Bezos's 2025 AI startup, Prometheus, has completed a $12 billion Series B round, reaching a $41 billion valuation.Author and source: Yuan Chuan Research Institute
In the tech landscape of 2026, NVIDIA and OpenAI remain at the top, while SK Hynix and Samsung rise strongly. Amid a flurry of established and emerging giants, the most intriguing player is Bezos.
After stepping down as Amazon’s CEO in 2021, Bezos stayed anything but idle: he traveled to space using Blue Origin’s space technology, invested in Altos Labs to explore longevity science, and co-invested in Figure AI to bet on humanoid robots.
After going through the Rich Man Game, Bezos’s only feeling is probably that the development cycle for complex physical products is simply too long.
Space industry, biomedicine, and embodied intelligence are all top-tier fields of advanced manufacturing. Although these sectors may seem distant from one another, they all point to a common long-term challenge: as the complexity of physical products grows to involve tens of thousands of components, the difficulty of moving from research and development to validation and mass production increases exponentially. For example, the Boeing 777X aircraft alone has a staggering 300,000 parts in its jet engines.
Increasing the thrust of a jet engine by 10% over current levels, due to the inherent complexity of engineering, could traditionally take a decade. Bezos aims to use AI to accelerate this cycle from concept to manufacturing by tenfold or more [1].
In November 2025, Bezos returned to entrepreneurship with the founding of Prometheus, marking his first time officially serving as CEO of a company since stepping down as Amazon’s CEO.
Just seven months later, Prometheus completed its latest Series B funding round, raising an impressive $12 billion—the largest Series B round ever for an AI startup—and reached a valuation of $41 billion.
Such a money-burning venture, yet capital is so eager to bet on it—what makes Prometheus so compelling?
What Bezos is handing to the market is far more than just a product—it’s a ticket to reshape the global industrial system.
Prometheus describes what he is doing as building a "future CAD (computer-aided design) system," where CAD is software engineers use to model anything before manufacturing it—but to Bezos, this still doesn't fully capture it [2].
It is closer to an engineering foundation—the closest real-world realization of J.A.R.V.I.S. from Iron Man—building an AI-powered tool system designed to assist engineers in designing and manufacturing physical products[2], thereby reimagining engineering and manufacturing through AI.
Prometheus's technological foundation lies in the physical world, which is tied to a broader industry theme: Physical AI.
The world is 3D.
At GTC Paris last June, Jensen Huang calculated the massive opportunity for physical AI: behind factories, logistics, and humanoid robots lies a $50 trillion industry table [3].
What does 50 trillion dollars mean?
According to Bloomberg’s latest forecast, the market size for broad generative AI, encompassing computing power, software, and related services, will reach no more than $2 trillion by 2030 [4]. Huang’s estimate for physical AI clearly represents a transformation of the real economy, encompassing industries such as manufacturing, logistics, and energy—sectors with market sizes in the hundreds of billions.
Currently, there is no strict definition of physical AI; broadly speaking, it refers to AI that can understand physical laws and interact with the physical world.
Autonomous driving and humanoid robots are currently the two most intuitive applications: the former enables machines to perceive, judge, and act on roads, while the latter seeks to give AI a physical body capable of interacting with the real world.
But regardless of whether the final vehicle is a car, a robot, or another industrial device, for AI to truly impact the carbon-based world from its silicon-based realm, it must confront one fundamental truth: the real world is 3D.
NVIDIA's robotics research focuses on exploration of the 3D world.
Huge market potential and entirely new competitive arenas—tech giants, leading scholars, and industry capital are all exploring their own pathways as AI enters the physical world.
It was Jensen Huang who first brought physical AI to the public eye. He has repeatedly emphasized at major tech events like GTC and CES that “the next generation of AI is physical AI.” Moreover, Huang has placed his daughter, Michelle Huang, in NVIDIA’s Omniverse division (focused on 3D simulation) and his son, Steven Huang, in charge of robotics—both unquestionably strategic departments central to NVIDIA’s physical AI roadmap for the next decade.
Li Fei-Fei, known as the "Godmother of AI," founded World Labs, betting on spatial intelligence with the goal of enabling AI to understand the 3D world; after leaving Meta, Yann LeCun founded AMI Labs, focusing on teaching AI to predict changes in the real world; and Bezos's involvement with Prometheus directly ties into building a foundational engine for manufacturing scenarios.
Over the past few years, AI's silicon-based capabilities have progressively advanced along the dimensions of human perception: text addressing linguistic expression, images addressing visual expression, and video adding the dimension of time. With each new dimension unlocked, capital and talent flood in, rapidly turning this new technological frontier into a high-stakes battleground.
Back in the 3D and physical world, everyone will eventually enter. But this time, the battlefield takes a form no longer entirely the same as that of generative worlds.
The physical world does not accept "approximately correct."
In exploring how to achieve 3D mapping in the physical world, two fields—gaming and industry—have long been deeply established.
Gaming is a natural use case, and the reason is straightforward. The gaming industry requires vast amounts of 3D assets, such as characters, environments, and props, while character animations and environmental interactions must adhere to physical laws. Real-time 3D engines like Unity and Unreal are designed specifically for this purpose.
Industrial 3D operates on an entirely different set of logic and represents a scenario more closely tied to the physical world.
In a generative probabilistic world, images and videos that are approximately correct can already be surprising—but in industrial applications, products with less than 99% accuracy may be considered defective. A screw hole off by half a millimeter can halt an assembly line; poor surface continuity can cause issues in both manufacturing and simulation.
Gong Minyan, founder of Ziqian Technology, has experienced both types of 3D worlds: one represented by Dassault Systèmes, a CAD giant in industrial software, and the other by Unity, a leader in gaming and real-time 3D.
This experience gave him a more direct sense of the difference between the two approaches, which he summarized as the distinction between “drawing skin” and “drawing bone”: 3D in the virtual world primarily serves perception and presentation, while 3D in the physical world must serve verifiable engineering outcomes [5].
In a "virtual-to-virtual" world, AI's most important capabilities are generation speed and expressiveness, addressing content production efficiency, with commercial pathways closely aligned with game assets, film and television production, marketing displays, and virtual spaces.
In the world of "virtual to physical," generation is only the first step—a 3D model must also be editable, measurable, simulatable, and manufacturable, while preserving engineering intent. It must also connect to the physical world: AI must deeply understand geometry, materials, constraints, forces, processes, tolerances, and manufacturing feedback, ultimately delivering a digital object into the real world and ensuring it withstands real-world validation.
The market ceiling for 3D in gaming is determined by content spending, whereas industrial 3D faces the entire manufacturing sector's R&D, design, and production budgets.
Bezos is re-entering the arena with Prometheus, while Huang continues to champion physical AI—both pointing to the same trend: the main battlefield of AI is shifting from generating content to generating products, from transforming the software world to transforming the hardware world.
In Iron Man, Tony Stark only needs to give a vague idea, and JARVIS can coordinate various devices to gradually turn a single sentence into a wearable suit. In reality, engineering systems are far from this seamless—human ideas must first be translated into geometric models, then undergo calculation, verification, manufacturing, and assembly; any loss of information at any stage could cause the final product to deviate from the original vision.
Over the past few decades, this industrial-grade high-precision entry has primarily been achieved through CAD (computer-aided design) software.
CAD detail drawing of a jet engine
CAD may appear to be a drawing software, but it is in fact the physical foundation of modern civilization. Industries such as architecture, mechanical engineering, aerospace, shipbuilding, medical devices, and textiles and apparel are all closely tied to it—indeed, CAD embodies the greatest concentration of industry-specific knowledge humans have accumulated in the physical world.
As AI enters the 3D realm, the closer it gets to the physical world, the more precision and constraints are required; the closer it gets to manufacturing systems, the more the value of CAD as an entry point will be highlighted. But this is also where the challenge of AI connecting to reality lies: how can we re-constrain general capabilities into engineering problems, enabling AI to transform the world faster and more reliably?
AI CAD: A New Gateway Connecting the Physical World
In the past, CAD was sold primarily to engineers for use. Today, with AI capabilities having evolved to this point, CAD now has a new type of user: machines.
This means CAD's business logic is no longer just about making drawing more efficient, but also about enabling AI to perform engineering tasks more reliably. This is a transformative impact of AI on CAD, changing not only the front-end user experience but also the distribution of value.
As machines become new users, CAD transforms from a human-computer software system reliant on interfaces and commands into an infrastructure directly accessible by AI.
The software foundation needs to modularize and interface tools so that AI can understand intent, break down tasks, and complete closed-loop operations. Furthermore, the business model will shift from charging by seat to charging by task, number of calls, and workflows.
Gong Minyan, founder of Ziqian, summarizes it more directly: Previous-generation CAD helped humans project their intentions into the digital world, while Ziqian aims to do the opposite—enabling digital systems to truly manipulate the atomic world, creating a closed loop between digital design, physical verification, and real-world manufacturing.
Global giants are also approaching this goal from different directions. NVIDIA is building a training ground for physical AI using Omniverse, simulation platforms, and computing infrastructure. Traditional industrial software giants like Dassault and Siemens are extending their engineering boundaries by focusing on industrial software and digital twins.
Traditional industrial software giants possess deep product ecosystems, but their legacy architectures are bulky, making it difficult to pivot; generative 3D companies excel at visual generation but lack engineering constraints and manufacturability; general-purpose large model companies have stronger intelligence, yet behind their broad versatility lies a lack of essential vertical industry tools and data assets.
For a team with many years of deep experience in the 3D world, Ziqian Technology believes that CAD sits at the intersection of the virtual and physical worlds—it is both a tool for humans to design products, describe spaces, and express engineering intent, and an infrastructure connecting simulation, manufacturing, and physical execution.
CAD models carry not only three-dimensional shapes but also physical principles and manufacturing processes, offering a digital representation closer to the essence of the physical world than images or videos.
Driven by this philosophy, Ziqian Technology has developed a technical roadmap spanning the generation, verification, and closed-loop physical world Dream Loop.
In the first phase, it's "point-and-shoot" Agentic CAD.
After large models empower software with understanding capabilities, Agentic CAD, leveraging Ziqian Technology’s cloud-native 3D CAD foundational tools, long-term accumulation of industrial data, and iterative evolution of the Harness system, can comprehend engineering intentions expressed in natural language, decompose user goals into a series of executable modeling actions, and invoke reliable geometric modeling tools to complete operations.
This process is akin to AI coding in a 3D world, transforming not only modeling efficiency but also the fundamental way humans interact with industrial software—from humans learning the machine’s operational language to machines understanding human creative intent.
Phase two: transitioning from physical structure to verifiable physical products.
Just as the core principle of implementing physical AI lies in understanding physical laws and interacting with the physical world, this step involves placing the product into Ziqian’s physical world simulation platform to verify the accuracy of physical principles such as mechanics, kinematics, dynamics, and motion control.
This process can autonomously iterate and close the loop with the product generation of the first phase until the most reasonable and optimal solution is delivered.
Stage three: moving toward manufacturing, achieving the ultimate goal of impacting the physical world.
The seamless operation across three stages will truly reduce the time from idea conception to physical realization to one-tenth of its original duration, or even less.
For Zi Qian, this is not merely a cloud-based CAD software, nor simply an efficiency improvement akin to Prometheus—it is the precise alignment of AI with the physical world.
When more judgment and imagination can be entrusted to human engineers, the CAD of the future will become a new gateway to human creativity.
Epilogue
Narratives around emerging industries often begin with a grand term and eventually grow into a comprehensive industrial ecosystem.
Looking back at several waves of technological change, shifts in entry points have often determined the power structure of industries. During the personal computer era, operating systems became the gateway for software development and hardware compatibility; in the mobile internet era, iOS and Android controlled the rules for developers; in the AI training era, CUDA transformed GPUs from mere hardware into an indispensable computing platform for developers.
The connection between AI and the physical world will be the same.
On this path reshaping the global industrial system, Bezos is not the only one who sees the opportunity. In China, a group of cloud-native industrial software companies have been integrating AI capabilities into real-world engineering and manufacturing processes even earlier. Today, the goals pursued by all parties have become increasingly clear:
Whoever becomes the first entry point for AI into the physical world truly holds the definition of "JARVIS."
Reference materials
[1] Prometheus, Jeff Bezos' AI startup, is now worth $41 billion, Axios
[2] Bezos's AI startup Prometheus raises $12 billion, valued at $41 billion — Wall Street Journal
[3] Nvidia believes physical AI systems represent a $50 trillion market opportunity, GamesBeat
[4] The generative AI market is poised to reach $2.3 trillion by 2032 as agentic systems proliferate and demand for infrastructure surges, according to Bloomberg Intelligence.
[5] Digital Twin: From "Skin Deep" to "Bone Deep"—Paving the Way for Industrial Internet and Smart Cities, Yicai