[AI influence. on the CLARiiON serverTaipeimessage]NVIDIACosmos 3 was officially released at the GTC Conference in Taipei on June 4. It adopts a dual Transformer architecture and is specially developed for robots, autonomous driving and visual agents. It supports collaborative generation of multimodal content and promotes the transformation of agents from perception to embodied cognition. This is the world's first open source embodied intelligence basic model, marking Nvidia's strategic transition from an "AI chip supplier" to a "embodied intelligent platform operator."
"Dual Transformer Architecture": Separation and Collaboration of World Models and Action Models
Cosmos 3's technological innovation lies in splitting the traditional end-to-end model into two transformers that work together:
World Model: Understand the physical environment, predict object motion, and simulate interaction results
Action Model: Generating specific motion control instructions based on predictions from the world model
This separation architecture allowsCosmos 3 can simultaneously process high-dimensional sensory inputs and low-latency action outputs, solving the "perception-action" coupling problem of traditional end-to-end models in complex physical environments.
Deeply bound with Yushu Technology's "chip + complete machine"
The release of Cosmos 3 and the cooperation between Nvidia-Yushu Technology form a strategic closed loop. The H2 Plus humanoid robot jointly launched by the two parties adopts the "Yushu hardware body + Nvidia AI brain" architecture, but the industry is worried that this model may lead to the "loss of technical sovereignty" of China's embodied intelligent enterprises-Yushu is responsible for hardware manufacturing, Nvidia monopolizes AI algorithms and control, and data security risks rise accordingly.
Yushu technologyIn 2025, revenue will be 1.708 billion yuan (a year-on-year increase of 335%), shipments of humanoid robots will exceed 5500 units, and the IPO issuance valuation will be approximately 42 billion yuan. Its binding with Nvidia allows China's tailored intelligent leaders to obtain a top computing power platform while also facing the structural risk of relying on foreign investment for core algorithms.
The release of NVIDIA Cosmos 3 is an "Android moment" in the field of specific intelligence-through open source basic model + chip ecosystem locking, NVIDIA is copying Google's platform strategy in the mobile Internet era. But unlike Android, Cosmos 3's "open source" is conditional: the core training framework is open source, but the most optimized reasoning version still relies on Nvidia GPUs and the Jetson platform. This strategy of "opening the core and locking the edge" allows China manufacturers to gain technological convenience while also falling into dual dependence on "chip-algorithm". When U.S. export control policies tighten, this reliance may escalate from a "commercial risk" to an "survival risk."