Welcome back to Level.UP, brought to you by UP.Labs.

This week, we dive into two of the week’s big stories. 

The first: NVIDIA introduced the first full-stack safety system for robots, and in doing so, moved to define the layer every robotics vendor and every operator will eventually have to build around. 

The second: OpenAI unveiled its first in-house chip, a move that says less about silicon than about where the real constraint on AI has gone: into power, land, and the physical build-out that operators like you already run.

Plus: BMW puts humanoid robots on the factory line, a $7B bet on the chips inside the machine, and more from the world of physical AI.

Think someone else needs this? Forward it to a friend or colleague navigating the same terrain.

On June 22, at the Automate conference in Chicago, NVIDIA announced Halos for Robotics, billing it as the industry's first full-stack safety system for physical AI. The move ports NVIDIA’s safety architecture for autonomous vehicles — what the company describes as more than 18,600 engineering-years of work — into a single stack for robots.

That stack spans three layers: compute (the IGX Thor platform), software (Halos OS and Halos Core), and a third-party certification path run through NVIDIA's new Halos AI Systems Inspection Lab, which holds ANAB accreditation and is recognized by TÜV Rheinland, UL Solutions, TÜV SÜD, SGS, exida, and CertX.

Agility Robotics is the first adopter, integrating IGX Thor and Halos Core into the human-detection system on its Digit humanoid — a robot NVIDIA says is already working with operators, including Amazon, GXO, the Schaeffler Group, and Toyota Motor Manufacturing Canada. NVIDIA also released an "Outside-In" safety blueprint that extends a robot's perception using external cameras and AI agents that can override the robot's behavior. The signal: in robotics, "Is it certified?" is about to replace "Does it work?"

First, NVIDIA isn't just selling chips here. By bundling compute, software, and the certification lab into one stack, it is positioning to define the safety layer every robotics vendor builds against — the same way it came to set the terms in autonomous vehicles.

For an operator, that turns a technical choice into a strategic one: standardize now on an emerging-standard safety architecture and accept the lock-in, or hold out for an open alternative that may or may not arrive. It's the same fork we keep returning to in these pages, now reaching the factory floor.

Second, and more useful: notice how NVIDIA got here. It took nearly two decades of hard-won autonomous-vehicle safety work and ported it into a new form factor, skipping years of groundwork. That is the real lesson for physical-industry operators, and it cuts in your favor. The validated safety knowledge, the edge cases, the operational discipline you've accumulated over decades — those assets transfer into adjacent, AI-native markets and compounds.

Carry your expertise across the gap into a new domain. That cognitive leap — seeing what you already own as the raw material for the next market — is exactly the leverage most operators hold and underuse.

So watch the certification path, not the capability demo. It's what your insurer, your safety officer, and your next RFP will actually gate the deployment on.

OpenAI and Broadcom have unveiled Jalapeño, OpenAI's first in-house chip.

OpenAI says it went from concept to tape-out in nine months — what the company believes is the fastest advanced semiconductor ASIC cycle ever — and that it used its own AI models to accelerate the design.

The chip's whole point is performance per watt: early testing shows roughly 50% cost savings compared to typical AI GPUs. It handles inference only; NVIDIA still powers the vast majority of OpenAI's work, especially training. Deployment begins by the end of 2026, with Microsoft expected to take about 40% of the initial run.

Getting Jalapeño into production means power, land, cooling, grid interconnects, electrical and thermal engineering, and the discipline to build it all on schedule. That is precisely the capability the frontier labs do not have in-house, and precisely what sits within operations like yours.

A few weeks ago, we said that physical-industry operators are an input to the AI economy, not just customers of it. Here's that idea with a price tag attached.

Ten gigawatts is about the power draw of 7.5 million homes. OpenAI can design a chip in nine months; it cannot conjure that much power, land, cooling, and grid connection on the same clock. None of that lives in a model. It lives in substations, water rights, poured concrete, and people who can bring an industrial site online without blowing the schedule.

So the relationship is quietly flipping. The most valuable AI company in the world is increasingly shopping for the things you already build and run. The operators who've noticed are at the table as suppliers of a scarce resource, not as buyers waiting for the tech to mature.

Worth sitting with: in this build-out, your stack of physical assets may be worth more to the labs than any model you could rent back from them.

Ready to work smarter? Here are the tools we're tracking this week:

Chipmakers Are Buying Their Way Into Physical AI. On June 25, ON Semiconductor agreed to buy Synaptics for about $7B. Robots, cars, and factory machines are starting to do their thinking inside the device instead of sending everything to the cloud, and that requires specialized chips working together to sense the world, figure out what's happening, and act on it. ON makes some of those chips; Synaptics makes the rest. Buying it lets ON sell the whole set instead of one piece, and it expects that to open up about $30B in new business by 2030. The signal: in physical AI, a lot of the real value sits in the small, unglamorous chips buried inside the machine — and the big chipmakers are racing to own them. → Read more

BMW Puts Humanoid Robots On The Real Line. The carmaker said it's rolling out Figure's newest humanoid robot, the Figure 03, at its Spartanburg plant to do real work: pulling loose parts from bins and sorting them into carts in the exact order the assembly line needs them. It builds on a 2025 trial in which an earlier model worked for about 11 months and helped produce more than 30,000 X3s. The tasks BMW is handing over are the repetitive, physically punishing ones it would rather not wear its people out on. The read: humanoid robots are quickly moving from lab demos into ordinary factory jobs, and much of that progress is happening within manufacturers that already run complex lines and can train robots on the floor. → Read more

China Uses Physics To Teach Robots. Shanghai startup Fysics AI, founded by a former NVIDIA manager, launched "Fysiverse" — a virtual world for training robots and self-driving cars. The twist is in how it works. Most systems (OpenAI's and Meta's included) learn how the physical world behaves by watching enormous amounts of video and inferring the rules. Fysics builds the actual laws of physics — gravity, friction, momentum — directly into the software, so the simulation can't drift into things that couldn't really happen. The tell: the simulator robots practice in is becoming a contest of its own, and the next edge may come from getting the physics right, not just feeding the model more data. → Read more