Join the movement and participate on Uniswap.
Participate
Nils recapped the Humanoid Hub keynote from San Francisco. The thesis is the same one we keep repeating because the industry still doesn't hear it: the next frontier of physical AI is not making the robots smarter. It's making the environments they work in smarter.
70% of the world economy — headcount and GDP — is tied to physical locations and labor. Those places are inaccessible to AI today. Physical AI has a larger TAM than AI in general. It's also harder to start, because the world of atoms has no URL.
Robots need three things: perceive, move, manipulate. Combine perception and locomotion and you get self-driving — we still call that a locomotion job. Combine perception and manipulation and you get laundry-folding — we still call that manipulation. Perception jobs exist too, and they're deceptively valuable.
The industry is funding manipulation. Data for manipulation. Models for manipulation. If you're not in those two buckets, investors don't know what you do.
The frontline needs brains more than it needs muscle.
Budbreak builds vineyard robots. They don't prune, harvest, or spray. They look at grapes. A crop-destroying disease takes over 4% of yield. Computer vision finds it better and cheaper than people. Matching that ROI with harvesting or pruning would mean cutting that labor 20–25% without losing quality. A perception job lifts yield and doesn't take anyone's job.
A well-established Chinese retailer — over 2,000 locations, about 3.5 billion in revenue — spent years taking human decisions off the floor. What to carry, how to price, where to place it: the computer. 24/7 monitoring. This year they published the flagship result: 258% sales lift in a big Beijing supermarket. Also 30% less operational workload, 85% less shrinkage, 26% less energy. That's their case study, not ours. A shelf is a web page that humans still design by hand. Retail is more complicated than chess.
The robotics industry treats robots as execution. A physical business also has to capture data, analyze it, decide, then execute. Physical AI has to be more than a motor cortex.
Together with BracketBot we built a store-manager robot that doesn't touch the products. It inspects tags and facings, feeds the store brain, and the store tells people what to do. We start shipping these to customers in January. Not a store worker. A store manager. Tiny lifts on thousands of SKUs beat automatic restocking.
The biggest tax on the human economy is what we lose by not coordinating. Beijing has more cars on the road than people in Los Angeles. Average commute: two hours. Back of the napkin: the time it took to build the pyramids, lost in Beijing traffic every week. Cheaper fuel doesn't fix that. Coordination does.
GPS is not precise enough for cars to tell each other what they intend. We think digital information anchored to physical space is the language — human to human, human to robot, robot to robot.
AR tasking has already saved customers more than 15 minutes per employee per day. Leave a note in space and a robot or a colleague can see it. Collaborative live mapping is robots agreeing on that language.
Figure can walk into an apartment it's never seen and make the bed. How do you tell it where the clean sheets are? "These new socks are for Maddie, put them in her drawer." "The caramel almond shipment was late — restock them now." You can't generalize that from the model. You have to give it spatially grounded context.
Dead spots, bandwidth that varies by site, latency if the brain is in the cloud, privacy if the data leaves the venue. The industry story is deploy, collect data, improve the model. The defensible thing is supposed to be proprietary data. Customers already know that. If they hand it over, the robot company might become the competitor. Amazon already opened grocery stores. Retail is heading toward a $25 trillion industry around 2030.
Your data flywheel is my moat.
That's why we've signed over a thousand locations. Japan requires camera shutters to make a sound. Europe banned untargeted facial tracking. The Netherlands is moving on camera glasses. Saudi Arabia stops them at the border. The US banned foreign humanoids on privacy and national-security grounds.
Brain on the robot: privacy, but weak compute and battery. Brain in the cloud: scale, but brittle links and a privacy problem. Since 2021 we've said there's a third option: domain-side compute — hyperlocal, hop-on hop-off, plugged into the wall nearby.
On the machine: perception and immediate control, as little compute as possible. In the domain: interpretation, mapping, shared context, shared memory. In the cloud: heavy reasoning that isn't time-sensitive or privacy-sensitive — and we don't think the cloud will be very involved in physical AI.
That grows the internet in three directions: an internet of spaces, of sensors, and of actuators. Co-embodiment: intelligence moving through a graph to borrow sensors and actuators in the domain. We call that the real world web. No GPS on the moon. Same protocol on Earth and beyond.
Open protocols so physical AI can browse, search, and navigate. Intercognitive for the standards. Cactus as the first harness — the store that improves itself. Over a thousand locations signed. We think we deploy over a thousand robots to that base in 2027. Retail spending $125 billion a year on physical AI in five years is half a percent of the predicted TAM. That may be conservative.
Auki is making the physical world accessible to AI by building the real world web: a way for robots and digital devices like smart glasses and phones to browse, navigate, and search physical locations.
70% of the world economy is still tied to physical locations and labor, so making the physical world accessible to AI represents a 3X increase in the TAM of AI in general. Auki's goal is to become the decentralized nervous system of AI in the physical world, providing collaborative spatial reasoning for the next 100bn devices on Earth and beyond.
X | Discord | LinkedIn | YouTube | Whitepaper | auki.com
