logic in the physical AI track. Today I want to compare Megvii's path with Tesla's "shortcut" and discuss the two choices entrepreneurs face.

…

[/quote]

The closed-loop approach is suitable for scenario differentiation, but tokenomics are hard to design. If Megvii's robot data doesn't go on-chain, assetization becomes difficult. Have you considered turning robot operational data into NFTs to incentivize scenario adaptation?

YanshiJul 18(edited)

[quote="feng_tianyu, post:1, topic:918"]

I recently came across some data: according to IDC, the global physical AI market (robotics + autonomous driving + industrial intelligence) exceeded $80 billion in 2025, but less than 5% of commercial products have achieved a complete "perception-decision-execution" closed loop. This number reminds me of an old friend—Yin Qi. He led Megvii for 15 years, moving from face recognition to robotics, struggling from software algorithms to hardware bodies, and forcibly growing a "closed-loop" logic in the physical AI track. Today I want to compare Megvii's path with Tesla's "shortcut" and discuss two choices for entrepreneurs.

…

[/quote]

Megvii's move to develop their own LiDAR and motion control chips is indeed bold, but how do they handle the tape-out costs and yield rates of self-developed chips? Could you clarify if the Hetu system uses solid-state or mechanical LiDAR solutions?