Roomba: Hitting the Wall | The New Space Race | 3
How do Chinese AI companies approach model distribution differently than U.S. companies?
Chinese AI companies release open-weight models that anyone can download freely , whereas U.S. companies restrict access through paid APIs. This fundamentally different distribution model allows Chinese firms to compete strategically by leveraging resources differently rather than trying to match U.S. companies' supercomputer budgets head-on.
The strategy reflects a pragmatic recognition of resource constraints. Chinese AI companies acknowledge they may not be able to compete directly with U.S. giants that have seemingly unlimited funding for massive computing infrastructure. By open-sourcing their models, they democratize AI access and build competitive advantage through volume and ecosystem participation rather than scarcity.
As discussed in this episode of Business Wars , this approach mirrors broader patterns in how different regions pursue technological leadership. U.S. companies guard their models as proprietary products, creating walled gardens that generate recurring revenue. Chinese companies instead share their AI with the world, betting that open distribution creates network effects and long-term competitive positioning.
The implications extend beyond AI distribution alone. This model reveals how Chinese companies are "working smarter" by recognizing where they cannot simply outspend competitors and instead building alternative pathways to market dominance. Free, downloadable models lower barriers to adoption globally and enable developers everywhere to build on Chinese AI infrastructure.
Christopher Mims — Tech Reporter, Wall Street Journal. Mims covers innovation, robotics, and artificial intelligence across multiple continents. He is the author of "How to AI: Cut Through the Hype, Master the Basics, Transform Your Work" and "Arriving Today: From Factory to Front Door, Why Everything Has Changed About How and What We Buy." His reporting spans the competitive landscape between U.S. and Chinese technology development, including in-depth coverage of the robotics industry and how companies like iRobot navigate global markets.
Open-weight models versus closed APIs: a strategic divergence
The technical difference is straightforward: open-weight models are fully downloadable, allowing users to run them locally without paying per query. U.S. companies' API model charges users every time they call the service, creating predictable revenue but also friction for adoption.
This distribution choice is not merely tactical—it reflects fundamentally different assumptions about competition. Chinese companies are betting that free distribution builds adoption faster than paid APIs, especially in developing markets where price sensitivity is high. They prioritize market penetration and ecosystem control over immediate monetization.
For developers and enterprises, the practical impact is significant. A team using a Chinese open-weight model incurs hosting and compute costs but no licensing fees. The same team using a U.S.-based API pays per transaction, which can become expensive at scale. This cost structure favors Chinese models in price-sensitive regions and among developers building on constrained budgets, as detailed in the full Business Wars conversation .
What is Moravec's paradox and how does it relate to current advances in robotics and AI?
Moravec's paradox states that it is easier to teach a humanoid robot to beat a human at chess than to physically pick up and move the pieces on a chessboard. For 50 to 100 years this counterintuitive reality has held true, but new kinds of AI—different from today's chatbots—are finally enabling robots to move and interact with the physical world in ways that were previously impossible.
Understanding Moravec's Paradox
Moravec's paradox is the observation that high-level reasoning tasks (like chess strategy) are computationally easier for machines to master than low-level sensorimotor skills (like grasping and manipulating physical objects). This reversal of what humans find intuitive has shaped robotics research for decades, leading engineers to focus on abstract problem-solving while struggling with basic physical interaction.
Why the physical world remains the hardest frontier
The reason behind this paradox runs deep: abstract thinking evolved recently in humans, while physical dexterity developed over millions of years of evolution. Chess rules are clean, digital, and deterministic—a computer can evaluate millions of positions instantly. But picking up a chess piece requires understanding weight distribution, surface friction, three-dimensional space, and countless unpredictable variables that human brains solve unconsciously every second.
As Christopher Mims explains in Business Wars , the real world was built for human hands and human perception. Manufacturing environments—where robots have thrived—are carefully controlled: flat surfaces, repetitive motions, predictable sequences. But the moment you move a robot into an unstructured space or ask it to handle soft, irregular objects, the difficulty spikes dramatically.
New AI breaking through the 50-year wall
What has changed recently is the emergence of new forms of artificial intelligence distinct from large language models like ChatGPT. These newer approaches provide robots with the flexibility and adaptability needed to handle unpredictable physical scenarios. Rather than relying solely on pre-programmed rules or rigid machine learning models, these systems can learn dynamic behavior and adjust to novel situations—something chatbots, for all their linguistic prowess, cannot do.
This represents a genuine inflection point in robotics. For half a century, Moravec's paradox seemed like an immutable law of physics. Now, as discussed in the episode , the gap is finally closing. Humanoid robots are beginning to acquire the kind of tactile intelligence and real-world agility that has eluded the field since its inception.
"The world is already made for humans. So if you want to automate what people do in the world that we have now, you need something that is roughly shaped like a human that has the same reach as a human."
Christopher Mims — Tech Reporter at the Wall Street Journal and author of "How to AI: Cut Through the Hype, Master the Basics, Transform Your Work" and "Arriving Today: From Factory to Front Door, Why Everything Has Changed About How and What We Buy." Mims has spent time with Colin Angle, the founder of iRobot, and closely observed the collision between Chinese and American robotics innovation, particularly following iRobot's bankruptcy in December 2025.
To fully understand how roboticists are solving this challenge and why companies like Aptronic and others are rethinking humanoid design from the ground up, listen to the full episode where Mims dives deeper into the industrial automation race between the U.S. and China.
How does Aptronic's approach to humanoid robot development differ from traditional methods?
Aptronic, based in Austin, Texas, is building the cheapest humanoid robot body by intentionally keeping it "mindless" —delegating artificial intelligence entirely to partners like Google's DeepMind. Rather than bundling hardware and brain together as most robot makers do, Aptronic focuses purely on the physical form and leaves the cognitive layer to AI specialists.
This split approach represents a fundamental departure from the integrated model that has dominated robotics for decades. Traditionally, companies developing humanoid robots attempt to control both the mechanical engineering and the AI stack in-house, treating them as a single product. Aptronic's division of labor treats the body and the brain as separate domains , with each partner optimizing for what it does best.
The philosophy behind this strategy is practical: a robot body exists to move and interact with a physical world designed for humans. As explored in detail in the Business Wars episode , the humanoid form factor itself solves a critical architectural problem. By leaving AI development to specialized firms like DeepMind—which are already building novel AI systems to drive robot bodies—Aptronic avoids duplicating expertise and can concentrate on cost reduction and manufacturing efficiency for the hardware layer.
"The world is already made for humans. So if you want to automate what people do in the world that we have now, you need something that is roughly shaped like a human that has the same reach as a human."
Christopher Mims — Tech Reporter, Wall Street Journal. Mims covers emerging technologies and has authored two books on AI and manufacturing: How to AI: Cut Through the Hype, Master the Basics, Transform Your Work and Arriving Today: From Factory to Front Door, Why Everything Has Changed About How and What We Buy . He has studied robotics innovation globally and spent time with industry figures including Colin Angle, founder of iRobot.
The financial incentive for this modular approach is compelling. Building a cheaper robot body expands the addressable market for humanoid automation, while leaving AI development to competitors who have already invested billions creates a leaner cost structure. This is not a weakness but a strategic advantage: Aptronic can iterate faster on hardware without getting bogged down in AI research timelines.
What makes Aptronic's bet particularly bold is the implicit faith in the AI partner ecosystem. Rather than owning the entire stack, the company is betting that as discussed in the podcast , breakthroughs in AI—whether from DeepMind or other research labs—will flow into robot bodies regardless of manufacturer. This stands in sharp contrast to older robotics ventures that treated proprietary AI as a defensible moat. The risk is dependency; the reward is speed to market and lower capital burn.
A Deliberate Rejection of Vertical Integration
Most humanoid robot startups have tried to own as much of the technology stack as possible. Aptronic's founders have instead chosen radical focus: build the best body cheaply, then plug in whatever brain the customer wants . This is closer to the smartphone model—where silicon design, OS development, and assembly are all separate industries—than to legacy robotics ventures that attempted to be vertically integrated from concept to deployment.
The partnership with Google's DeepMind signals confidence in the open-architecture approach. DeepMind specializes in developing novel AI systems, not manufacturing hardware. By joining forces, both companies avoid competing in domains where the other has no edge. This clarity of roles is unusual in the robotics industry and reflects a more mature understanding of where value actually lies. You can hear more about how this shift is reshaping competition in the full episode .
Robots excel at repetitive tasks on rigid surfaces, but laundry presents a fundamentally different challenge: an average T-shirt is a wobbly noodle that requires real hard thinking about where to grab it.
Why is a humanoid robot design preferred over specialized robots for specific tasks?
The world is built for human dimensions , so a humanoid robot can operate anywhere humans do—flipping burgers in restaurants, working in warehouses, adapting to new environments. A specialized robot does one job perfectly but only one, while a humanoid design sells to countless customers across industries.
One shape fits all: the economics of human-shaped automation
The logic is straightforward: existing spaces and tools are designed around human proportions and reach . A restaurant kitchen has door frames, work surfaces, and equipment configured for human bodies. An Amazon warehouse has ramps, aisles, and shelving systems that assume human-scale movement. Build a robot that matches those dimensions, and it can operate in any of these environments without modification.
A task-specific robot—say, one engineered solely for burger flipping—becomes economically viable only if enough burger restaurants buy it to justify the engineering cost. In reality, as explained in Business Wars , the market is fragmented across thousands of different kitchens, suppliers, and operations, each with unique layouts and quirks.
By contrast, a humanoid design can address multiple customer segments at once . The same robot that works the evening shift at a restaurant can be deployed to a warehouse during peak season, then retrained for packaging, assembly, or other labor-intensive tasks. This versatility dramatically widens the addressable market and accelerates the payback period on R&D investment.
Flexibility as a competitive moat
Specialized robotics excel in controlled, repetitive environments—car factories have proven this for decades. But beyond the factory floor, the world is messy and variable. A humanoid robot's ability to adapt to unpredictable layouts and changing priorities without hardware redesign is a strategic advantage.
As Christopher Mims discusses in the episode , this is precisely why companies across sectors from retail to logistics are betting on general-purpose humanoids instead of narrow-use robots. The investment pays off once adoption reaches scale.
"The world is already made for humans. So if you want to automate what people do in the world that we have now, you need something that is roughly shaped like a human that has the same reach as a human."
Christopher Mims — Tech Reporter, Wall Street Journal. Mims covers robotics, artificial intelligence, and global manufacturing trends, drawing on interviews with industry founders like Colin Angle of iRobot and observations of technological development across North America, Europe, and China.
The implication is larger than robotics alone. When you remove the constraint of matching human form, you also remove the ability to leverage existing infrastructure—doors, stairs, tools, workstations. Specialization gains precision but loses universality. Humanoid design trades some efficiency in any single task for the ability to operate profitably across dozens of customer segments .
This explains why despite decades of industrial robotics success, the recent wave of investment—from Tesla, Boston Dynamics, and others—centers on machines with heads, torsos, and limbs, not more optimized factory arms. The market has shifted from scarcity (anything robotic is valuable) to abundance (robots that don't fit your space are worthless).
The real driver, as detailed further in this podcast discussion on innovation and manufacturing strategy , is the economics of scale. One design, one supply chain, thousands of buyers across different industries—that is the formula for robotics profitability in an era when labor costs are high but deployment scenarios are diverse.
What institutional knowledge is lost when manufacturing is outsourced away from domestic designers and engineers?
When designers and engineers lose direct access to the manufacturing process, a critical feedback loop collapses. Manufacturers who gain control of production gain direct access to all intellectual property , enabling them to reverse-engineer your product and create cheaper versions without the innovation investment you made.
The Broken Feedback Loop
Domestic manufacturing creates a continuous exchange between the design team and the factory floor. Engineers can walk onto the production line, observe problems in real time, and immediately adjust their designs. This tight coupling between thought and execution accelerates learning and prevents costly mistakes from reaching scale. When that relationship exists, manufacturers have an incentive to improve your product because their success depends on your success.
Once manufacturing moves overseas—to China or elsewhere—that daily interaction evaporates. The designers stay home while the makers operate on another continent, separated by time zones, language barriers, and contractual distance, as explored in detail in this Business Wars episode about Roomba's struggle with Chinese competition .
When Suppliers Become Competitors
The real danger arrives when your manufacturing partner sees the entire blueprint of your innovation. Possession of full intellectual property is the gateway to commodification : they understand exactly how you built the product, what makes it work, and what costs you incurred. With that knowledge, they can manufacture a functionally similar device at a fraction of the price, undercutting you in markets where you once led.
This is not speculation—it happened directly in the case of iRobot and its relationship with Picea Robotics , where the manufacturing advantage eventually became the competitive advantage. The supplier who once took orders became the designer of its own product line.
Christopher Mims — Tech Reporter at the Wall Street Journal and author of "Arriving Today: From Factory to Front Door, Why Everything Has Changed About How and What We Buy." Mims has spent time with iRobot founder Colin Angle and has reported extensively on robotics innovation and the competition between domestic and international manufacturing strategies.
The institutional knowledge loss is not just about losing access to cheaper labor. You lose the speed of innovation when your feedback loop stretches across oceans, and you lose proprietary control when your supplier becomes fluent in your technology. The Roomba example shows this starkly: once Picea Robotics acquired iRobot in December 2025 following iRobot's bankruptcy, the supplier that once manufactured the product now owns the entire brand—a complete inversion of the original relationship that this Business Wars investigation explores in depth .
What role did government loans play in enabling Tesla's domestic manufacturing success?
Tesla would not be here today without multi-hundred million dollar loans from the U.S. Department of Energy, which wanted to support the manufacturer domestically of electric vehicles. This is called industrial policy, and China is great at it with its five-year plans that explicitly state what the country will prioritize and pour money into.
What does the future moonshot in the AI and robotics race look like according to technology experts?
Landing on the moon in this technological race means the first time you walk into a workplace, a restaurant, an office, and there is a robot standing there doing a job and no one finds it remarkable. This would represent the normalization of robots in everyday human spaces.