Soaring to $14 billion in 3 years! SoftBank and NVIDIA compete to invest in the $100 billion valuation "unicorn"

The fastest unicorn valued at over 100 billion in the AI circle by 2026. Skild AI completes Series C funding with a valuation exceeding $14 billion, just 3 years after its founding, with investments from SoftBank, NVIDIA, and Bezos. The founders come from Meta AI, focusing not on robot hardware but on creating a “general brain” called Skild Brain, which uses shared models to endow robots with physical world capabilities.

Capital Logic Behind a Tenfold Valuation Surge in Two Years

Skild AI is the fastest unicorn to surpass $100 billion this year. Its rapid development pace and explosive valuation growth are even considered phenomenal amid the fervent AI investment wave. The company was founded just two months ago and secured a $14.5 million seed round led by Lightspeed Venture Partners, marking a successful start. On its one-year anniversary, Skild AI completed a $300 million Series A funding, with post-money valuation soaring to $1.5 billion. In less than two years, its valuation increased nearly tenfold, with the latest Series C funding pushing it to over $14 billion.

The reason for capital backing Skild AI is simple—there is a severe gap in the global labor market. The manufacturing sector in the US alone is expected to face a shortage of 2.1 million jobs by 2030, and aging populations in Europe, Japan, and other developed economies make the situation even more critical. General-purpose robots capable of executing complex physical tasks are seen as key to solving productivity crises. However, the current robot industry is highly fragmented, with each manufacturer attempting to develop everything from mechanical structures to control systems independently, resulting in high R&D costs and limited cross-platform capabilities.

Skild AI’s “brain-only, no body” pure software model aligns perfectly with the industry’s shift from “hardware-oriented” to “AI model and software capability-oriented” trends. Sequoia Capital’s post-investment report states: “The core value of Skild AI is unlocking robots’ ‘emergent abilities’ in the physical world through shared foundational models. This is fundamentally different from the past ‘single-point controller’ approach, which lacked scalability.” This “emergent ability” is similar to GPT-3’s breakthrough in language understanding—when the model reaches a critical size, it automatically acquires capabilities not explicitly taught during training.

SoftBank’s involvement is particularly strategic. The Japanese tech giant acquired ARM for $32 billion in 2016 and has been active in the robotics field (such as acquiring Boston Dynamics). SoftBank Vision Fund is known for aggressive AI investments, and its heavy stake in Skild AI indicates it believes “general robot brains” are the next trillion-dollar market. NVIDIA’s participation provides computing power and ecosystem support, with Skild AI’s training infrastructure likely built on NVIDIA’s GPU clusters. Bezos’ personal investment is even more rare; the Amazon founder rarely participates in early-stage projects, and his endorsement adds intangible brand value to Skild AI.

The Technical DNA of a Team Coming from Meta

The robustness of Skild AI’s technological capabilities lies in the background of its founding team. Deepak Pathak, before founding Skild AI, was a renowned scholar and practitioner in AI and robotics, having served as an assistant professor at Carnegie Mellon University and published numerous widely cited papers. During his tenure at Meta AI, he was deeply involved in key projects related to adaptive learning, sim-to-real transfer, and large-scale robot data training.

Pathak firmly believes that true general AI must be built through interaction and trial-and-error in the physical world, rather than relying solely on digital text or image data. This philosophy faced skepticism within Meta, where the focus was on the metaverse and social AI, with relatively conservative investments in physical robots. This strategic disagreement ultimately led Pathak to start his own venture and realize his vision.

Abhinav Gupta, also from Meta AI, has made significant contributions in computer vision and robotics learning. He emphasizes learning physical commonsense from large-scale online video data, enabling robots to understand object properties, physical laws, and human intentions. In fact, Gupta and Pathak had multiple collaborations at Meta, exploring how to replicate the “emergent abilities” of large language models in physical robots.

Both believe that the current robot industry overly depends on task-specific, hardware-specific customized solutions, lacking a “general brain” that can be generalized and scaled. This severely limits robots’ application potential and adoption speed in the real world. Therefore, in the later stages at Meta, they began incubating a project to build a hardware-agnostic foundational robot model. In early 2023, Pathak and Gupta decided to leave Meta and pursue full-time entrepreneurship. They are convinced that the future of the robot industry lies not in creating more “bodies,” but in providing a powerful, shareable “brain.”

GPT-3 Moment in the Robot Industry

Rewinding to 2023, intelligent robots were everywhere, but each robot required training dedicated algorithms from scratch, leading to long R&D cycles, high costs, and inability to share capabilities across different robots. A persistent challenge in embodied intelligence industry is: how to achieve good generalization? Generalization here means that abilities learned on one robot can be quickly transferred to another.

This is extremely difficult because: the physical world is highly complex, uncertain, and dynamic, and robots need to generalize across perception, decision-making, and execution. For example, variations in lighting, weather, cluttered backgrounds, and occlusions can cause dramatic changes in visual sensor data distribution. Even if simple tasks (like grasping or walking) are learned, combining them into complex tasks (such as “open the fridge, take out a drink, and pour into a cup”) exponentially increases the decision space.

Three Major Technical Paths of Skild Brain to Solve Generalization

Large-scale Multimodal Pretraining: Learning physical commonsense from online videos, simulation environments, and real robot data to establish universal representations across scenarios

Hardware-Agnostic Architecture: Using abstraction layers to decouple perception and decision logic from specific mechanical structures, enabling the same model to be deployed on wheeled, legged, or arm-type robots

Continuous Learning Mechanism: Robots generate data during task execution, which is sent back to the cloud for ongoing model optimization, allowing experiences from all robots to benefit the entire network

Skild AI does not manufacture robot hardware but aims to install a “general brain” in all robots. Its founders claim they are creating a “GPT-3 moment” for embodied intelligence. Skild Brain can separate software from hardware, avoiding being constrained by a single hardware design. At the same time, it greatly lowers industry barriers, allowing other robot manufacturers or integrators to focus on hardware optimization and scene deployment, simply calling Skild Brain’s API to access advanced intelligence, greatly accelerating robot application adoption.

The commercial prospects are equally promising. In industrial and commercial sectors, robots on production lines won’t need to shut down for minor faults; in disaster rescue, robots with “limbs and parts” can still perform tasks; in consumer markets, a single “brain” can be “re-skinned” for different uses, significantly reducing costs. These technological foundations are reshaping perceptions of AGI—relying solely on digital knowledge cannot build true AGI; machine agents must learn through “practice” and understand real-world operation laws physically.