AMD Acquires World Labs, Qualcomm PickNik: NVIDIA’s Hugging Face Acquisition Triggers Renewed Physical AI Arms Race

In early September, NVIDIA acquired Hugging Face for approximately US$13 billion. Hugging Face supports more than 18 million developers and hosts 3 million models, alongside LeRobot, a collection of hardware- and Vision-Language-Action (VLA)-agnostic models and tools for robotics development. Spanning low-cost industrial arms to humanoids, LeRobot has become a de facto Physical Artificial Intelligence (AI) development and Research and Development (R&D) standard for emerging robotics.

On September 23 2026, Qualcomm agreed to acquire PickNik Robotics. PickNik’s core product, MoveIt, originated at Willow Garage, the renowned robotics incubator that gave rise to the Robot Operating System (ROS), The Open Robotics Middleware Framework (Open-RMF), and several other foundational robotics projects. The open-source MoveIt and commercially licensed MoveIt Pro primarily provide seamless motion control and path planning for robot controllers, alongside user-defined interfaces. Acting as a motion control layer, the software calculates forward and inverse kinematics for pick-and-place tasks, avoids collisions, and implements safety zones. While MoveIt Pro can support end-to-end pick-and-place tasks, PickNik’s offerings also allow third-party Physical AI models, including VLAs and Vision Language Models (VLMs), to be integrated above the motion controller. This is the draw for Qualcomm: remaining model-agnostic while possessing an optimizable framework that allows models to be deployed across various robot controllers. Qualcomm can introduce its chipsets either within the robot controller itself or as part of a separate “AI accelerator” kit.

AMD’s September 28 announcement that it would acquire World Labs for approximately US$8.2 billion indicates a greater ambition: to develop frontier models directly and compete more closely with NVIDIA and its Cosmos world-model platform. The acquisition builds on a relationship that began in 2025, when AMD and World Labs established a technical partnership focused on model training and inference optimization on AMD Graphics Processing Units (GPUs); AMD also invested in the company during its Series B. World models are increasingly viewed as a critical technology for scaling VLA models, particularly by generating the synthetic environments and training data needed to overcome the shortage of real-world robotics data. Their relevance extends beyond humanoids to Autonomous Mobile Robots (AMRs), drones, and other robotic systems requiring an understanding of and ability to interact with the physical world. By acquiring World Labs, AMD gains not only a world-model portfolio, but direct model-level insight that can be used to shape its future hardware, software, and Physical AI infrastructure.

Open-source technologies represent the primary opportunity for incumbent semiconductor vendors, including Intel, Qualcomm, and AMD, to catch up with NVIDIA in the Physical AI space. Physical AI development tools are currently open to developers because the technology is not commercially mature. However, this has created several different “open” development tools to choose from, each of which will become optimized for the semiconductor hardware of the silicon vendor funding it. This forces hardware lock-in and creates a fragmented landscape in which semiconductor vendors must provide funding both to maintain repositories and to support frontier laboratories using their stacks, hoping that the next algorithmic breakthrough, which could come from anywhere, happens on their hardware. The Physical AI arms race now encompasses the hardware and the full software stack, which is no longer truly open or semiconductor-agnostic.

Meanwhile, stakeholders continue to be rewarded for betting on long-horizon technologies. Companies shipping humanoid robots or offering commercially available Physical AI technologies often attract lower valuations than those pursuing a general-purpose humanoid future. At the same time, the extensive partnerships between semiconductor vendors and industrial robot Original Equipment Manufacturers (OEMs)—including Universal Robots, FANUC, KUKA, and Yaskawa—observed last year have quietened. They have been replaced by increased Physical AI rhetoric, but little explicit adoption of high-Tera Operations per Second (TOPS) Systems-on-Chip (SoCs).

The launch of Universal Robots’ seventh-generation Collaborative Robots (cobots) and controllers indicates that OEMs may be moving away from additional, SoC-dependent AI augmentation. Marketed as “AI ready,” the new controller allows partners, in some instances, to deploy vision guidance and other AI workloads natively, without intermediate SoC-backed accelerators. If this reflects the direction of the broader industrial and collaborative robotics markets, high-performance SoCs could be pushed away from established industrial applications and toward longer-horizon technologies such as humanoids, where compute-intensive world models may become a key enabler. These market signals indicate a further widening of the gulf between today’s manufacturing technologies and the compute-intensive strategies semiconductor vendors are currently pursuing.

Semiconductor vendors entering Physical AI must recognize that NVIDIA’s advantage increasingly extends beyond compute into the open-source models, datasets, and tools used by robotics developers. Competing vendors cannot rely solely on model-agnostic positioning if the principal distribution platform is owned by their largest competitor. Instead, vendors must reduce the technical and commercial risks associated with adopting alternative hardware. This requires investment in software tooling, validated integrations, and deployment partnerships alongside differentiated silicon. The objective should not be to replicate NVIDIA’s entire ecosystem, but to ensure that robot OEMs and developers can deploy emerging Physical AI models on alternative architectures without additional cost, complexity, or support requirements.

• Expect Further Acquisitions, Partnerships, and Mergers: Leading semiconductor vendors will continue to annex the models, algorithms, and Intellectual Property (IP) needed to compete across the full Physical AI stack. Those without world models will need to acquire them, partner with frontier laboratories, or provide hardware-specific optimization to remain in the race.
• Agnosticism Moves Elsewhere: Secondary semiconductor vendors supplying Microcontrollers (MCUs), Application-Specific Integrated Circuits (ASICs), and Microprocessor Units (MPUs) must remain open to interfacing with all major SoCs and compute platforms. NXP Semiconductors, Texas Instruments, Renesas Electronics, STMicroelectronics, and Infineon Technologies can differentiate by providing interoperable control, sensing, networking, and safety layers across increasingly fragmented Physical AI ecosystems.
• Watch for Further Broadsiding: Chinese semiconductor vendor D-Robotics raised US$400 million to expand its Sunrise robotics chips and software platform, targeting embodied AI applications ranging from industrial robots to humanoids. Given China’s strength in robotics manufacturing, competing semiconductor vendors should monitor Chinese-native suppliers and pursue partnerships with Western OEMs targeting defense and strategically supported manufacturing initiatives in markets with an explicit requirement to exclude China from their value chains.

Ultimately, scalable general-purpose robotics—the fruit of Physical AI—remains out of reach today. Whether broad generalizability emerges from new algorithmic approaches, such as World Action Models superseding VLA models or the next form of mathematical optimization; distributed architectures leveraging cloud-based compute; improved force detection and tactile sensing; or better imaging solutions remains unclear. The solution will likely combine several of these advances, developed by a disparate set of innovators across the Physical AI stack. At this stage of the technology development cycle, all players should hedge their bets.