Andreessen Horowitz (a16z) is shifting its AI investment thesis away from large language models toward a new frontier: AI systems that interact with the physical world. In a detailed technical essay, the venture firm outlines why robot learning, autonomous science, and novel human-machine interfaces represent the next major opportunity for AI scaling, arguing these domains share foundational technologies that could create compounding advantages .

What Technical Breakthroughs Are Making Physical AI Possible?

The shift reflects a maturing set of technical capabilities that didn't exist even 18 months ago. a16z identifies five core technical primitives that are converging to enable AI systems to understand and act in the physical world :

• Learned Physical Representations: AI models can now learn compressed, general-purpose understanding of how objects move, deform, collide, and respond to force, eliminating the need for each system to learn physics from scratch.
• Vision-Language-Action Models (VLAs): Systems like Physical Intelligence's pi-zero, Google DeepMind's Gemini Robotics, and NVIDIA's GR00T N1 extend pretrained vision-language models with action decoders that output motor commands, amortizing the enormous cost of learning to see across internet-scale image-text pretraining.
• World Action Models (WAMs): Built on video diffusion transformers pretrained on internet-scale video, these models inherit rich priors about physical dynamics and couple them with action generation, with NVIDIA's DreamZero demonstrating zero-shot generalization to entirely new tasks.
• Native Embodied Foundation Models: Generalist's GEN-1 takes a different path, trained from scratch on over half a million hours of real-world physical interaction data collected through low-cost wearable devices on humans performing everyday manipulation tasks.
• Spatial Intelligence: Companies like World Labs are building models that reconstruct and reason about the full 3D structure of physical environments, addressing a representation gap where existing approaches only capture 2D visual features or video projections.

The convergence matters because whether representations are inherited from vision-language models, learned through video, or built natively from wearable sensor data, the underlying capability is the same: compressed, transferable models of physical behavior. a16z argues the data flywheel for these representations is "enormous and largely untapped," encompassing not just robot trajectories and internet video, but the vast corpus of human physical experience that wearable devices are now capturing at scale .

Why Are These Three Domains the Next Frontier?

a16z identifies three specific application areas as representing the greatest opportunity for frontier AI development: robot learning, autonomous science (particularly in materials and life sciences), and new human-machine interfaces including brain-computer interfaces, silent speech interfaces, neural wearables, and novel sensory modalities like digitized olfaction .

The firm's reasoning is strategic. These domains sit at an optimal distance from the current AI paradigm: close enough to inherit infrastructure and research momentum from large language models, but distant enough to require non-trivial additional work. That distance creates a natural moat against fast-following competitors and defines a problem space that is richer, less explored, and more likely to produce emergent capabilities precisely because the easy paths have not already been taken .

Beyond technical progress, each of these areas has seen the beginnings of an influx in talent, capital, and founder activity. The technical primitives for extending frontier AI into the physical world are maturing concurrently, and the pace of progress suggests these fields could soon enter a scaling regime of their own .

How Are These Domains Reinforcing Each Other?

a16z emphasizes that these three domains are not entirely separate efforts. They share a common substrate of technical primitives, including learned representations of physical dynamics, architectures for embodied action, simulation and synthetic data infrastructure, an expanding sensory manifold, and closed-loop agentic orchestration. They are mutually reinforcing in ways that create compounding dynamics across domains .

The same representations that serve a robot learning to fold towels can also serve a self-driving laboratory predicting reaction outcomes or a neural decoder interpreting the motor cortex's plan for grasping. This cross-domain applicability suggests that breakthroughs in one area could accelerate progress in the others, creating what a16z calls a "structural flywheel for extending AI into the physical world" .

What Does This Mean for Founders and Workers?

While a16z is optimistic about the long-term potential of physical AI systems, the firm's co-founder Ben Horowitz has acknowledged a growing tension in the market. Founders are experiencing what he calls "AI anxiety from above," the fear of not moving fast enough in a market where execution timelines have collapsed. The window that once gave a strong software product 10 years of runway, then five years, has now compressed to "maybe five weeks," according to Horowitz .

"If you keep looking at it like the old world, and it's got completely different laws of physics, you are definitely going to die," said Ben Horowitz, co-founder and general partner at Andreessen Horowitz.

Ben Horowitz, Co-founder and General Partner, Andreessen Horowitz

Horowitz argues that the two competitive moats that software CEOs relied on for decades are both gone. "You can buy enough GPUs and solve basically anything in software," he explained, and customer lock-in through switching costs has evaporated because "it's very easy to replicate the code. It's very easy to move the data" .

Horowitz

On the ground inside companies, however, workers are experiencing a different anxiety. Rather than fear of moving too slowly, workers are afraid they're becoming irrelevant altogether. Roughly half of American workers now name AI-driven job loss as one of their primary fears, a share that has nearly doubled in a single year, according to KPMG research .

The behavioral consequence is already visible in enterprise data. A global survey of 3,750 executives and employees across 14 countries by WalkMe found that more than 54 percent of workers bypassed their company's AI tools in the past 30 days and completed the work manually instead; another 33 percent haven't used AI at all. Combined, roughly eight in ten enterprise workers are either avoiding or actively rejecting the technology their employers are spending record sums to deploy, even as average digital transformation budgets rose 38 percent year-over-year to $54.2 million .

The perverse irony, documented in the research, is that the fear itself accelerates the outcome workers dread most. Workers who resist AI adoption fall further behind peers leveraging the tools, in some cases by a factor of 10 or 20 to one in productivity .

Horowitz remains optimistic about the long-term trajectory, invoking the industrial revolution as a historical parallel. When more than 90 percent of Americans were farmers before virtually all of those jobs were automated away, technology eliminated jobs people recognized and created ones they could not yet imagine. "The history of technology is things have always gotten better," he said . However, he acknowledged the underlying tension: if AI systems become capable enough to replace human workers entirely, the fundamental business model of software companies selling to customers breaks down.

For now, fewer than 19 percent of U.S. establishments have adopted AI according to Goldman Sachs economists, meaning the revolution a16z is racing toward has barely begun. The gap between founders' anxiety about pace and workers' anxiety about purpose is where the real disruption lives, and right now almost no one is bridging it .