The Physical AI Dispatch, Issue #1
The Big Picture
Section titled “The Big Picture”This was the week physical AI stopped being a pitch deck and became an org chart problem.
Google swallowed Intrinsic back into the mothership and declared war on every robotics platform play in existence. BMW, already running Figure humanoids at Spartanburg, picked a completely different vendor (Hexagon’s AEON) for its Leipzig plant. Apptronik closed $520 million with John Deere on the cap table, which means humanoids are headed to cornfields, not just clean rooms. And NTT DATA and Ericsson started building the private 5G backbone that enterprise robot fleets will run on.
All of it points the same direction: physical AI is scaling fast. The infrastructure to manage it at scale isn’t there yet.
This Week in Edge AI for Robotics
Section titled “This Week in Edge AI for Robotics”Google Absorbs Intrinsic. The Platform War Begins.
Alphabet folded Intrinsic, its industrial robotics software company, back into Google on February 25. Five years as an independent “Other Bet,” and now it reports into the same org as DeepMind and Android. CEO Wendy Tan White stays. The Foxconn factory automation JV continues.
Google now has: foundation models (Gemini via DeepMind), simulation and orchestration software (Intrinsic), manufacturing partnerships (Foxconn), and cloud infrastructure. That’s the full Android playbook: open platform, hardware partners build on top, Google owns the ecosystem.
McKinsey pegs general-purpose robotics at $370 billion by 2040. Google is positioning to be the operating system underneath all of it.
Android won not just because of the OS, but because of what got built around it: MDM platforms, enterprise enrollment, OTA update infrastructure. The “Android of robotics” will need all of that too, and Google isn’t building it.
BMW’s Humanoid Strategy Just Got Interesting. Two Vendors. Two Continents.
Figure AI has been the poster child for humanoid deployment. Rightfully so: 30,000 cars produced, 90,000+ parts handled, 11 months of 10-hour production shifts at BMW Spartanburg. Five-millimeter precision. Two-second cycle times. On a live assembly line, not in a lab.
So when BMW announced its European expansion on February 27, you’d expect Figure to be the one crossing the Atlantic. They’re not.
BMW chose Hexagon Robotics’ AEON humanoid for Leipzig. Different robot. Different company. Different architecture (AEON moves on wheels, not legs). BMW also established a new “Center of Competence for Physical AI in Production” to coordinate it all, with the Leipzig pilot targeting high-voltage battery assembly starting April 2026.
One automaker is now running two different humanoid platforms, from two different vendors, across two continents, with different regulatory environments and different factory configurations.
The multi-vendor fleet is not a hypothetical. BMW is living it right now. Who coordinates OTA updates across Figure and AEON? Who monitors performance consistently across Spartanburg and Leipzig? Who handles compliance when German and US safety standards diverge? These are fleet operations problems, and they just became very real.
$520M for Apptronik. But the Investor List Is the Real Story.
Apptronik raised $520 million at a $5.5 billion valuation. Google DeepMind and B Capital were expected. Mercedes-Benz was already in from before.
The new names are the signal: John Deere and Qatar Investment Authority.
John Deere investing in humanoid robotics should change how you think about this market. It means the autonomous humanoid roadmap now includes agriculture, construction, and heavy industry: environments with no Wi-Fi, no clean power, and no controlled conditions. Edge AI isn’t a nice-to-have in a soybean field. It’s the only option.
Qatar Investment Authority (sovereign wealth fund money) signals that Gulf states see humanoid robotics as critical infrastructure, not speculative tech.
The investor list is a deployment map. Mercedes = automotive production lines. John Deere = agricultural and construction sites. QIA = sovereign infrastructure projects. Every one of these environments demands autonomous operation at the edge, fleet-scale monitoring, and operational resilience. These aren’t demo customers. They’re planning for thousands of units.
NTT DATA + Ericsson: Building the Connectivity Layer Nobody’s Talking About
NTT DATA and Ericsson announced a multi-year strategic partnership on February 26 to deliver private 5G as a fully managed service, specifically for enterprise physical AI deployments. The targets: manufacturing, mining, ports, airports, energy.
Everyone in robotics talks about models, chips, and simulation. Almost nobody talks about what connects the robots to each other and to the infrastructure around them. Wi-Fi wasn’t built for 200 autonomous machines moving through a 500,000-square-foot factory at the same time. It drops. It interferes. It doesn’t guarantee the latency you need for real-time fleet coordination.
Private 5G does. And now there’s a managed service to deploy it at enterprise scale.
Connectivity is physical AI’s invisible infrastructure layer: it’s boring until it fails and your entire fleet stops. The fact that NTT DATA and Ericsson are building managed 5G specifically for physical AI means the enterprise deployment stack is being assembled piece by piece: compute (NVIDIA, Qualcomm), connectivity (NTT/Ericsson), models (everyone). That’s real progress.
Embedded World 2026: The Moment Edge AI Became the Main Event
Embedded World has always been an industrial show: microcontrollers, sensors, embedded Linux. This year’s exhibitor lineup tells a different story.
e-con Systems is demoing Darsi Pro, an edge AI compute platform built on NVIDIA Jetson Orin NX, designed specifically for autonomous mobile robots. Full ROS 2 stack, multi-camera support, industrial-grade reliability. Synaptics is showing a Physical AI demo that unifies compute, connectivity, and sensing in a single robotics platform. Infineon is pitching PSOC and AURIX MCUs for deterministic real-time robotic control. Advantech is spotlighting edge HPC platforms for AMR development and logistics automation.
Every one of these companies was an “IoT vendor” two years ago. Now they’re all positioning for physical AI.
When embedded silicon vendors pivot their entire product strategy toward robotics, it’s a platform shift, not a trend. The edge is becoming the robot’s brain, and every chip company wants to be inside it. What none of them are building is the software layer that manages what happens after those chips ship, in production, at scale, across fleets.
The Numbers
Section titled “The Numbers”Edge Hardware Watch
Section titled “Edge Hardware Watch”The latest accelerated ROS 2 packages are live for Jetson AGX Thor. If you’re on the NVIDIA stack, this is your new baseline. Thor is the edge compute target for next-gen humanoids. This release matters. Source: NVIDIA Developer
SECO and Advantech are both showcasing edge AI solutions on Dragonwing (QCS6490, QCS5430, IQ-X) at Embedded World. Two years ago, “Jetson alternative” was a punchline. Now it’s a product category. Source: SECO
NVIDIA’s Blackwell-architecture module delivers 4x the energy efficiency and AI compute of prior gen. The edge compute arms race is accelerating faster than most people realize. Source: NVIDIA Newsroom
Open Source Corner
Section titled “Open Source Corner”NVIDIA’s latest generalist humanoid foundation model is now available via the LeRobot library. Integrates Cosmos Reason for step-by-step task planning. Supports bimanual manipulation on Agibot Genie-1 and Unitree G1. The open-source humanoid AI stack is getting very real, very fast. Source: NVIDIA Developer
AgiBot released SOP (Scalable Online Post-training), the first open framework designed specifically for pushing Vision-Language-Action model updates across entire robot fleets. Fleet-wide model updates sound elegant in a blog post. In practice: rollback, version pinning, canary deployments, audit trails across 200 robots at three sites. It’s an infrastructure nightmare. The fact that AgiBot built this tells you the pain is real. Source: AgiBot Research
The field is maturing. Trends: discrete diffusion architectures, reasoning-enabled models, benchmarks moving beyond simulation into real-world tasks. Two years ago there were maybe a dozen VLA papers at a major conference. The research pipeline is flooding. Source: Moritz Reuss
One More Thing
Section titled “One More Thing”Everyone wants to be the “Android of robotics.” Google is assembling the playbook with Intrinsic and DeepMind and Foxconn. NVIDIA is building it from silicon up. Qualcomm is making its bid with Dragonwing. Three credible companies, three full-stack plays, and that’s before you count the robotics companies building their own software stacks from scratch.
But here’s what none of them are talking about: what happens when a single customer (BMW is already there) has robots from multiple vendors, running different models, across factories in different countries?
The Android analogy is actually apt, just not in the way these companies intend. Android won the smartphone war. But the companies that made the real money weren’t just the ones who built the OS or the chip. They were the ones who built the management layer on top: MDM platforms, enterprise deployment tools, the security infrastructure that made it possible to run 50,000 devices across an organization without losing your mind.
Physical AI is heading for exactly the same moment. And the management layer doesn’t exist yet.