--- title: "The Physical AI Era" subtitle: "AI has a body now, and the discipline to run it is not yet in place." author: "Ritesh Vajariya" publication: "The Forward View" publishedAt: 2026-06-28T14:09:31.249Z updatedAt: 2026-06-28T14:12:44.084Z pillar: second-order pillarLabel: "SECOND ORDER" tags: ["physical-ai", "robotics", "industrial-ai", "governance", "liability", "foundation-models"] readingTimeMinutes: 10 canonical: https://theforwardview.com/essays/the-physical-ai-era --- # The Physical AI Era ## The Floor Did Not Match the Sign Mexico. May 2026. The visitor center was polished. Branded language about AI-piloted manufacturing, clean display cases, and the full picture of a company that had arrived at the future. I spent maybe 15 minutes in there before someone walked me through the door onto the production floor. What was actually running was a working system, but not a coherent one: automated lines and human operators in the gaps, all of it sitting on top of orchestration software stitched together over years of incremental decisions. I counted more humans than robots. This is not a failure. The plant ran. Parts moved. Vehicles got built. But I kept looking at the phrase on the wall in the visitor center, then at the floor, and the two pictures did not match. Two days later, I was watching Jensen Huang's keynote at GTC Taipei on a livestream. He named physical AI as the third wave (after perception AI and generative AI) and Cosmos 3 as the foundation model built to run it. The Mexican floor was still in my head. The keynote reframed what I had just seen on the Mexico floor. That is when I started paying attention to how often the phrase "Physical AI" was getting used as if everyone in the room had agreed on what it meant. --- ## What the Term Actually Covers [NVIDIA launched Cosmos 3](https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai) on May 31, 2026, at GTC Taipei. Open frontier foundation model, purpose-built for physical AI. A single architecture that brings together vision reasoning, world modeling, and the action generation layer that translates the first two into physical motion. Jensen's framing about it being the third wave dates back to [his CES 2025 keynote](https://blogs.nvidia.com/blog/ces-2025-jensen-huang/), but the model was shipped on May 31. ![Jensen Huang on stage at NVIDIA GTC Taipei announcing Cosmos 3, the open frontier foundation model purpose-built for physical AI — third wave after perception AI and generative AI.](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/physical-ai-era/image-1-jensen.png) *Cosmos 3 announced at GTC Taipei, May 31, 2026. Image: NVIDIA.* The category, as defined, covers four segments, each already in deployment. **Humanoid robots in industrial use.** [BYD and UBTECH](https://www.researchandmarkets.com/report/humanoid-robot) are running 100-200 units in what is currently the world's largest commercial humanoid deployment. [GXO and Agility Robotics](https://www.researchandmarkets.com/report/humanoid-robot) have 100 units contracted through 2026. [BMW and Figure AI](https://mlq.ai/research/physical-ai/) have 15-30 units running at the Spartanburg plant: 11 months of continuous operation, more than 90,000 parts handled. [Mercedes and Apptronik](https://www.researchandmarkets.com/report/humanoid-robot) have 10-20 units running tote delivery. **Robotaxis at commercial scale.** [Waymo](https://waymo.com/blog/2025/05/scaling-our-fleet-through-us-manufacturing/) is running more than 3,000 vehicles, [500,000 paid rides per week](https://techcrunch.com/2026/03/27/waymo-skyrocketing-ridership-in-one-chart/), across [1,400 square miles](https://electrek.co/2026/05/13/waymo-expands-coverage-1400-square-miles-11-cities/) of fully driverless coverage, an area larger than Rhode Island. **Warehouse and logistics robotics.** [Amazon](https://www.roboticscenter.ai/applications/warehouse-robotics) is running more than 750,000 robots across its global fulfillment network. [Symbotic](https://unteachablecourses.com/warehouse-robots-2026/) is carrying a backlog above $5 billion after the January 2026 Walmart deal. **Specialized embodied AI** in surgical robotics, agricultural systems, defense applications, and infrastructure inspection. Each of these operates in a constrained, high-consequence environment where the failure mode is not an incorrect text output. [Goldman Sachs](https://www.goldmansachs.com/insights/articles/the-global-market-for-robots-could-reach-38-billion-by-2035) revised its humanoid market projection sixfold to $38 billion by 2035, with the broader physical AI total addressable market running into the tens of trillions on [Morgan Stanley](https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050)'s longer view. In June 2026, [SoftBank's Masayoshi Son](https://www.techtimes.com/articles/317693/20260603/humanoid-robots-investment-race-heats-goldman-6x-forecast-china-leads-spy-law.htm) called physical AI the next trillion-dollar category. ![Three-panel comparison of physical-AI market forecasts: Goldman Sachs $38B humanoid market by 2035, Morgan Stanley $60T by 2050, SoftBank framing physical AI as the next trillion-dollar category.](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/physical-ai-era/image-2-market-sizes.png) *Three forecasts, three timeframes, one direction of travel.* And as of June 16, [Alibaba entered the model layer directly](https://technode.com/2026/06/17/alibaba-unveils-qwen-robot-series-with-three-foundation-models-for-embodied-ai/). Its Qwen Robot Suite spans navigation and manipulation, plus a separate world model that lets the robot predict physical outcomes before acting. The suite is already in pilot testing with select Alibaba Cloud enterprise customers. The race to own the software infrastructure underneath physical AI is not a Western story. None of that is in dispute. The disagreement is about what the category actually contains, and what changes when it scales. --- ## The Three Reads Circulating Right Now The standard readings go something like this. The first read goes that physical AI is just LLMs with bodies. You take the generative AI stack, add a sensor array and some actuators, and the machine can now act in the world instead of just describing it. The third wave framing encourages this read. A second view, common among skeptics, treats physical AI as Jensen Huang's positioning of NVIDIA to sell more Blackwell GPUs. The foundation model needs wafer-scale compute to train and run. The market-maker names the market, and the hardware vendor benefits. A third read says physical AI is robotics rebranded. The automotive industry has been running coordinated robotic systems for 40 years. Boston Dynamics has been shipping robots since before the generative AI wave. The new label is capital-raising language for a category that already existed. Each of these captures something true. None of them tells you what is structurally different about this wave, or what the liability picture looks like when a system with a body acts incorrectly in a workplace. --- ## What the Mexico Floor Was Actually Showing ### What physical AI is The closed loop is the load-bearing concept. A system that perceives the world and acts on it, with a reasoning layer mediating between the two, running continuously in real-world environments where failure has physical consequences that cannot be edited after the fact. What makes the closed loop distinct is what it is not. Digital AI handles text and image generation only. Agentic AI takes software actions inside digital systems. Programmed robotics runs pre-coded sequences without learning. Physical AI sits on top of all three but is reducible to none of them. The Mexico floor was running a version of this. Partial, stitched together, with humans closing the gaps. But the architecture was there. The system perceives the environment and acts in response, with a reasoning step in between. Then it perceives again. The loop does not stop. ### What makes this wave different The model operates in physics, not in a training distribution. When a language model produces a wrong answer, the output is editable. When a physical AI system misjudges a force vector or misreads a human operator's position on a production floor, the failure has a body attached to it. Latency is bounded by safety, not by user experience. Sensor fusion (camera, lidar, IMU, force sensors) is not an interface to the system. It is the system. Hardware and software are co-designed, not separable. The edge case problem is also different. Long-tail real-world conditions on a factory floor or a city street are not the same problem as long-tail token distributions in a text model. The tail is longer, and the consequences of hitting it are not a hallucination. ### The compute and the capital underneath it Training data is the binding constraint. Cosmos 3 addresses this directly: world simulation as a synthetic data source for physical environments that cannot be fully covered by real-world collection. Wafer-scale and Blackwell-class computing is what runs this work at the required scale. As I covered in [The Gigawatt Era](https://theforwardview.com/essays/the-gigawatt-era-ai-infrastructure-power-scale-explained), the infrastructure to run the current AI frontier is a multi-year physical buildout. Physical AI adds the embodiment layer on top of that infrastructure. The capital is committed. [Figure raised $1 billion at a $39 billion valuation](https://theaiinsider.tech/2026/05/27/ai-funding-in-2026-where-venture-capital-is-going/). [Apptronik raised $935 million](https://theaiinsider.tech/2026/05/27/ai-funding-in-2026-where-venture-capital-is-going/). [FieldAI raised $405 million](https://theaiinsider.tech/2026/05/27/ai-funding-in-2026-where-venture-capital-is-going/) for physical AI foundation models. [Global venture capital in Q1 2026 hit $330.9 billion](https://kpmg.com/xx/en/media/press-releases/2026/04/global-vc-investment-surges-to-record-330-9-billion-dollar-in-q1-26.html), with AI capturing roughly 80 percent of it, and physical AI emerging as the distinct breakout category inside that figure. Component costs are [declining roughly 40 percent annually](https://www.kaisoresearch.com/report-store/global-physical-ai-market), against an expected 15 to 20 percent. The current humanoid bill of materials runs between $30,000 and $150,000 per unit. Tesla has targeted $20,000 to $30,000 at production scale. Goldman's base case calls for more than 250,000 humanoid units shipping in 2030 and 1.4 million annually by 2035, with $50 billion in cumulative humanoid investment by 2030. ### What the Mexico floor also showed The orchestration software was stitched together over the years. The human operators in the gaps that the automation had not closed. The question nobody put on the visitor center wall: who is responsible when the system acts incorrectly? In February 2025, [a worker at Tesla's Fremont facility was pinned by a robotic arm during a maintenance window](https://theresarobotforthat.com/blog/humanoid-robot-safety-standards-2026/). That case is the first major documented humanoid workplace injury lawsuit in the US. [OSHA covers humanoid workplace deployment under the General Duty Clause](https://theresarobotforthat.com/blog/humanoid-robot-safety-standards-2026/) (Section 5(a)(1)), with no specific standard yet in place. [ISO 10218](https://news.aliasrobotics.com/humanoids-and-physical-ai-the-future-has-a-body-is-it-ready-to-be-secure/), which governs industrial robots, is widely acknowledged as insufficient for humanoids with adaptive AI. [ISO 25785-1](https://www.kitecompliance.ai/vertical-compliance/future-of-humanoid-robot-compliance), covering dynamic stability and fall-risk for legged systems, is still in development. ![Articulated industrial robotic arm on an automotive manufacturing line — the embodied compute layer already in production at scale across the auto industry.](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/physical-ai-era/image-3-robotic-arm.png) *Closed-loop industrial robotics — the substrate physical AI is now layering reasoning onto.* The [EU AI Act's high-risk obligations apply from August 2, 2026](https://humanoid.press/www-humanoid-press/Humanoid-AI-Safety/). [Colorado's AI Act is effective June 30, 2026](https://ai-frontiers.org/articles/the-robot-in-your-living-room-has-no-rulebook). Both are weeks away. The category exists. The discipline to run it has not yet been established. --- ## What This Looks Like in 2029 Push the timeline out to 2029, and the picture bifurcates along a line that is already visible. Goldman's base case puts humanoid units in industrial use in the low millions globally, with Wood Mackenzie projecting robotaxi fleets crossing 100,000 vehicles globally by the end of the decade. Warehouse and logistics robots have become the baseline expectation in any new fulfillment facility, not a competitive differentiator. NHTSA, OSHA, FDA, and EPA each issue physical AI-specific guidance in this window, drawing heavily on existing process safety frameworks from industries that have been running autonomous control of complex physical environments for decades. Insurance underwriters reprice physical AI exposure with higher granularity. The carriers that built the actuarial table early set the terms for everyone else. The companies that imported process-industry-grade governance discipline alongside their deployments look like they got lucky on safety. They did not get lucky. They made a decision when the cost of making it was low. The companies that did not face a compounding problem: the cost of the incident, plus the cost of catching up inside a regulatory environment that was written in response to that incident. As I noted in [Access Is Not Property](https://theforwardview.com/essays/access-is-not-property-frontier-model-revocation-enterprise-risk), physical AI deployments inherit the same permission-stack question as frontier models, with the additional load that the off switch is now connected to something that moves. The 2030 question is whether regulatory frameworks consolidate into a coherent operating envelope or remain fragmented across jurisdictions in a way that constrains deployment geography. Enterprises preparing for both outcomes compound differently from those betting on one. The repricing is already underway. The visitor center language gets written first. The governance layer gets written later, or not at all, until it has to be. --- ## What a Decider Should Do **If you sit on a board.** Add physical AI to the standing AI governance agenda. Ask management for a one-page inventory of any physical AI deployment in production, in pilot, or in active procurement. Each line should carry a deployment risk classification: industrial, public-facing, or safety-critical. If management cannot deliver that inventory in two weeks, the visibility gap is now on the record. **If you run a PE portfolio.** Add physical AI exposure as a named line item in the vendor and operational risk register for every portfolio company in industrials, logistics, transportation, healthcare, and consumer-facing operations. Most current registers contemplate cyber, financial distress, supply chain, and acquisition risk. The failure mode of a robot, an autonomous vehicle, or an embodied agent acting incorrectly in a workplace is not in them. The Tesla Fremont case is what it looks like when that gap closes. **If you are a GC or head of compliance.** Map your current physical AI deployments to the regulatory dates already in motion. Colorado AI Act: June 30, 2026. EU AI Act high-risk obligations: August 2, 2026. OSHA General Duty Clause: already applies. ISO 25785-1 is in draft and will shape the next compliance cycle. The institutional knowledge sits in your existing product safety, occupational health, and trade compliance functions. Those functions understand how to operate inside a framework that treats physical failure as a liability event. Bring them into the AI governance forum before the next board cycle. **If you are a CTO or head of architecture.** Run one process-safety-grade evaluation on a current physical AI deployment in the next 90 days. Layer of Protection Analysis, or a Bow Tie analysis, is the right starting point. These frameworks have been used in regulated industries for forty years. Importing the discipline into AI deployment is the lower-cost path. The Mexico floor was running orchestration software assembled over the years. The question worth asking now is whether that assembly has a documented failure mode. --- ## The Bottom Line Physical AI is not a marketing frame. It is what happens when AI gets a body and a license to act in environments where failure has physical consequences. The factory floor I walked in Mexico is one of those environments. A city street is another. The buildout is underway. Goldman revised its humanoid market projection sixfold. Waymo is at 500,000 paid rides per week. Amazon's 750,000 warehouse robots are the new baseline. The capital is committed, and the cost curve is declining faster than the models assumed. What is not in place yet is the discipline the category requires. Process industries spent the second half of the twentieth century learning what governed autonomous control of a complex physical environment actually demands. That discipline did not come from press releases. It came from incidents that made it expensive to skip. The Tesla Fremont case is not an anomaly. It is the first documented version of a problem that scales with the deployment numbers. The Mexico visitor center had a phrase on the wall. The floor had something more complicated. The category is the easy part. The discipline is the work. --- **P.S.** Pick one physical AI deployment or pilot anywhere in your organization or portfolio. A robot on a production line, an AV trial, a humanoid pilot, an automated inspection system. Ask one question: who at the company is responsible for the failure case if this system acts incorrectly in the next 90 days? If the answer is "the vendor," or "the engineering team," or a name with no defined authority to stop the deployment, you have just found the governance gap. Write it down. That gap has a price in 2029 that it does not have today. Next week's issue picks up that gap and walks through what the process industries already solved, and what it would take to import that discipline into the physical AI stack before the regulatory frameworks make it mandatory.