The move to physical AI

For the past two years, the cost of building software collapsed. Vibe coding tools, AI coding assistants, and hosted inference APIs made it possible for a solo developer with a $20 monthly subscription to ship a product that used to require a team of ten. The barrier to entry for software fell to almost nothing. And because the barrier fell for everyone simultaneously, the result was predictable: a flood of near-identical products, rapid copying, and a race to the bottom on differentiation. Physical AI is the opposite game entirely. When AI moves from screens into the physical world, into robots, autonomous systems, and embodied machines, the rules change. Not everyone can play. You need capital, connections, deep technical knowledge, and access to manufacturing infrastructure just to get started. The stakes are higher, the feedback loops are slower, and the cost of failure isn't a wasted weekend of prompting. It's millions of dollars in hardware that doesn't work.

Software AI became a commodity

The democratization of software AI happened fast. Claude, ChatGPT, Gemini, and their competitors made code generation accessible to anyone who could describe what they wanted in plain language. A wave of tools like Cursor, Windsurf, and Replit turned natural language into functioning applications. The result was an explosion of shipping speed. Small teams and solo builders could prototype, iterate, and launch in days instead of months. This was genuinely transformative. But it also meant that the protective moat around software products eroded almost overnight. If anyone can build an app in a weekend, then the app itself isn't the defensible asset anymore. The cost of creating software approached zero, which meant the cost of copying software approached zero too. Startups found themselves competing not on technical capability but on distribution, brand, and speed of iteration, because the underlying product could be replicated by anyone with the same AI tools. The software layer of AI became, in economic terms, a commodity. Valuable, yes. But not scarce.

Physical AI plays by different rules

Physical AI is the application of artificial intelligence to systems that interact with the real world: humanoid robots, autonomous vehicles, industrial automation, drones, surgical systems. NVIDIA CEO Jensen Huang declared at GTC 2026 that "every industrial company will become a robotics company" and called the moment "the big bang of physical AI." Gartner named physical AI one of its top 10 strategic technology trends for 2026. The momentum is real. But the barrier to entry is orders of magnitude higher than anything in software. Consider the capital requirements alone. Figure AI has raised approximately $1.9 billion across all funding rounds, reaching a $39 billion valuation after its Series C in September 2025. Physical Intelligence, a two-year-old robotics startup founded by ex-Google DeepMind researchers, raised $1 billion in late 2024 and was reportedly in talks for another $1 billion at an $11 billion valuation by March 2026. Advanced Machine Intelligence, a Paris-based startup building world models for physical AI, closed a $1.03 billion seed round, the largest seed round in European history. These aren't software companies burning through cloud credits. They're building physical things that need physical factories. Figure AI announced BotQ, a dedicated high-volume manufacturing facility for humanoid robots, staffed by manufacturing engineers who design assembly lines, select tooling, and optimize production cycle times. Tesla is converting production lines at its Fremont factory into Optimus robot manufacturing capacity, targeting one million units per year. These are industrial operations that require supply chains, quality control, safety certifications, and enormous upfront capital expenditure. In the AI chip space, the numbers are even more staggering. As of April 2026, 18 out of 19 disclosed chip funding rounds over the past 12 months were $50 million or larger, with a median round size of $400 million. Cerebras Systems raised $1 billion in its Series H. Zero disclosed deals sat under $20 million. This is a financing profile you normally see in mature industrial sectors, not startups.

Why you can't vibe code a robot

The fundamental difference between software AI and physical AI comes down to atoms versus bits. In software, iteration is almost free. You write code, deploy it, see what breaks, and fix it. The feedback loop is minutes. The cost of a mistake is a bug report. You can ship ten versions in a week and keep the one that works. In physical AI, iteration is expensive, slow, and sometimes dangerous. A robot that drops an object in a factory doesn't just generate a bug report. It damages equipment, delays production, or injures someone. Every failure costs real money, and the feedback loop isn't minutes. It's weeks or months of redesign, retooling, and retesting. The hardware itself creates barriers that software never faces. You need sensors that can perceive depth, texture, temperature, and force in real time. You need actuators precise enough to handle delicate objects without breaking them. You need battery systems that can sustain hours of continuous operation. You need mechanical designs that survive the physical stresses of real-world environments. Each of these components has its own supply chain, its own manufacturing constraints, and its own failure modes. Then there's the data problem. Software AI models train on text and images scraped from the internet, essentially free and abundant. Physical AI models need data collected from real-world interactions: robots navigating actual environments, manipulating actual objects, responding to actual variability. This data is expensive to collect, hard to scale, and can't simply be downloaded. As one investor noted, every dollar of equity spent gathering physical training data dilutes the cap table before you've even scaled deployment. Simulation helps bridge the gap, with NVIDIA's Cosmos and Isaac platforms enabling robots to train in virtual environments before touching the real world. But simulation-to-reality transfer remains an active research challenge. A robot that performs flawlessly in simulation can still fail unpredictably when confronted with the messiness of actual physics.

The capital moat is real

This creates a competitive dynamic that looks nothing like software. In software AI, a talented individual with a laptop and a subscription can build a product that competes with well-funded startups. In physical AI, the minimum viable entry point is measured in hundreds of millions of dollars. Over $34 billion in private capital flowed into robotics-related companies in 2025, more than double the amount in 2024. That capital represented 9% of all venture funding globally. But the distribution was heavily concentrated. A handful of companies, Figure, Physical Intelligence, Boston Dynamics, Tesla, and a few others, absorbed the majority of investment. The long tail of smaller players faces a structural disadvantage that no amount of clever engineering can overcome without matching capital. CB Insights and other analysts have pointed out that physical AI has a fundamentally different capital structure problem than software. Software AI scales on intellectual property, with low marginal deployment costs. Equity financing works well for that model. Physical AI requires hardware manufacturing, fleet deployment, and real-world data collection at every stage of growth. The capital intensity doesn't shrink as you scale. It compounds. This means the competitive landscape for physical AI will likely look more like the automotive or semiconductor industry than the software industry. A small number of well-capitalized players with manufacturing expertise, supply chain relationships, and deep engineering teams will dominate. The long tail of indie builders and small startups that thrives in software AI will find the barriers in physical AI far harder to overcome.

Geography matters again

Software AI is location-agnostic. You can build a competitive AI product from a laptop in Lisbon or a coworking space in Bali. Physical AI ties you to geography in ways that software never did. Manufacturing capacity is concentrated. Semiconductor fabs take years to build and billions to finance. Robotics supply chains depend on specific component manufacturers, many of them in East Asia. Access to testing environments, whether factory floors or logistics warehouses, requires physical proximity and partnership with industrial operators. Geopolitics adds another layer of complexity. Export controls on advanced semiconductors, competition for rare earth materials, and national strategies for robotics development all shape who gets to build physical AI and where. China's robotics ecosystem demonstrated its scale at a half-marathon event in Beijing where 300 humanoid robots from 26 brands competed, not as a technology demo but as a display of industrial capacity. The United States, Europe, and Japan each have their own clusters of capability, shaped by their respective manufacturing bases and policy environments. For builders accustomed to the borderless world of software, this reintroduction of geography as a constraint is jarring. Your ability to compete in physical AI depends partly on where you are, who you know, and which supply chains you can access.

What this means for builders

If you're a software developer or indie builder watching the physical AI wave and wondering how to participate, the honest assessment is sobering. You can't bootstrap a humanoid robot company the way you can bootstrap a SaaS product. But the ecosystem around physical AI does create opportunities that don't require building robots from scratch. Simulation tools, training data infrastructure, middleware for robot fleet management, safety certification services, and specialized software for physical AI workflows are all emerging niches where software expertise translates. The startup Antioch raised $8.5 million to build simulation tools for robot developers, positioning itself as "the Cursor for physical AI." That's the kind of opportunity that exists at the intersection of software capability and physical AI needs. The broader implication is that we're entering a period where the AI landscape bifurcates. Software AI will continue to democratize, with costs falling and accessibility increasing. Physical AI will consolidate, with capital requirements rising and the number of viable players shrinking. The two domains will coexist but operate under completely different economic logic. Vibe coding made it possible for anyone to build an app. Nobody is going to vibe code a factory robot. The move to physical AI isn't just a technology shift. It's a shift in who gets to participate, and that changes the entire game.

References