I. The Showdown Between Vertically Integrated Closed Loops and Open Standard Alliances
The rise of Physical AI and Embodied Intelligence marks the evolution of semiconductor demand from mere "general-purpose computing" to a "perception-reasoning-control closed loop." Unlike traditional data centers, embodied intelligent agents require chips to provide highly deterministic physical feedback while processing massive amounts of multimodal data.
In this paradigm shift, the strategic philosophies of NVIDIA and AMD exhibit a fundamental divergence:
NVIDIA's "Capital-Vertical Integration" Moat: NVIDIA not only locks in developers by building a closed-loop ecosystem from DGX (training) and Omniverse (simulation) to Jetson (execution), but also implements a "Capital-as-a-Moat" strategy. In fiscal year 2026 alone, NVIDIA invested up to $17.5 billion in embodied intelligence startups, deeply binding the supply chain through equity. Its core logic is to treat Physical AI as the infrastructure for "Sovereign AI," committed to building irreplaceable "AI factories."
AMD's "Open Adaptive" Alliance: AMD leads the UALink and Ultra Ethernet Consortium (UEC) open standards, uniting giants like Broadcom and Microsoft to counter NVIDIA's closed system. Its strategic focus lies in reducing the "vendor lock-in" risk for industrial customers through the universality of the x86 architecture and the flexibility of FPGAs, while providing industrial-grade reliability with a longer lifecycle (15-year supply commitment).
These two philosophies directly determine the initial selection logic of developers: startups prefer NVIDIA's "one-stop" efficiency to shorten time-to-market; whereas traditional industrial giants pursuing sovereign security, deterministic control, and supply chain autonomy view AMD's open platform as a strategic safe haven.
II. Core Hardware Architecture: Computational Divergence of the Edge "Brain" and "Spinal Cord"
In biological evolution, when we attempt to catch a bouncing ball, the body employs a dual logic:
"Brain" (High-level Decision/Perception): Responsible for processing multimodal perception (vision, touch, radar). It utilizes large models for semantic understanding and task planning. This belongs to "non-deterministic" computing, pursuing computing throughput (TFLOPS).
"Spinal Cord" (Low-level Control/Feedback): Responsible for motor servo and sensor synchronization. When the ball touches the fingertip, the spinal cord must complete the command closed loop at the microsecond level without going through the brain's thinking. This belongs to "hard real-time" computing, pursuing determinism and extremely low jitter.
Merely computing fast is not enough. If there is latency in the low-level control, even if the brain calculates a perfect obstacle avoidance path, the robot will still fall in the physical world due to command jitter caused by "cerebellar atrophy."
The core hardware challenge of embodied intelligence lies in balancing the "breadth of perception" with the "precision of control."
2.1 In-Depth Architecture Comparison: Perceptual Hegemony vs. Deterministic Execution
NVIDIA Jetson AGX Thor (Blackwell Architecture): This chip integrates a 14-core Arm Neoverse-V3AE enterprise-grade CPU, with the core highlight being its mixed-precision Transformer engine (MVFP4). Under FP4 sparse precision, its AI computing power reaches up to 2,070 TFLOPS. Despite the astonishing perceptual computing power, limited by the Arm architecture's constraints in ultra-high-frequency serial timing logic processing, when facing ultra-high-frequency trajectory replanning, its control loop frame drop rate averages 1.6 times/second.
AMD Ryzen AI Embedded X100 Series: Adopts an integrated "brain, spinal cord, and body" approach. It integrates a Zen 5 CPU (multi-threaded throughput 1.5 times higher than competitors), an RDNA 3.5 GPU, and an XDNA 2 NPU (50 TOPS). Its unique advantage lies in processing sensor physical alignment through FPGA fabric (programmable logic). Based on the high single-thread throughput of x86 cores, its control loop frame drop rate is only 0.45 times/second, and it retains 82% of idle bandwidth for user logic, supporting a 125μs hard real-time control loop.
The table below provides an in-depth comparison of the core hardware specifications of NVIDIA and AMD, ranging from large-scale data center AI accelerators to edge robotic computing platforms.
2.2 Quantitative Comparison of Core Performance Parameters
NVIDIA possesses absolute hegemony in multimodal visual perception (the brain) with its MVFP4 engine, making it the preferred choice for humanoid robots processing complex environments; whereas AMD's physical advantages in deterministic servo control (the spinal cord) make it irreplaceable in high-precision industrial assembly and extreme operating conditions such as explosion-proof and impact-resistant scenarios (e.g., Foundation Robotics).
III. Software Stack and Toolchains: The CUDA Moat and the AI Programming Revolution of ROCm
The software ecosystem builds the highest migration costs through "developer inertia."
3.1 NVIDIA: The End-to-End "Three Computers" Gravity Field
NVIDIA builds a closed-loop gravity field through Omniverse (physical simulation), Isaac Sim (robot training), and Project GR00T (foundation models).
Technical Path: Focuses on "Domain Randomization," utilizing the ovphysx library and the Cosmos world model to generate massive biased training data in simulation, narrowing the Sim-to-Real gap.
Ecosystem Lock-in: The dividend of 7.5 million CUDA developers ensures that all the latest embodied large models prioritize adaptation to the NVIDIA architecture.
3.2 AMD: The AI Agent-Empowered Decentralized Path
AMD is leveraging AI to change the game and make up for the shortcomings of the software ecosystem:
· AMD-Anthropic Strategic Alliance: AMD invested $5 billion in Anthropic, with the most core "black swan" move being: Anthropic utilizes the Claude model to reconstruct the ROCm compiler pipeline. Through AI-automated optimization of operators and instruction allocation, AMD attempts to eliminate CUDA's leading advantage at an exponential rate.
· ROCm 7.x & AMD Skills: Provides a 70%-80% automatic code conversion rate (HIPIFY). AMD Skills allows programming agents like Claude/Cursor to natively understand the AMD hardware architecture, significantly reducing R&D labor costs for CTOs.
3.3 Comparison of "Sim-to-Real" Migration Paths
The maturity of the software ecosystem determines the strength of the computing platform's "gravity field."
Comparison of Sim-to-Real Technical Paths
Algorithmic Path (NVIDIA): Trains the model's robustness to physical deviations through massive synthetic data generated by large-scale parallel simulation.
Hardware Path (AMD): Emphasizes "hardware-level physical alignment," ensuring that the robot's action rhythm in reality is mathematically identical to the simulation model through ultimate deterministic control.
IV. Partners and Industry Moats: Imperial Systems and Ecological Alliances
According to the latest market analysis, computing chip manufacturers are establishing deep moats by binding with industry giants.
NVIDIA (Industry Lighthouse Camp): Holds approximately 50% share on the edge "brain" side of embodied intelligence. Partners include Figure AI, Agility Robotics, and Boston Dynamics. It holds an 80% OEM share in the autonomous driving field and tightly locks cloud simulation and terminal deployment through proprietary NVLink interconnects (1.8 TB/s).
AMD (Industrial Adaptive Camp): Holds approximately 25-30% share in the edge "spinal cord/joint control" field. It has absolute advantages in industrial automation (SICK, ADI), defense, and extreme operating condition robotics (Foundation Robotics). AMD, together with Broadcom, Meta, and others, promotes the UALink open standard, attempting to break NVIDIA's memory wall monopoly.
4.1 Value Chain Distribution
· NVIDIA (Frontier Benchmark): Deeply binds top humanoid robots (Figure AI, Agility, Boston Dynamics) and autonomous driving giants (Mercedes-Benz, BYD). It builds a "chip-as-a-system" monopoly through proprietary NVLink/NVSwitch protocols, suitable for research laboratories pursuing ultimate performance.
· AMD (Industrial Hardcore): Penetrates into the industrial "spinal cord" segment. Partners include perception giants (SICK, ADI) and defense pioneer Foundation Robotics. The latter's Phantom series robots utilize AMD solutions to achieve physical performance of "IP67 waterproof, 100G impact resistance," proving AMD's robustness in extreme operating conditions.
4.2 Standard War: UALink vs. NVLink
NVIDIA's moat is facing challenges from the Ultra Ethernet Consortium (UEC). AMD, together with Broadcom, Microsoft, and Intel, promotes UALink over Ethernet, aiming to eliminate InfiniBand's latency advantage and achieve free networking compatible with multiple vendors.
V. China Market Landscape and Autonomous Confrontation
Export controls are forcing China's embodied intelligence industry into a period of "underlying reconstruction."
5.1 Competitive Landscape and "Volume Cap" Restrictions
The 2026 export rules introduce a "Volume Cap": Chips with TPP (Total Processing Performance) below 21,000 or bandwidth below 6,500 GB/s (such as compliant versions of H200/MI325X) are allowed for export, but the quantity must not exceed 50% of the sales volume to domestic U.S. customers.
Share Changes: NVIDIA's accelerator share in China shrinks to 47%, while Huawei Ascend captures 35.5% of the market with its Atlas 950 SuperPod system (1 EFLOPS FP8 cluster performance).
5.2 Key Manufacturer Comparison: Leaders in Autonomous Driving and Embodied Intelligence
Huawei "Qiankun" (ADS 3.0): Relying on the Ascend 910C/950 series, Huawei builds a closed-loop empire. Its vehicle-mounted MDC system achieves a 15.2% market share in NOA in 2025.
Horizon Robotics "Journey 6": As the dominant player in the pre-installed autonomous driving market, Horizon Robotics holds up to a 28% market share in China's L2/L2+ segment. Its Journey 6P single chip provides 560 TOPS computing power, specifically designed for end-to-end Transformer hardening.
5.3 The "Disintermediation" Threat from Autonomous Driving OEMs
The storm of in-house chip development by OEMs such as NIO, Li Auto, and XPeng has reached a large scale. Li Auto's Mach M100 chip achieves an 82% hardware utilization rate through instruction hardening for BEV/Occupancy algorithms. This "vertical integration" trend is squeezing U.S. giants toward the transformation crossroads of "Software as a Service" (SaaS).
VI. Strategic Selection Strategies
The second-half competition in embodied intelligence lies in the convergence of "computing power redundancy" and "control determinism."
1. Frontier Humanoid Robot Startups: It is recommended to choose the NVIDIA path first. Fully utilize the Cosmos world model and Isaac Sim simulation pipeline to quickly break through the generalization algorithm barriers from text to action.
2. Industrial Heavy-Duty and Safety-Sensitive Manufacturers: It is recommended to choose the AMD path first. Utilize the ASIL-D certification of Versal Gen 2, high-throughput x86 CPUs, and sub-microsecond clock jitter control of FPGAs to ensure robustness in extreme operating conditions.
3. Domestic Chinese OEMs: It is recommended to adopt a "domestic + compliant" dual-track system. While benchmarking against advanced international VLA models, deeply integrate into the Huawei or Horizon Robotics ecosystem, and accelerate the development of customized operator chips to maximize PPC utilization.
The winning hand for future embodied intelligent computing platforms will not lie in the mere stacking of TFLOPS, but in who can perfectly map the "infinite simulation" in the cloud to the "absolute physical determinism" at the edge through the deep convergence of heterogeneous architectures.