For two years, the physical AI story has been about the brain: bigger models, better world simulation, humanoids learning to walk. This edition is about everything the brain still can't do on its own.

The labs can generate intelligence. What they can't generate is the manufacturing depth to put it in a body, or the interface to let a worker talk to it without stopping the job.

Both live inside operations like yours.

This week, we look at why that makes physical-industry operators an input to the AI economy rather than just a customer, and why the next moat may be the channel between your people and the autonomous systems already running around them.

Plus: more tools we're using internally, the startup cleaning homes to harvest training data, the saturation of "tokenmaxxing," and NVIDIA’s $6.5B bet on photonics.

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

Caitlin Kalinowski spent 15 years building physical products. She was a technical lead on the MacBook Air at Apple and ran the hardware teams behind Meta's Quest, Rift, and Orion AR programs. In November 2024, she joined OpenAI to build its robotics and hardware teams from scratch, leaving in March 2026.

In a recent interview on Lenny’s Podcast, Kalinowski laid out why she believes digital AI is approaching saturation, and why the next frontier is physical. She also named what's actually blocking it. Humanoid robots aren't held back by software. The gating constraints are actuators, thermal envelopes, battery density, and the absence of a manufacturing base capable of shipping them at consumer-grade reliability and volume. 

If you run a physical-industry operation, you’re not just a customer to the AI economy. You’re a critical input to it.

The labs can generate intelligence. What they cannot generate is the manufacturing depth, supplier relationships, thermal engineering expertise, and production-line discipline required to put that intelligence into a body and ship it at volume. That capability lives inside companies like yours; you hold the data that AI needs to work.

Physical industries have accumulated data across decades of production, from failure rates to supplier performance, line throughput to field telemetry. That data does two things for you: it tells you which adjacent markets you're already positioned to enter, and it gives you the leverage to set terms with frontier labs rather than accept theirs.

There's a form-factor opportunity that most operators haven't yet priced in. The lines you run today — built for vehicles, appliances, complex assemblies — were optimized for durable goods while consumer electronics moved to Asia. That specialization is now a strategic asset. Those same lines can be adapted to manufacture new physical AI outputs, or made dramatically more efficient by embedding AI into the production process itself. Both paths extend an advantage you already own.

Ultimately, the question isn't "How do we catch up to OpenAI?" But rather, "What part of our operational depth becomes a moat once the intelligence layer is a commodity input anyone can rent?"

A field technician on a wind turbine, harness clipped and both hands on a wrench, needs to send a command to the diagnostic unit at their belt. A worker on a loading dock, gloved, eyes on the pallet, needs to redirect a connected lift. Neither moment calls for a smarter robot. Each calls for a faster way to be heard by the machines already running around them.

For two years, the physical AI conversation has centered on one side of that loop: make the machine better. World models, dark factories, humanoids closing the sim-to-real gap. Meanwhile, the channel humans use to talk back has barely moved since 1984. Screens, buttons, voice. All three assume the worker can stop, look down, and translate intent into a structured command. On a turbine, a dock, or a hospital floor, that assumption breaks.

A new category of companies is building for the gap. 

Austin-based Wetour Robotics fuses wrist-worn biosignals, spatial position, and computer vision into a single real-time command, with inference on-device and full-chain latency under 100 milliseconds. The bet underneath it is architectural: make the human body a node in the computing network, contributing intent at the same speed any connected device already does.

For most of the last few years, the race has been about giving robots better eyes, hands, and a brain. The harder problem, though, and the one with more durable enterprise value, is giving the human a low-latency way to participate in systems that increasingly run on their own. You cannot augment a field technician if your only input channel is a screen they can't look at.

The cost of that mismatch is already on the P&L. It just doesn't have a line item. It hides in training time, error rates, and the productivity lost whenever a worker breaks from a physical task to operate a system. Most operators are paying it without ever naming it. Augmentation, the thesis we keep coming back to, depends on a working channel between the worker and the system. Right now, that channel barely exists.

This is also where augmented reality finally earns its keep. AR has been pitched as an operations and safety tool for a decade, with little success, because it solved only the output half of the problem and ignored the input half. An overlay that shows machine status is useful. But an overlay that also reads a worker's gesture, position, and gaze and feeds a command back without breaking the task is a closed loop. 

Again, the deeper prize here is the data. Every natural exchange between a worker and an autonomous system is a training signal for embodied AI, and almost all of those exchanges are invisible to any computer today. Whoever owns the interface captures them, structured and in the wild. That company won't just sell better wearables. It will own the data that trains everyone else's robots.

We're an AI-powered company that builds AI-powered companies, so we often use ourselves as the first test case. Lately, that's meant building our own agentic tools to clear the busywork so our engineers can spend their time architecting the AI systems our clients actually need.

This week, our CTO, Rob Patrick, breaks down a few tools we’re running internally with UP.Code, our engineering team of developers across Santa Monica, San Francisco, and remote. They build the corporate software that runs UP.Labs and powers the work inside our portfolio companies.

Free Cleaning, For Training Data. AI startup Shift will clean your home at no charge, in exchange for recording the work and using the footage to train robots. The company is already paying tens of thousands of people across 15 countries to record their tasks, and cleaning is just the start. Next on their roadmap is plumbing, cooking, and even building. The pattern: the Web 2.0 subsidy was "lose money to acquire users." The AI-era version is quickly becoming "lose money to acquire data." → Read more

“Tokenmaxxing” Is Reaching A Saturation Point. In May, Amazon shut down "KiroRank," an internal leaderboard ranking staff by AI token usage. Microsoft is pulling most of its Claude Code licenses by the end of June. And Uber's COO, Andrew Macdonald, said the company burned its entire 2026 AI coding budget in four months, without a correlation in features shipped. The thread: agentic AI is breaking the cost curve, and companies are starting to recognize it. The metric that survives remains the same: output, not activity. → Read more

NVIDIA Is Betting $6.5B On Light. Since March, NVIDIA has poured at least $6.5B into photonics — moving data with light instead of electricity over copper. The catch is timing. The parts are difficult to manufacture, and analysts don't expect them to be produced at scale before 2028. The signal: the bottleneck has shifted from the chips to the wiring around them — and, like this edition's main stories, the build-out, not the model, is the hard part. → Read more