By Sarah Xiao 
On 26 October, Unitree Robotics (688836.SH) traded as high as ¥737.50 intraday and closed at ¥672, up 7.69% on the day, for a market capitalisation of ¥271.8 billion. Measured from 16 October, the stock is up 65.89% across seven trading sessions.
The price move is the easy story. The more interesting document is the research note behind it: on 28 September JPMorgan initiated coverage on Unitree at Underweight with a ¥300 target; on 26 October, 27 days later, it upgraded the stock two notches to Overweight with a target of ¥995. Both targets are dated December 2027. That is a 3.3x change in a valuation that, on the bank’s own numbers, has not changed its multiple at all.
What the second note says changed
The trigger was not a quarterly result. It was a conference. At the World Physical AI Conference in Hangzhou, Unitree founder and CEO Wang Xingxing compressed his own timeline for a physical-AI breakthrough from “up to ten years” to two to three years, and in some remarks to one to two years. The event drew more than 3,000 participants across city governance, emergency response, education, entertainment and industrial sessions.
The analyst’s reasoning follows from that compression. If commercialisation is still a decade away, buying hardware leadership today is a bet on a long, contested race in which rivals have time to catch up. If the applications arrive in two years, then model companies and developers need working platforms now — robots that can be shipped, programmed, instrumented for failure data and iterated on. Installed base and developer relationships stop being a manufacturing statistic and start being an asset in the software race.
That is a coherent argument. It is also a claim about probability, not a fact, and the note’s own numbers show how much of the target depends on it.
The model change is almost entirely a 2028 volume change
Laid out side by side, the two notes differ in one year and barely at all in the others.
The 2026 forecast is identical at 14,200 units. The 2027 forecast rises from 26,200 to 31,700. The 2028 forecast goes from 41,600 to 111,600 — a 169% increase in a single revision. Revenue for 2028 moves from ¥5.269 billion to ¥13.383 billion, and adjusted net profit from ¥1.161 billion to ¥3.202 billion.
Two things follow from that pattern, and both matter more than the headline target.
First, the upgrade is not a near-term call. Nothing about the next twelve months changed in the model. The bank is paying for 2028 two years early, before corporate repeat purchasing has been demonstrated at anything like that scale. The note itself flags this: enterprise deployment after trial, maintenance cost over a robot’s service life, and whether the platform can convert installed base into revenue are all still open, and only operating data will close them.
Second, the 2028 number implies a much larger market, not a larger share. The same note forecasts Unitree’s share falling from 35% in 2025 to 15% in 2028. A 15% share with 111,600 units implies a global humanoid market of roughly 744,000 units in 2028 — the forecast is a statement about the size of the market, not about Unitree winning it. Readers who take “Overweight” as “Unitree will dominate” have the note backwards.
There is a third check worth doing: price. The average selling price fell from ¥593,000 in 2023 to ¥166,000 in 2025, and the revised 2028 revenue divided by the revised 2028 volume gives roughly ¥120,000 per unit. So the upgrade assumes continued deflation of about 28% from 2025 levels, with gross margin already contracting two to three percentage points year on year and net margin holding near 22%. Volume is doing all the work; price and mix are not helping.
The arithmetic behind ¥995
The number looks like a leap. It is mostly a multiplication, done twice.
The September note valued 2027 revenue of ¥4.038 billion at 30x price-to-sales, which produces a market value of roughly ¥121 billion and, on about 404 million shares, a target near ¥300. The October note values 2028 revenue of ¥13.383 billion at the same 30x, producing roughly ¥401 billion and a target near ¥995.
Split the difference into its two components. Rolling the valuation year forward while keeping the old 2028 revenue forecast of ¥5.269 billion at 30x gives about ¥391 — a mechanical gain of ¥91, or 30%, which is simply one year of expected growth. The remaining ¥604, about 87% of the total increase, comes from raising the forecast itself.
That decomposition is the useful thing to take away. A “3.3x target” sounds like a violent change of view about a company. Seen properly, it is a modest calendar adjustment plus a very large, explicitly probabilistic bet that 2028 volumes will be 2.7 times what the same analyst thought four weeks earlier.
What the note still says against the stock
The bear case did not disappear; it moved into the forecast. The October note still models market share falling by 20 percentage points in three years, still describes average selling prices in decline, still notes gross margin compression, and still flags the regulatory problem: US FCC restrictions on foreign-produced advanced robots now cover complete humanoid and quadruped systems as well as components, require detailed disclosure and conditional approval, and land on a company that booked 43% of 2025 sales overseas, about a third of it in the United States.
The other pillar is scarcity. The note argues that the IPO pipeline for robot companies is slower than it expected, so the premium attached to the only large listed pure play in A-shares should persist longer. That supports a high multiple; it does not add a yuan of revenue. It is also the part of the thesis most exposed to someone else’s decision to list.
The part that is genuinely new
Strip out the marketing and one idea remains, and it is worth taking seriously regardless of what happens to the share price.
For two years the industry assumed that progress in foundation models would commoditise the body and let software capture the value. The October note inverts that: as models get better at planning and reasoning, the bottleneck moves to execution in the physical world — grasping, contact, locomotion, the unglamorous last mile where a policy meets friction, tolerance and a dropped robot. In that world, a company that has shipped 10,000 to 20,000 G1 units since May 2024 and has a developer ecosystem around them owns something that cannot be bought quickly.
The caveat is the one the note itself supplies: cumulative production is not paid delivery, an installed base is not a data flywheel, and a data flywheel is not revenue. Each link in that chain requires a commercial proof that does not exist yet. The 995 target is a position on how fast those links will close.
What to watch, if you care about the industry rather than the ticker
- Repeat orders. The first purchase is a trial; the second is a business. Watch enterprise re-orders, not unit shipments.
- Cost per operating hour. Uptime, maintenance and spare parts decide whether a robot competes with labour or with a demo budget.
- Whether the data argument becomes an invoice. Platform fees, model licensing, or managed services — anything that shows value moving from the body to the stack.
- The IPO queue. Every new humanoid listing tests the scarcity premium that a large part of this valuation rests on.
This article summarises and analyses a public sell-side research note and the market reaction to it. It is not investment advice, and it is not a recommendation to buy or sell any security. Figures are as reported in the two notes dated 28 September and 26 October 2026; the arithmetic in the third chart is our own.
FAQ
Why did JPMorgan upgrade Unitree by two notches in 27 days?
The stated trigger was the World Physical AI Conference in Hangzhou, where founder Wang Xingxing shortened his expected timeline for a physical-AI breakthrough from up to ten years to two to three years, and in some remarks to one to two. The analyst concluded that hardware platforms and developer ecosystems become more valuable when applications arrive sooner, and raised the 2028 volume forecast by 169%.
Where does the ¥995 target price come from?
From 2028 estimated revenue of ¥13.383 billion at a 30x price-to-sales multiple, the same multiple the September note used. The September target of ¥300 applied 30x to 2027 estimated revenue of ¥4.038 billion. Both targets are dated December 2027.
Is the whole increase from the higher forecast?
No. About 30% of the step from ¥300 comes simply from rolling the valuation year forward — applying 30x to the old 2028 revenue forecast of ¥5.269 billion gives roughly ¥391. The remaining move to ¥995 is the forecast revision.
Does the note expect Unitree to gain market share?
No, the opposite. It models Unitree’s share falling from 35% in 2025 to 15% in 2028. The upgrade is a bet on the size of the 2028 market, which at 111,600 units and a 15% share implies roughly 744,000 humanoid units sold globally that year.
Is this investment advice?
No. This is a summary and analysis of a public research note. It contains no recommendation to buy or sell any security.