by Theo Encrypted

Unitree 7x in a day: the market pays for iron

On August 19 Unitree opened near $66B, 7.3 times its listing price. The same day Figure's Brett Adcock asked what to do with a retired robot, and Schwarzenegger replied, "You should melt them." Between those two posts sits the whole physical-AI market, record money, scarce iron, and a brain that gets cheaper faster than it can pay for itself.

On August 19, 2026, Unitree Robotics, the humanoid maker from Hangzhou, started trading on Shanghai's STAR Market. The listing valued the company at $9.04B: it sold 10% of its shares for $905M. The first trades printed roughly 7.3 times higher. CNBC clocked a 542% jump in early minutes, Wall St Engine caught a 629% peak and a market cap near $66B.

It is the world's first public humanoid maker that ships them by the thousands and makes money doing it: 5,215 humanoids and $41M of net profit in 2025. At the listing price alone the company was worth 36 annual revenues and 219 annual profits. At the opening price, roughly 7 times more.

The same day Brett Adcock, founder of rival Figure, posted a farewell: "We loved Figure 02, it was a workhorse. We're out of room at HQ. It has to go. It deserves a proper sendoff. Any ideas?" Arnold Schwarzenegger answered in one line: "You should melt them." The Terminator gave the industry advice from his own finale. The reply drew 558k views, and Adcock reposted it with two exclamation marks.

The joke and the candle describe one market. Money is flowing into robots at record pace, while the robots in the news are more often running the 100 meters than standing in a paid shift.

This article sorts out what the market is paying for: where physical AI has a moat that survives the next few years, and where there is a shell waiting to be melted down.

How this was counted

Three sources. First, the Silicon Valley Bank report "Physical AI and the Future of Robotics," July 2026, built with Prologis Ventures on a survey of 400+ warehouse and supply-chain executives and an analysis of 100+ suppliers in the AI-accelerator chain; we worked from the Axevil Capital breakdown. Second, Unitree's listing prospectus in the RoboStrategy review of August 12. Third, X posts with metrics as of August 20, collected by our harvester, including Andrew Kang's thread and intern's article.

The robot-hour math, the multiples and the valuation comparisons were recomputed by the desk, with assumptions named right in the blocks. Where a number is an estimate rather than a reported figure, the text says so.

The turn: venture is about iron again

For 10 straight years hardware's share of US venture sat in a 15-21% band. For 2026 SVB expects about 30% excluding Anthropic and OpenAI. The reversal took 1 year. In dollars that is $120B for 2026 at the current pace, about 33% of all US venture investment.

The better tell is not the total but fund behavior.

The report names the cause directly. In the last hardware cycle, the era of Cisco, Intel and Qualcomm, a robot executed a hard-coded program in ideal conditions: a flat floor and identical parts. Now the iron got smarter. A robot copes with situations it never saw in training, and that widens the set of sites where automation applies at all.

Mark Harris, SVB's head of frontier technology, puts it this way: "What distinguishes this hardware cycle from the past is intelligence. Robots have for the first time gained the ability to generalize and act in unfamiliar environments. This shift to physical AI expands the robotics market like never before."

Two roads: the brain gets cheaper, the robot does not

If AI becomes the core of computing, the token is its unit of cost. Here the economics fork.

Cloud work gets cheaper. Writing an email costs about 5 cents; having an agent digest an annual report, about $15 at the report's blended rate. Both numbers are falling, because the same GPUs in a data center serve millions of users and split their cost between them. The same work on the open model Kimi K3 via Fireworks runs roughly 20 times cheaper.

A robot gets none of that economics. A delay that is harmless for a chatbot means an accident for a humanoid: by the time the request round-trips to a data center, the hand has already missed. The computer has to sit on the machine itself. It works for that 1 robot and its full cost lands in the robot's price. The robot buys the same scarce components as the data centers, and it has no reason to get cheaper.

The scarcity is measurable. Since 2022 chipmakers have added 13 gigawatts of annual chip power draw, 2/3 of it NVIDIA. More than 830 data centers are under construction in the US, and the key markets, Northern Virginia and Dallas, have under 1% vacancy. The median wait for a grid connection has reached 45 months. 90% of high-bandwidth HBM memory goes to 4 chipmakers, and a second buyer now competes for that same memory: machines that compute for themselves.

The desk has a separate breakdown of how Peter Thiel trades this same scarcity: the $418.7M Thiel Macro portfolio where 7 of 8 positions are power utilities.

Unitree: what $66B actually bought

Wang Xingxing's company differs in one simple way: you can buy its humanoid directly, and cheaply. Shipments grew from under 10 units in 2023 to 5,215 in 2025. The average selling price fell from $87,900 to $24,650 over the same stretch.

The cheapness rests on architecture, not dumping. Unitree builds the critical parts itself: brushless motors, planetary gearboxes, LiDAR, depth cameras. Bought-in components take only 14-18% of bill-of-material cost.

A SemiAnalysis teardown put a high-end G1's BOM at $8,976 against a pretax price near $27,300; the base G1 lists at $13,500. Humanoid gross margin holds near 60%, higher than most AI labs. Kang makes that point separately.

The question is not whether Unitree can build cheap iron. The question is who buys it.

73.6% of humanoid revenue comes from universities and research groups. Industry contributes 9%, and 50-70% of that slice is factory tours and entertainment rather than shifts on a line.

The prospectus is honest. The company's own models have not been deployed in products at scale, and data collection from real deployments has not begun. Of the $623M spending plan, 48% goes to AI models: Unitree itself names the brain as its next bottleneck. For that task it signed an agreement with DeepSeek for priority access to model training.

Kang's thread: the brainless body was the product

The most discussed voice of listing day was Andrew Kang, co-founder of Mechanism Capital and manager of the robotics fund RoboStrategy, with $19M invested in Figure. His thread explains why "a robot without a brain" was a market strategy, not an oversight.

Kang names the fork of outcomes. If Unitree climbs from the research market into the industrial one, it repeats DJI and Toyota. If not, it becomes the Raspberry Pi of robotics: a company permanently registered with experimentalists.

Kang adds one more point: thousands of sold robots are also a data-collection fleet that pure model companies do not have.

Adcock replied inside that same thread: "Majority don't understand this - they are moving in a much different direction than we are." Figure runs the opposite bet: a finished product aimed straight at deployment, with no developer-platform stage. The 1,000th F.03 came off its own BotQ line in July, and a pilot is running at BMW's Spartanburg plant.

That post is a direct jab at 1X, whose tendon-driven, 25-degree-of-freedom NEO hands we broke down in July. The industry has not even agreed on how to build a hand. A whole form factor, even less so: Adcock in parallel declared wheeled robots "an utter dead end."

For an investor the conclusion is simple: paying a premium for a specific body is premature.

The hour math: $1.66 versus $30

The second text of the day Kang reposted with the caption "good summary of the major points of our robotics investment thesis": intern's article on how to think about the robotics market. Its frame: thanks to AI, robots finally have a brain; labor is a market of roughly $60T, about 52% of world GDP; today's robotics is worth around $200B. The gap between those numbers is 300x.

The heart of the piece is the cost-per-hour math. We recomputed it and added 2 cases of our own.

Intern's base gives $1.66 an hour: a $50k robot working 20 hours a day for 7 years, with upkeep at 10% of its price annually.

The desk's conservative case gives $7.44: a $100k machine including integration, 16 hours and 300 days a year, 5 years of service, 15% upkeep. Even that is 3-4 times cheaper than a US warehouse worker at $20-30 an hour before taxes and benefits.

The third case is leasing. Kang's phrase: "robots are the GPUs of physical labor." At $50k per robot per year, 20 thousand machines make $1B of annual revenue, and 1 robot over a 7-year life brings the maker $350k instead of a one-time $50k sale.

The consequence for makers: while demand runs into supply, renting out labor beats selling bodies. The most valuable companies in the sector would then look less like factories and more like owners of fleets of on-demand labor.

The desk's caveat gets its own paragraph. Both computations assume the robot actually works its 16-20 hours a day. No humanoid maker publishes the realized utilization of deployed machines, and that is the single most unverifiable number in the whole math. The reason is spelled out in "What we cannot see."

The other side: the warehouse, the wall, and the price of self-grading

The warehouse. Prologis and SVB surveyed 400+ warehouse executives, the people who sign purchase orders. 2 of 3 grew more confident in automation over the year; skeptics are 1 in 10. But 3 of 4 say humanoids today promise more than they deliver, and what decides a purchase is the payback period, not how impressive the technology is. What is already automated is conveyors and sortation, where flow is predictable and payback is computed in advance. Everything harder is still pilots.

The wall. On August 20, at Beijing's robot games, a humanoid pushed to a speed record failed to brake and folded in half against a safety cushion. The Breaking911 clip drew 3.5M views. A day earlier Unitree had shown a new machine: a 2-meter standing jump and 12.66 m/s, formally faster than Bolt's record.

Researcher Eren Chen framed the problem better than any skeptic: "China's robotics industry may be over-optimizing for speed. Knowing when to slow down, when to stop, and how to stop safely matters just as much as how fast you can run. A robot that can sprint incredibly fast but has no idea when to brake feels like a student who aced 1 subject and ignored the rest."

The price of self-grading. Eric Jang of 1X, congratulating Dyna on its 1M-hour video pretrain, added a warning the desk considers the line of the month: "Robotics is entering a scale-up era. The scale of investment is very serious, and so warrants serious rigor when companies make claims about models that only they can verify. Otherwise, we risk vaporizing billions of VC dollars underwritten by self-reported evaluations of capabilities."

Geopolitics: the market got cut in half

On July 28 the FCC barred imports of new Chinese humanoid and quadruped robot models into the US. Bernstein calls it the start of US-China robotics decoupling, with possible extensions to chips and investment, and a Chinese answer through rare-earth magnets.

State money is already in the game. China runs a 1 trillion RMB robotics fund. A former US national security advisor proposes a mirror $50B fund; Kang's comment is "I tend to agree."

For Unitree the ban is a ceiling. The US supplied 13.3% of revenue and, more importantly, hosts the physical-AI labs that built on its iron.

For American makers the same ban works as a subsidy: demand for "made in USA" gets regulatory protection. Hence Path Robotics' $600M+ contract with shipbuilder Huntington Ingalls, against a 400,000-welder shortage with an average age of 55. Hence Standard Bots' bet on manufacturing industrial robots in the US.

The portfolio consequence: the "China discount" and the "America premium" in valuations are no longer sentiment, they are a tariff. Identical iron on opposite sides of the FCC list is priced differently, and that is here to stay.

Capital gathered in a handful of names

Venture investment into hardware supply chains and the models for them reached $25B, and nearly all of it came in large checks: 84% of capital in deals above $500M, versus 4% in 2023. The market moved to very large bets on autonomy and handed them to a short list of names.

The same table has a flip side. 48% of exits happen before round B; 7 of 10 exits are acquisitions, the median one at $56M. Companies get bought while they solve a narrow problem. The ones that grow to standalone scale are those whose solution carries from one site to the next.

Unitree's opening valuation equals 60 Boston Dynamics at the 2021 price. The derivatives market missed in the other direction: a week before the listing Hyperliquid priced Unitree at $35B, half the open.

The day after the debut Binance listed a UNITREE perp with up to 20x leverage. Crypto got a 24/7 bet on humanoids before most brokers could offer the stock itself: the STAR Market is effectively closed to foreign retail.

Where the moat is, and where the shell

The moat in physical AI for the next few years is built from 4 layers. "How human the robot looks" is not one of them.

Hardware cost curves. Unitree's vertical integration works as a loop: lower cost, lower price at the same margin, higher volume, lower cost again. Adcock says the same from Figure's side: "Figure had to build our own supply chain. There was no off the shelf humanoid parts." The most concentrated piece of the loop is actuators, 40-60% of a humanoid's BOM.

Scarce components. HBM memory, actuators, rare-earth magnets, grid connections. This layer earns under any outcome of the form-factor dispute. Our position from the physical-AI breakdown has not changed, it has strengthened: the SVB report and corporate checks now stand on it too. Amazon alone invested in or bought 10 robotics and drone companies in 12 months.

Real-shift data. Dyna showed scaling laws on 1M hours of human video. NVIDIA taught a humanoid with 22-DoF hands from 20k hours of egocentric video with no robot in the loop. Teleoperation, per Jim Fan, is dying, and 16TB of manipulation data already sits open-source.

It follows that raw data commoditizes faster than iron. What gets expensive is what you cannot download: actual shifts at customer sites, with failures, recoveries, and metrics verified by an external client.

Portability. Axevil's formula after the report: the market pays for generalizability, not for the robot. SVB and Prologis suggest watching 3 things: does the solution run on different sites without rework, do the computer and the model sit on the machine itself, does the company accumulate real shifts instead of demos. That is the buyer's checklist, and legs are not on it.

Where to invest, and what to hold

The desk's position for the next few years, from the conservative layer to the risky one.

The public core is unchanged: the semiconductor chain, HBM memory, power and grid connection, industrial robotics with an installed base. This layer takes a cut of the whole physical-AI economy no matter which body wins. It is also the only layer where the numbers are checked by filings, not demos.

Unitree is the sector's public benchmark, but not a position. 36x revenue at listing became roughly 260x at the open, foreigners cannot access the stock, and the UNITREE perp is a news-trading instrument, not a hold-for-years one.

Watching it is still mandatory. Unitree's filings will be the first public data series on real humanoid demand. The second series arrives with Agility's SPAC in Q4: at $2.5B premoney against Unitree's 66, the market will for the first time price the difference between "sells by the thousands" and "leases a fleet."

The private upside is where portability lives: model companies with externally verifiable results, and companies that own real-shift data. After Jang's warning the desk filter is simple: if a model's results can only be verified by its authors, the valuation rests on self-grading.

Avoid: humanoid shells with no deployments and no cost-curve loop of their own. The story of Sally, the $57k robot teacher showed what price parity without capability parity looks like. 3 of 4 warehouse executives say the same thing with their purchase orders.

What to watch and listen to

2 fresh conversations cover both sides of the argument: the maker and the investor.

Adcock on My First Million talks about the next generation of AI hardware and models, humanoids, and the school weapon-detection systems his company Cover builds. The context for watching comes from this article: the 1,000th F.03 off the BotQ line, the BMW pilot, and the abandoned tendon hand. It is worth hearing how the "finished product first" bet sounds when the maker himself defends it.

Kang on Mad Society, August 12, a week before the Unitree listing. From the episode's timestamps: how he got a Figure allocation alongside OpenAI and Bezos and raised his check from $500k to $19M in a week; why Figure is "the Apple of robotics" to him at $39B versus Unitree's 9; why manufacturing is the sector's bottleneck and hardware difficulty is the moat; and how RoboStrategy works as "the Strategy for robotics," a public venture fund. Best listened to in a pair with his thread above: the thread is the conclusion, the podcast is the reasoning.

What we cannot see

Realized utilization. Neither Unitree, nor Figure, nor 1X publishes how many hours a day their deployed robots work. The whole labor-hour math stands on a 16-20 hour assumption; if the real number is 4, the cost per hour rises 4-5x. The bias favors the robots.

Data return from sold machines. Kang's "data fleet" thesis works only if data from sold G1s flows back to Unitree. Most of its machines are research units running scripted motion in unrepresentative settings, and the prospectus concedes that collection from real deployments has not begun. The bias favors the data thesis.

The makeup of "industrial" revenue. Even within Unitree's 9% industrial slice, 50-70% is tours and shows. How many humanoids worldwide are standing in a paid shift on a line right now is not knowable from any public source. The bias favors the whole sector.

Private valuations. Figure's $39B and Skild's $14B are prices of their last rounds, not of a market; they do not reprice between rounds. Comparing them to Unitree's exchange candle is comparing a photograph to a video. The direction of the bias is unclear: after the public benchmark's 7.3x, the private marks could just as well be too low.

What would have to be true for us to be wrong

If Unitree lifts industry's share of humanoid revenue from 9% to even 25% within 4 quarters, the "research demand does not scale" bet is dead and the opening premium was right. Checked against STAR Market quarterly filings.

If Figure publishes verifiable shift metrics for F.03 at BMW (hours, throughput, cost per hour) and they come in under $10, the desk's conservative case was too harsh and the "shell" was underpriced. Checked by the publication of the metrics themselves.

If at least 1 model company exposes an open inference endpoint on which an external lab reproduces the claimed results, the "self-grading = discount" filter must be relaxed. Checked by the endpoint appearing; Jang calls it the next phase.

If raw iron starts getting cheaper faster than volumes grow, say the average humanoid price drops below $10k while maker margins fall under 30%, then the cost-curve moat is thinner than we think and the value migrates entirely to the model layer.

Corrections made along the way

An early draft attributed "You should melt them" to a standalone Schwarzenegger post. It is actually a reply inside Adcock's thread, so the article paraphrases it instead of embedding it: the desk's rule does not quote thread replies.

The first version compared Unitree with UBTech on the "valuation" row, but the desk had no verified UBTech market cap for the date. The comparison was replaced with Agility Robotics and its disclosed SPAC price. UBTech's $104M loss and the 5x shipment gap stayed in the text: those numbers are in the prospectus review.

The opening multiple "about 260x revenue" is computed against 2025 revenue. The desk has no 2026 consensus, and any "forward" number would have been an invention, so the article does not contain one.

TT desk thoughts

We are dropping the split of the sector into "robots" and "not robots." The working split is now 4 layers: hardware cost curves, scarce components, real-shift data, portability. Every new company in the desk tracker gets a layer tag. A humanoid shell without a layer is news, not a position.

We are putting 2 numbers under watch. First: industry's share of Unitree's humanoid revenue in quarterly filings, the only public series that separates research fashion from labor. Second: the spread between Agility's SPAC valuation and Unitree's multiple. It will show what the market pays for Chinese cost curves versus American access.

We are changing the demo-citation rule. After the wall in Beijing and Jang's warning, a clip without an externally verifiable metric passes through desk articles as marketing, not evidence. That includes clips from companies on our own "private upside" list.

Ticker specifics, by layer. Memory: Micron (MU), SK hynix (000660.KS), Samsung (005930.KS): the 2026 HBM output is sold out in advance, and SK Group expects the shortage to last into 2030. Compute: NVIDIA (NVDA), 2/3 of the chip power-draw added since 2022 is theirs. The outlet: Vistra (VST) and American Electric Power (AEP) from Thiel's portfolio, broken down in the link above. Industrial installed base: Fanuc (6954.T), Yaskawa (6506.T), ABB (ABB).

What the desk would actually hold: the memory-plus-outlet pair, MU and VST, as the multi-year core, and NVDA as the single bet on all of physical AI at once. Pure humanoid tickers are a different story: TSLA carries Optimus inside a $1T+ company, the UNITREE perp is for news trading only, and we watch the Agility SPAC (CCXI) without paying until the deal terms are disclosed. Data and models still live only on the private market; that layer has no public ticker.

Check 1 number for us: $1.66 an hour. Take our assumptions from the robot-hour block, put in your own machine price, hours and service life, and see where a robot-hour stops being cheaper than a warehouse. If your threshold lands below $50k per machine, write to us: it means the desk's conservative case is still too generous.

Keep reading

AI × Crypto Financial Infrastructure · All notes