Ep. 016 - What Unitree's Evolution Means For Robotics (Robotics)

ELI5/TLDR

A Chinese company called Unitree started by making cheap robot dogs nobody thought were useful. Now it makes cheap humanoid robots and is about to go public. The SemiAnalysis crew argues the interesting thing isn’t that these robots are good yet — they break every 20 minutes and mostly give people tours — but that Unitree keeps making them cheaper and better faster than anyone in the West, riding the same Chinese manufacturing ecosystem that gave us cheap drones and EVs. The bet: humanoids don’t need to be perfect to be worth buying, just a little cheaper than hiring a person who keeps quitting.

The Full Story

The robots aren’t good yet, and that’s the point

The conversation opens by puncturing its own hype. The hosts — Nico and Raik from SemiAnalysis, with Jordan moderating — are quick to say that “industrial deployment” of Unitree’s robots is, for now, mostly a fancy phrase for robots showing people around buildings.

“At best, we’re talking about the frontier of deployment is like teleop to pick up boxes and stuff of this nature… we’re in baby days to put it lightly.”

Teleop here means teleoperation — a human somewhere else holding the controls, like a very expensive remote-control car. The robots still “burn out” (overheat and need rest), carry small payloads, and aren’t precise. So why the excitement? Because the rate of improvement is fast, and because the bar for usefulness is lower than people assume.

Why bother going public

Unitree is filing for an IPO. Part of the reason is structural — Chinese companies often go public faster, sometimes with a go-public clause baked into the original investment terms. Part of it is that the business is real. The hosts note the robots have been wildly profitable: a humanoid priced around $27,000 still carries roughly 67% gross margins.

“It’s kind of just Unitree in there gouging the living hell out of the market while they can.”

That margin exists because almost nobody else can build these at scale yet. Prices have already collapsed — from robots that cost $54,000 and “would burn out in 5 minutes” down to around $30,000. The revenue looks healthy partly because the old prices were absurd, and partly because there’s a bigger research-and-hobbyist market than outsiders realize.

The drone playbook, run again

The recurring analogy is DJI, the drone giant. Before DJI, there was no consumer drone market — just hobbyists soldering parts in basements or governments buying $20,000 military-grade units. DJI shipped one merely-good, affordable, camera-equipped drone and revenue went from $4 million to $130 million in about two years. The market didn’t exist; DJI invented it by proving the thing was useful and cheap enough.

The hosts argue Unitree is running the same move. Make the robot cheap enough that curious engineers and even non-technical people save up to buy one just to see what it can do. They call this Unitree’s claim to the “mandate of heaven” — sell a product people love, earn the right to build the next, harder one, and climb the capability ladder while cutting cost at each rung. The phrase they keep returning to:

“Economies of scale is like China’s scaling law.”

Think of it like compound interest on manufacturing: each cheaper, better product funds a wider market, which funds the next product.

Why “a little better than a person” is enough

A key idea borrowed from their earlier “levels of autonomy” paper: a robot doesn’t need superhuman skill to be worth deploying. It needs to break even against a human on cost for one specific, narrow task.

Their example: even assuming 100% teleoperation, a failure every 20 minutes, and 5 minutes to repair each time, a Unitree humanoid moving a box from here to there comes out slightly better than a human doing the same thing.

“That’s a whole job, right? Like that’s a whole job.”

The economic lever they emphasize is the loaded cost of labor — not just the wage, but the full cost of hiring, onboarding, training, and replacing someone. In the US and Europe that cost has gone non-linear because people quit constantly. So a robot that runs at 70% of human speed but works two shifts and doesn’t quit can win even while being objectively worse. Each narrow task that crosses this threshold — stirring onions in a kitchen, scanning a construction site, swapping server racks in a data center — opens a “demand shock” that pulls capital and talent in.

The real moat: the supply chain

The deepest part of the discussion is about why China specifically. The hosts point to Huaqiangbei in Shenzhen — a seven-story building where the entire consumer-electronics supply chain sits under one roof.

“You show up with your yuan… throw it down, and you walk out, and you have every single part you need to build a drone now, right? Like all within a single building.”

Humanoids inherit most of that ecosystem — standardized motors, injection-molded plastics, controllers. What’s new is that specialized parts like planetary gearboxes (the geared joints that let an arm move with force and precision) are now being made by countless small Chinese firms because ~200 humanoid companies suddenly need them. Drones never needed gearboxes; now every province has someone cutting gears. The hosts call this a derivative ecosystem — built on top of decades of consumer-electronics and automotive manufacturing, not from scratch.

They describe the underlying culture as “a sea of a thousand bosses” — thousands of small specialists, each the world’s best at one component, competing brutally on price and yield. That internal pressure, not government decree, is what produced the quality and speed.

The H1 was a robot dog standing up

A delightful detail: Unitree’s first humanoid, the H1, was essentially a quadruped (four-legged robot) redesigned to stand on two legs. Its legs stay bent like a dog’s, and it walks with a slight patter. They reused the actuators and mechanics that were already “good enough” for robot dogs and transferred them to a humanoid. The scaling is visible in real time — the S1 went from ~400 units in early 2025 to ~4,000 over nine months, then ~6,500 a few months later.

The bear case and the wake-up call

Asked for the bear case, the hosts concede Unitree will face serious competition — “a new humanoid company in China every week.” But they argue Unitree benefits no matter who wins, because the Western supply chain itself leans on Chinese components. Their bet: Unitree becomes “a good business,” probably a great research-tools business.

The closing note is geopolitical. The US is described as massively under-invested — weak on PCBs, neodymium (the magnets in motors), metal processing, the chemicals for actuators. The article was meant as a wake-up call: you don’t need to believe Tesla’s Optimus or Figure’s full-fledged humanoid dream to see that the market is already moving, and China’s structural advantage isn’t going away soon.

Key Takeaways

Teleoperation means a human remotely controlling the robot; the deployment “frontier” today is teleoperated box-picking, not autonomous work.
Unitree humanoids reportedly carry ~67% gross margins at a ~$27,000 price — a sign of near-zero competition, not a durable moat.
Robot prices collapsed from ~$54,000 (burning out in 5 minutes) to ~$30,000 with far better endurance.
The DJI precedent: a single affordable, good-enough product can create a market that didn’t previously exist (DJI: $4M → $130M revenue in ~2 years).
Usefulness threshold is low — a robot only needs to break even against a human on one narrow task, not match human skill.
Loaded cost of labor = wage + hiring + onboarding + training + churn; it has gone non-linear in the West, making slow-but-tireless robots competitive.
A robot at 70% human speed can still win if it runs two shifts, doesn’t quit, or makes fewer visual (vs manipulation) errors.
Planetary gearboxes are the geared joints giving robot arms force and precision; a whole Chinese supplier ecosystem is forming around them now because ~200 humanoid firms demand them.
Drones don’t use gearboxes (just high-speed motors), so the gearbox supply chain is genuinely new — but it’s a derivative of existing electronics/auto manufacturing.
Huaqiangbei (Shenzhen) is a seven-story building holding the full consumer-electronics supply chain — buy every drone part in one place, same-day.
“Sea of a thousand bosses” — thousands of small, specialized component makers competing on price and yield; quality emerged from survival pressure, not mandate.
Unitree’s first humanoid (H1) was literally a quadruped redesigned to stand on two legs — bent knees, pattering gait — reusing robot-dog actuators.
Unit scaling is visible: S1 went ~400 → ~4,000 → ~6,500 units across roughly a year.
US gaps: PCBs, neodymium magnet production, metal processing, actuator chemicals — most still depend on China even if final assembly happens domestically.
“Economies of scale is China’s scaling law” — the oversupply strategy: cheaper product → wider market → next product, compounding.

Claude’s Take

This is a genuinely good conversation that mostly resists its own genre. The SemiAnalysis crew clearly love hardware and have done the homework — the bill-of-materials gushing, the gearbox-ecosystem observation, the H1-was-a-dog detail, and the “loaded cost of labor” framing are the kind of specifics you only get from people who actually read the supply-chain tea leaves. The repeated insistence that the robots are not good yet is refreshing and earns the rest of the argument credibility.

Where to apply the BS filter: the core thesis rests entirely on the DJI analogy, and they wave away the obvious objection themselves before fully answering it. Drones had latent hobbyist demand; humanoids are being asked to invent demand and clear a much higher reliability bar in safety-critical environments. “It’s slightly better than a human at moving a box under 100% teleop” is a real claim, but it quietly concedes there’s a human in the loop — which is a labor-arbitrage story, not an autonomy story. The “demand shock” language is doing a lot of hand-waving; every example (cooking, data centers, logistics) is hypothetical. And there’s an unexamined assumption that software reliability will arrive on schedule, which is exactly the part nobody can underwrite.

Score: 7. Substantive, specific, intellectually honest about current limitations, and strong on the one thing they truly understand — the manufacturing ecosystem. It loses points for leaning hard on a single analogy and for the back-third drifting into banter about robots dancing and folding laundry. Worth the listen for the supply-chain mental model alone.