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

This week, we dive into two stories. The first: Harvard Business School researchers priced 1.8M patents, and we distilled the takeaways to help you optimize your AI spending. The second: in the race for robotics, Google’s new model has extended from tabletop tasks to whole-body control 

Plus: SpaceX’s new phone network for robots, and $30M for a data-recording head rig. 

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

Harvard Business School researchers examined 1.8M US patents and identified the ones filed by “AI integrators” — companies with no AI research operation — building on top of AlexNet, the 2012 deep learning breakthrough. Those patents carried a 6.9% value premium over the same companies’ non-AI inventions.

Notably, the examples are not just software firms. 

Gains followed high-value AI patents in gross margin, market share, and return on sales in later years. And it grew fastest at firms whose day-to-day work was well suited to AI in the first place.

Capital markets are pricing application, not just invention. 

For operators in the physical AI industry, the useful question isn't how much to spend on AI. It's whether your spending produces something you still own after your vendor contract ends.

The two examples in the study both cleared that bar, and neither involved building a model. Chevron pointed an existing technique at reservoir behavior. Johnson Controls filed on a maintenance method. What each patented was an operating approach — the intersection of a general-purpose tool and a problem only they understood well enough to specify. 

That intersection is where the premium sits, and physical industries hold more of it than anyone — deep process knowledge and terabytes of data that are only increasing in value today. 

The practical consequence is a change in how AI gets underwritten. A pilot justified on labor hours saved produces a savings number and nothing else. The same budget aimed at instrumenting a process — making a line's actual behavior legible, structured, and yours — produces an asset that compounds and can be filed on. 

Both look like AI spend in the budget. Only one shows up in margin and shares two years later.

You can apply this practically in your next portfolio review: for each line item, name the thing that remains if the vendor disappears. If the answer is a dashboard, it's a cost program. If it's a dataset, a spec, or a filing, it's an investment.

Google DeepMind released Gemini Robotics 2 on July 30, extending its robotics models from tabletop tasks to whole-body control. Earlier versions assumed the robot stood still — arms and hands only — with work brought to the machine. This model enables robots to walk to the work, crouch, reach a bottom shelf, and maintain balance while doing it.

The same trained model ran three different machines: Apptronik’s Apollo 2 humanoid with two different sets of hands, and a Franka Duo, a stationary two-armed unit with a simple gripper. 

Normally, each robot needs its own version, retrained on its own hardware. A companion model that runs onboard, without a network connection, adapts to an unfamiliar robot in a few hours on fewer than 200 demonstrations.

Here are its reported success rates, by task:

The release lands in a crowded position.

Figure AI raised over $1B at a $39B valuation and has humanoids billing hourly at BMW. Nvidia's GR00T is the open humanoid foundation model, while Physical Intelligence and Skild sell hardware-agnostic brains built to run across many bodies. 

Note that DeepMind is playing the model layer without owning a body, which it buys from Boston Dynamics.

What that crowd is actually competing over has shifted. Walking on flat ground is solved, and agile movement (recovery, rough terrain) is close. What remains unsolved is coordinating movement and dexterity in new, unfamiliar environments. 

Which reframes Figure's BMW strategy. It's a data position — hours of a model failing and recovering inside a live production environment, which no simulator generates, and no lab reproduces. Apptronik had to build its own dedicated facility, aptly named Robot Park, to manufacture a thin substitute for it.

As the race for real-world data intensifies, the scarce input in physical AI is increasingly in your hands — permission to operate somewhere that matters. Every operator in a physical industry holds some of that permission.

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

$30M To Record People For Robots To Copy. Singapore's Ropedia raised two pre-A rounds on July 23 for HOMIE, a head-mounted rig that records first-person video, depth, hand tracking, gaze, and body motion on a single clock. The takeaway: robot training data barely exists, so the scarce input is footage of people doing the work. Before a vendor instruments your floor, settle who owns what the cameras record. → Read more

SpaceX Is Building A Phone Network For Robots. On August 4, SpaceX detailed a 2027 satellite service that reaches devices directly. Instead of towers, small base stations would hang on existing Starlink dish mounts. The signal: a phone in a city has three carriers; a robot in a mine or a rail corridor has none. The bid is to sell coverage where nobody built any — which is where most autonomy plans stop. → Read more

Boeing To Sell Autonomy Programs (But Keep The Keys). On August 10, Archer Aviation announced plans to acquire Wisk Aero, Insitu, and SkyGrid from Boeing, folding roughly 2M combined flight hours, 3,500 fielded unmanned aircraft across 35 countries, and a $200M-a-year defense business into its ZEE aerospace model. Boeing takes an equity stake and keeps access to Wisk's autonomous flight tech for its own aircraft. The read: an autonomy program is worth what it has flown, and Archer skipped a decade by buying someone else's log book. Ask any vendor pitching autonomy for the fielded record behind it before you look at the demo. → Read more