AI needs a good "brain" to develop a body.

Article author and source: Semi-Industry Spectrum

Recently, Muxi Co., Ltd. and UBTECH signed a strategic cooperation agreement in Nanjing to jointly establish Xuanxuan Chuangzhi Technology (Wuxi) Co., Ltd., focusing on the research and development and mass production of embodied intelligence edge chips. Muxi is a domestic high-performance GPU company, while UBTECH is known as the “first listed humanoid robot company.” With Muxi providing the computational foundation and UBTECH offering humanoid robot platforms and real-world application scenarios, Xuanxuan Chuangzhi plans to tape out its chip in the second half of 2027 and achieve mass production by 2028.

Over the past few years, large models have driven computing demand to the cloud. Now, as AI begins to enter vehicles, robots, industrial equipment, and various mobile devices, embodied intelligence chips have become an urgent need.

The robots haven't made money yet, but the embodied AI chips are already driving fierce competition.

Currently, NVIDIA’s Jetson Thor, Qualcomm’s Dragonwing IQ10, Digu Robotics’旭日S600, Chiplet R1, Black Sesame Intelligence’s SesameX, Li Auto’s Mach M100, and XPeng’s Turing chip are all approaching embodied intelligence from different directions.

Embodied intelligence ≠ humanoid robots

First, it’s important to understand that embodied intelligence is not humanoid robotics.

Of course, humanoid robots grab the most attention—they have arms, legs, faces, can walk, carry boxes, and stand at the center of the exhibition. But the core of embodied intelligence isn’t about appearance; it’s whether the machine can complete a closed loop of “perception—judgment—action” in real-world environments. Therefore, cars can be a platform, AMRs can be a platform, industrial mobile platforms can be a platform, and so can inspection robots, service robots, and autonomous devices.

NVIDIA discusses Physical AI, covering robotics, autonomous vehicles, and industrial systems. Qualcomm's DragonWing IQ10 series (Qualcomm's most advanced robotics processor to date) is designed not only for full-sized humanoid robots but also for home robots, industrial AMRs, and edge intelligent devices.

Li Xiang, Chairman and CEO of Li Auto, refers to "embodied intelligent vehicles" not to turn Li Auto into a humanoid robotics company, but to redefine the car as an embodied intelligent terminal. Cars are equipped with sensors, computing power, actuators, energy systems, and operate in complex real-world road environments. They drive daily in the real world, entering scalable data loops earlier than many robots in laboratories.

Xpeng and Tesla are similar. Xpeng has integrated intelligent vehicles, Robotaxi, humanoid robots, and flying cars into its "Physical AI" framework. Tesla has transferred its FSD expertise to Optimus; Optimus is essentially a form of "embodied intelligence," fundamentally similar to FSD. FSD in vehicles is a "four-wheeled robot," while Optimus is a "humanoid robot."

Therefore, embodied intelligence chips cannot be narrowed down to just robot chips.

Humanoid robots are an entry point, but not the entire market. Embodied intelligence chips truly serve a new class of terminals: they must operate on-site, process sensors, perform inference, control actuators, and operate within constraints of power consumption, heat dissipation, cost, and security. This is far more complex than simply “equipping a robot with an AI chip.”

Embodied intelligence enters the "chip moment"

The biggest challenge for embodied intelligence remains the limitations on the edge.

The primary battlefield for AI chips is in the cloud, where insufficient computing power can be addressed by adding more racks, cooling systems, and power. But robots are different—battery capacity is limited, weight cannot exceed thresholds, cooling space is constrained, and costs have strict limits.

The current mainstream solution for humanoid robots is NVIDIA’s Jetson series, specifically the Jetson Thor and Orin edge AI chips. Taking Thor as an example, its power consumption can be configured between 40W and 130W—this level of power draw is very high when deployed on a robot.

Currently, the Tesla Optimus carries a 2.3 kWh battery, consuming about 100 W when idle and approximately 500 W when walking briskly; the Unitree H1 is equipped with an 864 Wh battery, with an official stated motion endurance of just over one hour. When fully powered, Thor’s brain alone could drain the H1’s entire battery within six hours—this is only the chip’s processing unit, not even accounting for motors or sensors. Therefore, there is an urgent need for embodied intelligence chips: they must not only solve “whether computation is possible,” but also “sustain computation within limited power constraints.”

The machine perceives its environment, assesses actions, executes them, and then adjusts based on feedback. This shifts the chip from pure AI inference to a heterogeneous system: the CPU must coordinate tasks, the GPU or NPU must run models, the ISP must process images, the MCU must handle real-time control, interfaces must connect to sensors, and communication must ensure determinism. At the level of robotic joints, dexterous hands, and chassis, this further involves EtherCAT, CAN, real-time buses, and motion control.

Chiplet's R1 "brain," D9 "cerebellum," and E3-R actuator represent this trend: the brain handles perception and planning, the cerebellum manages motion coordination and real-time control, and the underlying MCU controls joints and actuators.

The automotive industry is no stranger to this. Over the past few years, cars have evolved from distributed ECUs to domain control, and now to centralized computing. Chip manufacturers have learned multi-sensor fusion, real-time control, functional safety, long-term supply, and automotive-grade validation along the way. Now, with the rise of embodied intelligence, automotive-grade chipmakers are naturally turning their attention here.

This is why companies in the automotive industry chain, such as ChipSky, Black Sesame, Li Auto, and XPeng, are all entering the field of embodied intelligence—they are extending their accumulated expertise in automotive intelligence to broader physical intelligent devices.

Big players are entering—what are they competing for?

From the current embodied intelligence chip market, new entrants are exploring different directions. Although all are called embodied intelligence chips, some are high-performance modules, others are robotic reference designs, some stem from automotive-grade SoCs, and others are new projects jointly defined by original equipment manufacturers and chipmakers.

Let’s start with the iconic example: NVIDIA. NVIDIA’s chip designed for embodied intelligence is the Jetson Thor series module, targeted at Physical AI and robotics. It delivers up to 2070 FP4 TFLOPS of AI computing power and is equipped with 128 GB of memory, with power consumption configurable between 40W and 130W. Compared to the previous-generation Jetson AGX Orin, NVIDIA claims a 7.5x improvement in performance and a 3.5x improvement in energy efficiency. The Jetson Thor is positioned as a high-end computing platform for advanced robotics and complex edge AI workloads.

For robotics companies, the direct benefit of adopting NVIDIA is a mature development pathway: tools and platforms such as CUDA, TensorRT, Isaac, Isaac ROS, GR00T, and Cosmos already cover the entire robotics development lifecycle—including simulation, training, deployment, and inference. However, this solution has clear limitations: the power consumption range of 40W to 130W is not trivial for certain humanoid and lightweight service robots; the cost associated with 128GB of memory and high-end modules will also appear on the BOM at scale. As a result, NVIDIA currently serves more as a high-end industry benchmark—continuing to shape the embodied AI development ecosystem—but will not cover all price segments.

Qualcomm’s product offering is different. This year, Qualcomm officially launched the DragonWing IQ10 Robotics Reference Design at Computex 2026. This reference design delivers up to 700 TOPS of AI performance, featuring 18 Oryon CPU cores, a multi-core NPU, and a GPU, supporting up to 12 GMSL2 cameras and integrating sensors such as LiDAR, ToF, and IMU. For control interfaces, it supports PCIe, TSN, USB, CAN, EtherCAT, and CAN-FD, targeting industrial robots, AMRs, and humanoid robots.

Qualcomm is not just selling 700 TOPS—it’s selling a “reference design.” That means Qualcomm isn’t merely providing robot manufacturers with a single processor; instead, it’s delivering an integrated system that combines computing, sensor interfacing, motion control, network connectivity, and a software stack. For robot companies transitioning from prototypes to products, the first major hurdle is system integration: how to synchronize cameras, how to feed sensor data into the compute unit, how to ensure timing integrity in control loops, and how to enable overall device connectivity and OTA updates—tasks that consume significant development time. Qualcomm’s accumulated expertise in smartphones, XR, automotive, and IoT is perfectly suited to address these challenges in such endpoints. Currently, Qualcomm’s DragonWing IQ10 is not targeting the highest-end humanoid robots competing directly with NVIDIA’s top-tier offerings, but rather a broader range of robotic devices: AMRs, commercial service robots, industrial mobile platforms, and home robots. These applications may not require the highest computational power, but they do demand comprehensive interfaces, low power consumption, reliable connectivity, and stable supply.

In China, different companies interpret embodied intelligence chips differently.

Under the Horizon umbrella, Digua Robotics officially launched its flagship embodied intelligence high-performance platform, the Sunrise S600, in the first quarter of this year. With a powerful 560 TOPS (INT8) computing capacity and a 4× BPU Nash multi-core heterogeneous architecture, the Sunrise S600 enables efficient inference of diverse models—including language, vision, perception, and control—on a single chip, fully meeting the practical demands of large-scale robot production and deployment. Digua has also introduced an all-in-one development platform encompassing data闭环, training environments, and Agent development services. This product suite primarily serves as a foundational platform for robot development.

To date, the旭日S600 has achieved deep adaptation and optimization for flagship models including Qwen3, Qwen3-VL, Whisper-medium, YOLO26x, as well as multiple proprietary SOTA algorithms such as OmniOCC and VO-DP. Baidu’s advantages stem from its technical accumulation within the Horizon ecosystem. Over the past few years, intelligent driving companies have extensively deployed edge-side models, optimized toolchains, and adapted systems for mass production—experience that translates with continuity to robotics.

ChipXia Embodied Intelligence Product Roadmap

Chipletech's products focus more on low-level control. This year, Chipletech launched its full-stack embodied intelligence solution, covering the complete architecture of "brain-cerebellum-trunk-joints," including the R1 "brain" SoC, D9 "cerebellum" SoC, and E3-R series MCU for joints and trunk. The D9-Max features a multi-core heterogeneous architecture integrating CPU, GPU, NPU, DSP, and MCU, combining motion control, human-machine interaction, and EtherCAT master functionality on a single chip.

Zhang Xitong, General Manager of the MCU Product Line at ChipEast Technology, stated that embodied intelligence is a highly efficient and lucrative opportunity for ChipEast. “I believe there will definitely be a reuse rate of 60%-70%. The requirements for robotics and automotive applications are largely aligned, but there will still be some differentiated needs.”

ChipsX previously focused on intelligent cockpits, vehicle control, and MCUs, with deep expertise in automotive-grade chip reliability, functional safety, and long-term supply. Upon entering embodied intelligence, rather than competing for the highest AI compute power, the company began by targeting motion control and actuation endpoints. The R1 "brain" chip is still in development and will feature an ARM V9.2 architecture CPU and a high-performance NPU, designed for deploying edge-side models such as MLLMs and VLAs. If the R1, D9, and E3-R can eventually form a complete ecosystem, ChipsX aims not to create just a single robot AI chip, but to establish the foundational chip architecture within a robot’s electronic and electrical system.

Black Sesame Intelligence’s move also stems from the spillover of automotive chip technology. In 2025, Black Sesame Intelligence strategically expanded into embodied intelligence, launching the industry’s first SesameX Multidimensional Embodied Intelligence Computing Platform designed for commercial robot deployment. SesameX is a comprehensive, end-to-end computing platform “from edge modules to full-brain intelligence,” with full-stack proprietary development spanning hardware, software, toolchains, and model ecosystems. The three core computing modules—Kalos, Aura, and Liora—correspond to three tiers of robot development: vision-driven, sensorimotor coordination, and cognitive evolution, collectively forming a complete hierarchy of robotic intelligence. Aura delivers 70 TOPS of computing power; Liora approaches 600 TOPS.

Black Sesame’s advantage is not high-end humanoid robot “brain computing power.” It places greater emphasis on the commercial deployment of embodied intelligent devices—such as delivery robots, reception robots, inspection robots, cleaning robots, and educational robots in low-speed wheeled scenarios, as well as legged robots, inspection and maintenance robots (for industrial, power, and port applications), intelligent robotic arms, collaborative robotic arms, humanoid or teleoperated robots. These products demand computational power but prioritize interfaces, stability, cost, and mass production timelines. Drawing from its Huashan series of autonomous driving chips and Wudang series of cross-domain chips, Black Sesame has deep expertise in automotive perception, heterogeneous computing, and serving mass-production clients. The significance of SesameX lies in transferring automotive-era perception fusion, toolchains, and engineering experience to robotics. In 2025, Black Sesame Intelligent’s revenue from embodied intelligence solutions surged significantly year-over-year to RMB 96.3 million, indicating its entry into a phase of substantive commercial deployment.

Car manufacturers haven't been sitting idle either.

Ideally, embodied intelligence finds its ultimate application in automobiles. Li Auto defines an embodied intelligent vehicle as: an electric vehicle + a professional driver + an AI computer + a life assistant. The newly launched Mahe M100 chip, built on a 5nm automotive-grade process, delivers 1,280 TOPS of single-chip computing power. The autonomous driving system of the all-new Li Auto L9 Ultra is powered by Li Auto’s proprietary Mahe M100 chip; the L9 Livis features dual Mahe M100 chips, four lidar sensors, and integrates the MindVLA large model and 3D ViT Encoder. Additionally, building upon the Mahe M100, Li Auto has developed a complete embodied intelligent operating system—StarRing OS. Automobiles are currently the most mature embodied intelligent terminals, equipped with sensors, computing power, actuators, data feedback loops, and a clear business model. In embracing embodied intelligence, Li Auto has chosen the most prudent path.

The Xpeng Turing chip has a broader application scope. At the 2025 AI Day, Xpeng revealed that starting in 2026, select models will launch Robo versions equipped with four Turing AI chips, delivering a total computing power of 3,000 TOPS. Xpeng’s Robotaxi will also feature four in-house developed Turing chips, with plans to initiate trial operations in 2026. If viewed solely through the lens of automobiles, 3,000 TOPS is already high-end; but when including Robotaxi and humanoid robots, the chip becomes Xpeng’s AI foundation. Cars, Robotaxis, and humanoid robots will share certain models, sensor understanding, and edge inference capabilities, enabling R&D investment to be amortized across these platforms.

China's advantages, closed-loop scenario

For Chinese chip manufacturers, where are the opportunities in entering the embodied intelligence chip market?

It’s not about copying NVIDIA. Many want to beat NVIDIA, but they merely shout slogans and only create a “domestic Jetson,” which makes it hard to truly win.

Embodied intelligence differs from cloud-based AI chips. It is closer to real-world scenarios and closer to complete systems, and this is where China’s advantage lies. China has the world’s most active smart vehicle market, a complete robotics manufacturing supply chain, and high-density application scenarios such as factories, warehouses, logistics, inspection, security, and commercial services. These scenarios will, in turn, define the design of chips.

What interfaces does a warehouse robot require? How much power consumption can an inspection robot handle? What level of real-time control is needed for an industrial mobile platform? Which models should be deployed on the edge for an embodied intelligent vehicle? These questions cannot be answered in the abstract—they require placing the chip into the robot, the robot into the scenario, and using scenario feedback to refine the chip. This is the scenario闭环.

Cars may be the earliest mature example. Cars already have a clear business model, sensors, actuation systems, data feedback loops, supply chains, and mass production systems. Li Auto, XPeng, and Tesla view cars as embodied intelligence platforms. In contrast, humanoid robots are still seeking product-market fit.

The first wave of demand for embodied AI chips may not come from household humanoid robots, but rather from automobiles, Robotaxis, AMRs, industrial mobile robots, inspection systems, warehousing, and commercial services. This is precisely where Chinese manufacturers should focus their efforts. Don’t limit your goal to merely “replacing NVIDIA.” In many low- to mid-power, highly reliable, and application-specific terminals, customers don’t need the highest computational power—they need the most suitable system solution.

According to Omdia, global humanoid robot shipments in 2025 are projected to reach approximately 13,000 units. While this number is growing rapidly, from the perspective of the semiconductor industry, it remains an early-stage market. For a single chip, tens of thousands of units are not the end goal—they are just the beginning.