Physical AI Is the Next Frontier

The AI conversation is about to change. For three years, we've focused on what AI can say and generate. The next decade will be defined by what AI can do—physically, in the real world, with hands, wheels, and sensors that interpret touch, force, and motion.

For enterprise leaders, the strategic question has shifted: not whether AI will enter the physical world, but whether your organization is positioned to benefit when it does.

Most professionals use AI assistants the same way they'd use a search engine—one question at a time, starting from scratch with every conversation. They're leaving 80% of the value on the table.

The real power of Claude Projects, ChatGPT CustomGPTs and Projects, and Google Gemini Gems isn't having a smarter assistant—it's building persistent, menu-driven systems that codify your expertise and automate your workflows.

In this session, you'll learn how to transform general-purpose AI into purpose-built tools that replicate complex tasks with consistency and precision.

Join us on January 27, 12:00 PM EST and leave with practical frameworks you can use immediately!

🎙️ AI Confidential Podcast - Are LLMs Dead?

☕️ AI Tangle - Ads, Agents, Healthcare Models & $100B Lawsuits — AI’s Real World Hits the Mainstream

🔮 AI Lesson - Take Your Google Workspace to the Next Level with Gemini

🎯 The AI Marketing Advantage - Google’s AI Is Running Your Ads Now (Whether You Like It or Not)

💡 AI CIO - Fresh Minds Outsmart the Experts

 📚 AIOS - This is an evolving project. I started with a 14-day free AI email course to get smart on AI. But the next evolution will be a ChatGPT Super-user Course and a course on How to Build AI Agents.

Why the shift from language models to sensory models will determine which enterprises lead—and which get left behind

The evolution of AI follows a clear trajectory. First came language models—systems trained on text that could predict the next word. Then came multimodal models—systems that could process images, audio, and video alongside text. Now comes physical AI: systems that can perceive, reason, and act in the three-dimensional world.

Each stage requires capabilities that the previous stage lacked. A language model can write instructions for assembling furniture. A multimodal model can watch a video of someone assembling furniture and describe what's happening. A physical AI system can actually assemble the furniture—picking up pieces, applying appropriate force, adjusting when something doesn't fit.

That last capability—acting in physical space with real-time sensory feedback—represents the frontier that matters for enterprise.

Physical AI refers to artificial intelligence systems designed to operate in and interact with the physical world. Unlike traditional AI that processes digital inputs and produces digital outputs, physical AI must handle the messy, continuous, unpredictable nature of reality.

The key differentiator is sensory integration. A physical AI system needs to process multiple streams of real-world data simultaneously: visual input from cameras, depth information from lidar, force feedback from tactile sensors, proprioceptive data about its own body position, and often much more. It then needs to translate high-level goals into low-level motor commands that account for physics, safety, and uncertainty.

This is fundamentally different from generating text or images. When ChatGPT produces an incorrect answer, you can ask it to try again. When a robot applies incorrect force while handling a component, it may damage equipment, injure someone, or destroy itself. The margin for error in physical systems is measured in millimeters and milliseconds.

NVIDIA's Jensen Huang captured the shift in his CES 2026 keynote: "The ChatGPT moment for robotics is here. Breakthroughs in physical AI—models that understand the real world, reason, and plan actions—are unlocking entirely new applications."

What makes this moment different from previous robotics hype cycles is the convergence of three capabilities: foundation models that can reason about physical scenarios, simulation environments that can generate training data at scale, and hardware platforms that can run these models in real-time on moving systems.

The clearest example of physical AI at scale is Waymo. The Alphabet subsidiary now provides over 450,000 paid robotaxi rides per week across five U.S. cities—Phoenix, San Francisco, Los Angeles, Austin, and Atlanta. By the end of 2026, Waymo aims to reach 1 million rides per week. In 2025 alone, the service completed 14 million trips.

These numbers matter because they represent real commercial deployment, not demonstrations. Waymo vehicles navigate construction zones, respond to ambulances, handle San Francisco's steep hills and dense pedestrian traffic—all without a human safety driver. The company's safety data shows its robotaxis are involved in five times fewer injury-causing accidents per million miles than human drivers.

What makes this possible isn't just cameras and lidar. It's the AI stack that processes sensor data into situational understanding. Waymo's vehicles don't just detect objects—they predict behavior, model uncertainty, and make thousands of decisions per second about speed, trajectory, and risk. They understand that a ball rolling into the street probably means a child is about to follow. They recognize a pedestrian's body language when they're about to jaywalk.

This kind of contextual physical reasoning is what separates true physical AI from earlier automation. And it's now being applied beyond vehicles.

Figure AI completed an 11-month deployment at BMW's Spartanburg plant in South Carolina, with its Figure 02 humanoid robots running 10-hour shifts loading sheet metal into welding fixtures. The robots contributed to the production of over 30,000 vehicles. More importantly, they demonstrated the minimum viable product for factory humanoids: performing a limited set of tasks reliably, over extended periods, with measurable results.

The technical requirements for physical AI are divided into three categories:

NVIDIA's new model releases target each layer. Cosmos Reason 2 handles perception and reasoning—it's a vision-language model that enables machines to see, understand, and contextualize physical environments. Isaac GR00T N1.6 handles action—it's a vision-language-action model purpose-built for humanoid robots that translates reasoning into whole-body control. Together with simulation frameworks like Isaac Lab-Arena, they form a complete stack for training, testing, and deploying physical AI systems.

For most enterprises, physical AI implementation will follow the pattern set by Waymo and factory robotics: start narrow, prove value, expand scope.

Phase 1: Identify Structured Physical Tasks

Physical AI today excels at tasks that are repetitive, rule-governed, and occur in controlled environments. Warehouse picking. Assembly line loading. Last-mile delivery in geofenced areas. Inspection routes that follow predictable paths.

The questions to ask: What physical tasks in your operation are high-volume, well-defined, and currently performed by humans? What's the cost of errors in those tasks? What data exists about how those tasks are currently performed?

Figure AI's success at BMW came from choosing sheet metal loading—a classic pick-and-place task with clear success criteria (cycle time under 84 seconds, placement accuracy within tolerance, minimal interventions). The task is boring, repetitive, and ergonomically challenging for humans. It's perfect for a first deployment.

Phase 2: Establish Measurement Infrastructure

Physical AI deployment requires granular measurement that most organizations lack. You need to know: How long does the task currently take? What's the error rate? What causes variations? What safety incidents occur?

Waymo's advantage comes partly from the 100+ million miles of autonomous driving data it has accumulated. Every edge case, every unusual scenario, every near-miss feeds back into training. Organizations deploying physical AI need similar feedback loops.

This often means instrumenting processes before automation. Cameras, timers, quality sensors, and incident tracking create the baseline against which AI performance will be measured—and the data that will improve it.

Phase 3: Pilot with Clear Success Metrics

Physical AI pilots should have quantified targets before deployment begins. Figure AI's BMW project tracked three metrics: cycle time (37-second target for sheet metal loading), placement accuracy, and the number of human interventions required. Everything else was secondary.

Conservative success criteria matter. Physical systems that fail can cause injury, damage equipment, or halt production. A pilot that achieves 80% of human performance while demonstrating safety and reliability is more valuable than one that occasionally matches human performance but introduces unpredictable failures.

Phase 4: Build for Fleet Learning

The economics of physical AI improve with scale because learning compounds. When one Waymo vehicle encounters an unusual situation—a parade, a flooded intersection, an aggressive driver—the learnings propagate to the entire fleet. When one factory robot masters a tricky assembly sequence, the skill transfers to every robot in the network.

This requires architecture decisions upfront. How will data flow from deployed systems back to the training infrastructure? How will model updates reach deployed robots? What testing validates that updates improve performance without introducing regressions?

Organizations that treat physical AI deployment as a one-time project will find themselves constantly re-implementing. Those who build for continuous learning will see compounding returns.

The economic case for physical AI rests on three factors: labor arbitrage, capability extension, and data generation.

The shift from language AI to physical AI will reshape enterprise technology strategy over the next five years. Three planning considerations deserve immediate attention.

First, the workforce strategy must account for augmentation before replacement. Physical AI systems will initially work alongside humans, handling specific tasks while humans manage exceptions, quality, and coordination. Organizations that frame robotics as immediate headcount reduction will underinvest in the human capabilities—supervision, maintenance, process optimization—that successful deployment requires.

Second, data infrastructure becomes a competitive asset. Every physical AI deployment generates operational intelligence. Organizations with substantial data collection, storage, and analysis capabilities will extract greater value from robotics investments and improve more quickly. Those without will find themselves dependent on vendor learning curves rather than their own operational data.

Third, the geographic dimension matters. China's 5-to-1 patent advantage in humanoid robotics and its cost leadership in robot components signal where manufacturing capability is concentrated. U.S. organizations face choices about supply chain resilience, technology partnerships, and build-versus-buy decisions that have geopolitical implications.

The question is no longer whether AI will operate in the physical world. Waymo's 450,000 weekly rides settle that. The question is whether your organization will be among those defining how physical AI integrates into your industry—or among those adapting to standards set by others.

Author’s note: This week’s complete edition—including the AI Toolbox and a hands-on Productivity Prompt—is now live on our website. Read it here.

Join us at All Things AI 2026 in Durham, North Carolina, on March 23–24, 2026.

This is where AI gets real. No sales pitches—just 4,000 builders, engineers, operators, and execs sharing how AI actually works in practice, from hands-on workshops to real-world sessions, right in the heart of RTP and Duke’s AI ecosystem.

Early registration is open, and prices go up soon.

I appreciate your support.