Welcome to the rabbit hole.
If you’re here, you already suspect that robotics isn’t just another engineering field. It’s the physical embodiment of intelligence itself, the moment when bits escape the screen and start moving atoms at scale. Once that happens, everything changes that used to take centuries will take years, and changes that used to take years will take weeks. This feed is my attempt to map that phase transition in real time, one post at a time.
Let’s start with the single most under-appreciated fact in modern technology:
The cost of robotic actuation (moving real stuff in the real world) has been collapsing faster than almost anyone in the commentariat has noticed.
In 2010, a high-end industrial robotic arm with a 10 kg payload cost roughly $150,000–$200,000, plus another $100k+ for integration, safety fencing, and custom programming. By 2025, a Chinese humanoid (e.g., Unitree H1 or Figure 01 with comparable specs) can be had for under $90,000 in small batches, and Boston Dynamics Atlas (the version doing parkour in 2024 videos) is reportedly being offered to select partners for under $200k in volume. Meanwhile, the new wave of $20,000–$30,000 class humanoids (Agility Robotics Digit with the new hands, Tesla Optimus Gen 2 projected pricing, Sanctuary Phoenix, etc.) are already in pilot deployments.
That’s a ~10× price drop in 15 years for something that used to be the most expensive capital good on earth outside of jet engines and MRI machines. For context, solar panels only fell ~8× in the same period, and lithium-ion batteries ~15× since 2010. Robotics is matching or beating the canonical “hard-tech” cost curves we usually reserve for semiconductors and renewables.
But price is only half the story. The other half is dexterity and intelligence.
Dexterity
Watch any 2025 clip of Figure 01 walking through a BMW factory, picking up a battery tray, inserting it with sub-millimeter precision, then walking away. Or Tesla Optimus folding a T-shirt (slowly, yes, but perfectly) using nothing but vision and its own two hands. These are not cherry-picked demos on a static table anymore; they’re happening on factory floors with vibration, lighting changes, and unexpected objects. The new hands (Apptronik, Google DeepMind’s ALOHA lab, Tesla Gen 2) have 20+ degrees of freedom, tactile arrays with thousands of sensing points, and can exert anything from 0.1 N (to pick up an egg) to 100+ N (to tighten a bolt). We have crossed the threshold where robotic hands are no longer look like clumsy claws; they look disturbingly human.
Intelligence
The foundation models that crushed language (GPT-4o, Claude 3.5, Grok 2) and image/video generation (Sora, Veo 2) are now being poured straight into embodiment. The training loop is simple but brutal:
- Throw millions of hours of teleoperated or simulated data at a giant transformer.
- Let it learn physics, affordances, and task decomposition implicitly.
- Fine-tune on a small amount of real-world data (often just dozens of demonstrations).
- Deploy and continue learning from every mistake in the fleet.
This is exactly how Waymo cracked self-driving: fleet learning at scale. The difference is that a car has maybe 15 degrees of freedom and a very constrained environment (roads, lanes, traffic rules). A humanoid has 30–40+ degrees of freedom and is expected to eventually operate in kitchens, construction sites, hospitals, and your living room. The data advantage compounds much faster.
The result? Companies are reporting 10–100× reductions in the number of human demonstrations needed to teach a new task compared to 2022-era methods. Sanctuary AI claimed in mid-2025 that their Phoenix system can learn a completely novel manipulation task (e.g., “insert this specific USB-C cable into this specific laptop port you’ve never seen before”) in under 20 demonstrations, sometimes as few as five. That is not incremental; that is the slope going vertical.
The Economic Phase Change
Labor is Labor is ~50–70% of global GDP once you include services. If even 20% of human labor becomes automatable by general-purpose humanoids at $25k–$50k purchase price and $5–$10/hour operating cost (electricity + maintenance + cloud inference), the capital expenditure pays for itself in months, not decades.
At that point you are no longer in the realm of “productivity enhancement.” You are in the realm of being able to spin up entire virtual workforces overnight. Need 10,000 workers for disaster relief? Lease them. Need 100,000 seasonal pickers for harvest? Buy them outright and store them in a warehouse the rest of the year. Need caregivers for an aging population that is already causing societal collapse in Japan, Korea, Italy, and soon China? Deploy them to every home.
This isn’t 2050 sci-fi. The first humanoid factories are breaking ground in 2025–2026 (Tesla Texas, Figure South Carolina, Agility Oregon). The BOM (bill of materials) for Optimus Gen 2 is reportedly already under $20k at scale thanks to Tesla’s vertical integration of batteries, motors, and casting. Figure is raising at a $2.5B valuation on track for similar numbers. The moment one credible player ships 100,000 units at $30k all-in, the race becomes “who can build the most actuators and hands the fastest,” because the neural net part will be largely solved and copyable.
The Counter-Arguments (and Why They’re Weakening)
“Humanoids are unnecessary; just redesign the world for specialized robots.”
That worked for factories (conveyor belts, fixtures, jigs), but ~80% of labor is in unstructured environments: homes, restaurants, retail, construction, healthcare. Redesigning all of those is more expensive than building general robots.
“Data bottleneck; you’ll never get enough human demonstrations.”
VR teleoperation + simulation + foundation-model priors have already broken the bottleneck. 1X’s EVE robots are operated by humans in VR headsets from the Philippines at $5/hour, generating perfect training data 24/7. Simulation is catching up fast (NVIDIA Isaac Sim, MuJoCo updates, DeepMind’s Genesis). The data flywheel is spinning.
“Safety and regulation will stop it.”
Self-driving cars were supposed to be banned too. Instead we got a patchwork: California slow, Texas/China fast. Humanoids will follow the same script: pilot programs in controlled environments (factories, warehouses) → gradual expansion into public spaces → eventual normalization. China has already declared humanoid robots a national strategic priority and is subsidizing heavily.
Where We Are in December 2025
- At least six companies have humanoids that can walk reliably, manipulate objects with five-fingered hands, and learn new tasks from <50 demonstrations.
- At least three (Tesla, Figure, BMW + Figure) have announced factory pilot lines starting 2026.
- The actuator supply chain (harmonic drives, frameless motors, linear actuators) is the new GPU: every robotics startup is hoarding inventory and signing 5-year contracts.
- Foundation model labs have quietly pivoted: the next frontier after “video prediction” is “action prediction.” OpenAI invested in 1X, Google DeepMind in Physical Intelligence (π0), Anthropic is rumored to be building its own embodiment team.
This is the calm before the storm. In 2026 we will see the first “iPhone moment” for robotics: a device that is clearly superhuman at some narrow but economically critical task, shipped in tens or hundreds of thousands of units, and suddenly everyone realizes the exponential is real.
Next post: “Actuators: The New Oil” – why the companies that control strain-wave gears and high-performance electric motors will be the Saudi Arabia of the 2030s.
If you want to come along for the ride, stay subscribed. The rabbit hole only gets deeper.