XPeng's Robot Unit Is Now a $6.3 Billion Company. XPeng's CEO Runs It and Bought 8.75% for $100 Million Down.

Abstract

On 7 September 2026 XPeng announced that it had commissioned production lines for its IRON humanoid robot and that the first machine off them had walked away under its own power. More than 80% of core processes are automated, the company said; scaled production is due by the end of this year, with the first robots going into XPeng's own stores and campuses and outside deliveries in China and overseas from 2027. It is a real manufacturing milestone and it is the part of the story everyone has covered.

The part nobody has covered is the document underneath it. Two weeks earlier, on 24 August, XPeng filed the announcement for the financing that pays for the line. It is a discloseable and connected transaction, and it prints its own cap table. Four outside investors — IDG Capital, Alibaba, Tencent and Gaorong Ventures — subscribed for 296,025,600 Series A preferred shares in Dogotix Inc., the Cayman company that now holds the robot business, for US$600 million. Two executives, through companies they wholly own, subscribed for 49,337,600 ordinary shares for US$100 million and for warrants over a further 246,688,000 ordinary shares at a purchase price of US$123.35. Add the warrant shares and the executives hold 296,025,600 shares. The same number. To the share.

Both blocks come out at 8.75% of Dogotix fully diluted. The difference is when the money moves and what the shares are. The investors wired US$600 million in August. He Xiaopeng and Brian Gu wired US$100 million and hold an option to wire the remaining US$500 million at any point before Dogotix lists, at the price the round set in August 2026. In exchange for giving up that option, the investors took preference: a liquidation preference, anti-dilution, and a redemption right that lets them require Dogotix — or XPeng itself — to buy the shares back if there is no qualified listing within seven years. And He Xiaopeng is not an outside buyer. Since June 2026 he has been chief executive of the robotics business as well as chairman and chief executive of XPeng, so the man taking 8.75% of Dogotix is the man who runs Dogotix and the man who runs the company selling it to him. That is why the subscription is a connected transaction, and why he abstained from the board vote on his own half of it.

Two more numbers fall out of the same filing. The Dogotix 2026 Equity Incentive Plan reserves 469,502,968 shares, 14.96% of the enlarged company. At the price this round set, that pool is worth US$952 million: more than the US$900 million the round raised. And US$444 million of the headline "over US$6.3 billion" post-money valuation is those plan shares, counted at the round price before anyone has been granted one. Shares that someone actually paid cash for total US$5.9 billion.

None of this is hidden. All of it is in an announcement XPeng published in full, with the independent non-executive directors stating the terms are fair and reasonable. It is simply not what "raised $900 million at a $6.3 billion valuation" conveys. What follows reads XPeng's robot vertical through the document that carved it out: how IRON got to a line, the cap table the announcement prints, the incentive pool that is larger than the round, the seven-year clock attached to the outside money, and what the line has to build for any of the mark to be worth what it says.

Table of contents

1. What XPeng commissioned on 7 September
2. How IRON got to a production line
3. The cap table the announcement prints
4. The same 296,025,600 shares, bought two ways
5. The incentive pool is larger than the round
6. Where the $6.3 billion comes from
7. The redemption right, and the seven-year clock
8. What the line has to produce
9. What the retained stake is worth
10. Implications

1. What XPeng commissioned on 7 September

The announcement is short and its claims are specific. XPeng has commissioned production lines for advanced general-purpose humanoid robots at its manufacturing facility; the first robot completed line assembly and walked off under its own control; more than 80% of core processes on the lines are automated; the lines combine automotive-grade quality systems with precision manufacturing. Mass production is scheduled by the end of 2026, first into XPeng's own stores and campuses, with an official launch and deliveries in China and overseas markets in 2027.

He Xiaopeng's own framing, in the release: "The robot production lines were created from scratch with no precedent to follow. Today's step is small, but XPENG is building the production lines for an entirely new product category." He posted the footage himself the same day.

IRON, the robot coming off the line, carries 76 degrees of freedom across the body and 21 in each hand, and runs three of XPeng's in-house Turing chips for a combined 2,250 TOPS of on-device compute. Those figures are the company's, repeated consistently since its November 2025 AI Day and on the second-quarter call.

What the announcement does not contain is a rate. Not units an hour, not units a shift, not units a year. That absence is worth holding onto, because every other announced humanoid factory has led with exactly that number: Figure says BotQ can build one robot an hour and 12,000 a year; Agility says RoboFab peaks at 10,000 a year; Elon Musk is reported to have described the first Fremont Optimus line, in March 2026, as a million-unit line. XPeng has given a percentage of processes automated and a date. Nobody outside the company knows what the line can build in a week, and this piece is not going to guess.

2. How IRON got to a production line

IRON is not new. XPeng showed a first-generation humanoid at its November 2024 AI Day. At the 2025 AI Day on 5 November, it showed the next generation, and the demonstration produced the strangest publicity the category has had: the walk was fluid enough that a large share of online viewers insisted there was a person inside the suit. He Xiaopeng answered by having the robot's leg cut open on stage to show the actuators and wiring, then released a video the next morning unzipping its back.

On the third-quarter 2025 call that same month, He put dates on it: mass production by the end of 2026, and a floated ambition of a million units a year by 2030. In June 2026 he took the chief executive role of the robotics business himself. "Since June, alongside my responsibilities as CEO of XPeng, I have also taken up on the role of the CEO of our Robotics business," he told the second-quarter call. It is the clearest signal available of where the company thinks its optionality sits, and it is what makes everything downstream a related-party matter: the man running Dogotix day to day is also the chairman and chief executive of its controlling shareholder, and, two months later, the buyer of 8.75% of it.

Then, on 24 August 2026, XPeng announced the financing on the same morning it reported second-quarter results. The results were a margin-repair quarter that we covered at the time — half the gross profit came from 14% of the revenue — and the stock fell 8.53% in that session, to $11.15. On the call, He described the round as setting "a new private financing record in China's humanoid robotic industry" and said XPeng would "build the world's most valuable humanoid robot company."

Two weeks later the line was running. The sequence matters: the money and the machine are the same story, announced a fortnight apart.

XPeng's own footage of the walk-off is the only public look at the line. In its newsroom note the company puts the automation of core processes at over 80% and specifies IRON at 76 degrees of freedom across the body, 21 per hand, on three Turing chips rated 2,250 TOPS — figures that describe the robot rather than the factory, and which nobody outside XPeng has verified.

3. The cap table the announcement prints

Because the transaction is discloseable and connected under the Hong Kong Listing Rules, XPeng had to publish the shareholding structure before and after. It did. Dogotix Inc. — a Cayman company, wholly owned by XPeng before the round — had 2,466,880,000 ordinary shares in issue. After the subscription, with the incentive plan's mandate fully used and the warrants unexercised, there are 3,137,420,425 shares and XPeng holds 73.80%. With the warrants exercised in full and the additional investor reserve taken up, there are 3,384,108,425 shares and XPeng holds 68.41%. The announcement calls the difference "a deemed disposal and/or possible disposal of approximately 31.59% equity interest."

The subscriptions themselves:

• XPeng Dogotix, a wholly owned XPeng subsidiary, subscribed for 98,675,200 Series A preferred shares for US$200 million. This is XPeng putting its own cash into its own subsidiary; it is not outside validation of anything.
• IDG Capital, Alibaba, Tencent and Gaorong Ventures subscribed for 296,025,600 Series A preferred shares for US$600 million. IDG took half of it; the other three took 49,337,600 shares each.
• Executive Subscriber A, wholly owned by He Xiaopeng, subscribed for 39,470,080 ordinary shares for US$80 million, and a warrant over 197,350,400 ordinary shares for US$98.68.
• Executive Subscriber B, wholly owned by Hongdi Brian Gu, subscribed for 9,867,520 ordinary shares for US$20 million, and a warrant over 49,337,600 ordinary shares for US$24.67.
• An unnamed additional investor may take up to 7,400,640 further preferred shares for US$15 million within four months.
• The Dogotix 2026 Equity Incentive Plan reserves 469,502,968 ordinary shares.

Everything in the round was priced at the same number: US$600 million divided by 296,025,600 shares is US$2.026852 a share, and every other subscription, and the warrant strike, works out to the same figure. That single price is what makes the rest of the arithmetic here checkable.

4. The same 296,025,600 shares, bought two ways

Set the two private blocks side by side and the symmetry is exact rather than approximate.

The four institutional investors end with 296,025,600 preferred shares — 8.75% of the fully diluted company — for US$600 million, paid at closing.

He Xiaopeng and Brian Gu end with 296,025,600 ordinary shares — 8.75% of the fully diluted company — for US$100 million paid at closing, US$123.35 for the warrants, and US$500 million payable only if and when they choose to exercise. The warrant purchase price is not a discount to anything: the announcement says it "represents the par value" of the shares it covers, which works out to five ten-millionths of a dollar each. The strike is the round price, unchanged.

The exercise window is the part that carries the value. The warrants are exercisable "at any time after the date of issue and prior to one business day prior to the qualified initial public offering of Dogotix." There is no annual schedule and no expiry other than a listing. If Dogotix is worth four times as much in 2030, the executives can pay August 2026 prices for 246,688,000 shares in 2030. If it is worth less, they need not pay at all.

What is that option worth? Pricing it needs a volatility assumption and an expected tenor for a private company that has never sold a product, and this is exactly where analysis usually invents a number. We are not going to. What can be said without inventing anything is the direction: an at-the-money call with a multi-year window and no obligation to exercise is worth strictly more than zero, and the whole of that value sits with the executives rather than with XPeng's ordinary shareholders.

The announcement does not pretend otherwise. It says, in terms, that the executives "enjoy the time value of not having to contribute the exercise price of the Dogotix Warrants at the same time as the Investors' subscription," and argues the benefit is offset by two things: the preferred shares carry rights the ordinary shares do not — redemption, liquidation preference, pre-emption, rights of first refusal, co-sale, anti-dilution and tag-along — and each executive signed a non-competition undertaking running until eighteen months after he leaves. The independent non-executive directors concluded the terms are fair and reasonable and on normal commercial terms. He Xiaopeng abstained from the relevant board resolutions, as he had to: he is chairman and chief executive of the seller, chief executive of the company being sold, and — through a company he wholly owns — four fifths of the executive block doing the buying. Brian Gu, XPeng's vice chairman and president, is the other fifth.

That is a defensible package, and readers can weigh it. It is also, plainly, a different bargain from the one the headline describes.

5. The incentive pool is larger than the round

The Dogotix 2026 Equity Incentive Plan has a scheme mandate limit of 15% of the share capital as enlarged by the subscription — 469,502,968 ordinary shares, 14.96% of the post-round company. At US$2.026852 a share, that pool is worth US$952 million. The round raised US$900 million.

The plan is funded two ways, and the split matters more than the total:

• 250,401,583 existing shares transferred out of XPeng's own holding — 8% of the enlarged capital, US$508 million at the round price. XPeng's ordinary holding falls from 2,466,880,000 to 2,216,478,417 as a direct consequence.
• 219,101,385 newly issued shares — 7% of the enlarged capital, US$444 million, which dilutes every holder including XPeng.

Roughly half a billion dollars of the pool therefore comes out of the parent's stake rather than out of new issuance. That is not a criticism — an operating company being carved out needs equity to hire against, and a pool funded partly by transfer is a normal way to build one. It is simply the largest single equity movement in the transaction, larger than the outside investors' cheque, and it appears nowhere in the coverage.

Two qualifications belong here at first use, not in a footnote. The US$952 million is a reservation at the round price, not a grant: grant price, exercise price and vesting all sit with the plan administrator and none of them has been set. And a ten-year plan is a ten-year plan; nothing says the mandate will be fully used. The announcement's own percentages, including the 68.41%, assume it is.

6. Where the $6.3 billion comes from

The valuation everyone quoted is a defined term in the announcement, and the definition is where the interesting part lives.

Pre-transaction valuation: the round price applied to the 2,466,880,000 shares Dogotix had before the round. US$5.0 billion exactly. Post-transaction valuation: the same price applied to the share count immediately after the subscription and the full utilisation of the incentive plan's mandate, expressly excluding the warrants and the additional investor's reserve.

So the reconciliation is clean. Five billion of pre-existing value, US$900 million of subscriptions, and US$444 million of newly issued incentive shares that nobody has paid for and nobody has yet been granted. That is the "over US$6.3 billion."

Strip the plan out and the post-money on shares that someone actually paid cash for is US$5.9 billion. Strip out XPeng's own US$200 million as well — money that moved from one pocket of a consolidated group to another — and the outside world put US$700 million into Dogotix in August 2026, US$600 million of it from institutions and US$100 million from two of the company's own officers.

For completeness: on the earnings call the same morning, He Xiaopeng described the round as "over USD 6.2 billion post-money valuation" while the written announcement says US$6.3 billion. The share counts in the announcement give US$6.344 billion. The difference is immaterial to the argument and is noted so that a reader comparing the two sources is not left wondering.

7. The redemption right, and the seven-year clock

Under the Dogotix shareholders' agreement, each investor is granted a redemption right: the investor "shall be entitled to require Dogotix, the Dogotix Material Subsidiaries or XPENG to redeem, purchase, cause to be purchased or otherwise acquire all or any portion" of its preferred shares if Dogotix fails to complete a qualified initial public offering within seven years of that investor's first closing, or on other triggering events.

Three consequences follow, and none of them is speculative.

The US$600 million is not permanent equity from XPeng's point of view. It is capital with a put attached, and the announcement names XPENG among the parties who may be required to honour it. The grant of the redemption rights was itself classified as a discloseable transaction, on the basis that the rights are treated as if exercised at the time of grant.

There is a clock. Seven years from the first closing of the investors' subscription. The agreement is dated 24 August 2026 and completion was still subject to conditions, so the outside date is not fixed to a day — but it lands in 2033, and it converts "build the world's most valuable humanoid robot company" from an ambition into a deadline with a financial consequence attached.

It sits alongside a stated benefit that reads the other way. Among the reasons XPeng gives for the transaction is that it "broadens the sources of capital available to fund the XPeng Robotics Business without recourse to the Company's own balance sheet." The redemption right is a recourse to the Company's own balance sheet, contingent on a listing that has not happened. Both sentences are in the same document. A reader is entitled to hold both.

None of this makes the structure unusual. Redemption rights are standard in Chinese growth rounds and the seven-year window is at the generous end. It does mean that describing the round purely as equity understates what XPeng has taken on.

8. What the line has to produce

The line is where the argument stops being about paper. So what does the field actually ship?

The gap between the two ends of that chart is about three orders of magnitude, which is why the axis is logarithmic — on a linear scale the bar for units that genuinely left a factory would not be visible at all.

The one audited figure on it is UBTech's. In the six months to 30 June 2026 UBTech sold 921 full-size humanoid robots and booked RMB 590.3 million for them, up 1,445% year on year and 46.5% of group revenue. That is RMB 641,000 a unit, about US$95,000. It is the only per-unit price in the industry that comes from a listed company's interim report rather than from a spokesperson, and even at that price and that growth UBTech lost RMB 338.8 million in the half. Unitree is the volume leader — it has said it shipped more than 5,500 humanoids in 2025, the year its revenue across all products reached RMB 1.71 billion — and it listed in Shanghai in August 2026. A research house, Smart Analytics Global, puts the whole world's first-half 2026 humanoid shipments at 19,100 units; that is an estimate, not a disclosure, and should be read as one.

Against that, XPeng's stated ambition. On the second-quarter call He Xiaopeng said scaled production would begin by the end of 2026 into XPeng's own stores, that external commercialisation follows in 2027, and that "starting from the mid and second half of next year… the mass production units will be at multiple of several thousand units per month." Three thousand a month is eighteen thousand in a half-year — on its own, roughly what the entire industry is estimated to have shipped in the first half of 2026.

He has given no price. Asked directly on the call about unit cost and pricing, he answered with a market convention rather than a number: robots are "generally speaking… about 2.5 to 3x of the [BOM] material," more than 85% of the robot supply chain overlaps with XPeng's automotive suppliers, and he expects hardware gross margin better than the car business and lifetime revenue and gross profit per IRON "substantially higher than the current average selling price and gross profit per vehicle." Those are company expectations, offered as such. They are not guidance and there is no disclosed price to test them against.

So the second curve is straightforward. The financing has put a US$6.3 billion mark and a seven-year listing clock on a business whose entire industry, on the best available estimate, shipped 19,100 machines in six months — against the 103,295 cars XPeng delivered in the June quarter alone. The line commissioned on 7 September is the first physical evidence that the smaller number can start moving.

9. What the retained stake is worth

At the round price, XPeng's retained interest in Dogotix is worth US$4.69 billion — and it is worth that on either basis, because the warrants are struck at the round price, so exercising them changes XPeng's percentage without changing the value of what it holds. 73.80% of US$6.36 billion and 68.41% of US$6.86 billion are the same US$4.69 billion. The dilution that costs XPeng value is the incentive plan, not the warrants.

XPeng's ADSs closed at $10.95 on 4 September 2026, the last session before the US Labor Day holiday. On the 956,367,190 diluted ADSs in the second-quarter release, that is a market capitalisation of about US$10.5 billion. The August mark on the robot stake is therefore equivalent to roughly 45% of XPeng's entire market capitalisation.

Three things stop that from being a valuation argument, and they should be said plainly:

• Dogotix stays consolidated. Its results run through XPeng's accounts and its value is already inside the share price in whatever proportion the market chooses. This is not a separately tradeable asset and nobody has to mark it.
• The US$6.3 billion is a private mark, set in a round where US$200 million came from XPeng itself and US$100 million from two of its own officers. US$600 million of it is genuinely third-party price discovery.
• The outside US$600 million ranks ahead of XPeng's ordinary shares on liquidation, and carries the redemption right.

One note on our own model, and then we will leave it alone: the forward model we publish for XPeng carries two verticals, vehicle sales and services-and-others, and states in its own text that there is no robotics or robotaxi revenue line anywhere in it. There is no robot price, no robot volume and no unit economics in it to reconcile against, because none of those exist in public. The US$4.69 billion has nowhere to land, and inventing a line so that it could would be the opposite of useful.

10. Implications

In descending order of how confident we are:

1. The disclosure is complete and the summary of it is not. Every figure in this piece comes from one announcement XPeng published in full. "Raised $900 million at a $6.3 billion valuation" is true and conveys something materially different from what the document says.
2. The incentive pool, at US$952 million, is the largest equity item in the transaction. Half a billion of it comes out of XPeng's own holding by transfer. It is also the item most likely to move: none of it has been granted, and the assumption of full utilisation is what produces the 68.41% and part of the US$6.3 billion.
3. The executives' warrants are the cleanest asymmetry in the structure. Identical share counts, identical price, one side paying now and one side choosing later — and the larger half of the executive block belongs to the man who has run the division as its chief executive since June. The board discloses the offset it relies on; the option value itself is not quantified anywhere, by them or by us.
4. The redemption right converts part of the round into a contingent obligation, and starts a seven-year clock on a Dogotix listing that runs to 2033.
5. The production line is the first thing in this story that can be verified from outside. Volumes, unlike valuations, eventually show up in someone's revenue line.

What would settle each of these is specific. XPeng's third-quarter report, due 16 November, is the first set of accounts to carry Dogotix as a non-wholly-owned subsidiary, and the first place a non-controlling interest and any accounting for the redeemable preferred should appear. The four-month window for the additional investor's US$15 million closes on 24 December 2026. The first grants under the 2026 plan will show whether the 15% mandate is a ceiling or a plan. And the first IRON delivered to a customer outside XPeng, promised for 2027, is the number that makes all of the above either cheap or expensive.

Primary sources worth reading

• XPeng's 24 August 2026 announcement of the Dogotix subscription, redemption rights and 2026 equity incentive plan — the source of every share count, percentage and price in this piece, including the shareholding table.
• The 7 September 2026 release commissioning the humanoid production lines.
• XPeng's own pressroom note on the robotics financing.
• "IRON, the first general-purpose humanoid robot, rolls off XPeng's new production line", XPeng's newsroom, dated 8 September 2026 — the company's own account of the commissioning, and the source of the 80% automation, the 76/21 degrees of freedom, the 2,250 TOPS and the end-2026 mass-production and 2027 delivery dates. It puts the raise at over $900 million and the post-money above $6.3 billion, matching the filing.
• XPeng Global's video of IRON walking off the line, embedded above.
• Our capture of XPeng's second-quarter 2026 results and call, which carries He Xiaopeng's production and pricing remarks in full, and the sentence in which he confirms he has run the robotics business as its chief executive since June 2026.
• UBTech's first-half 2026 interim results, published 28 August 2026 on the Hong Kong exchange's filings site — the only place a listed humanoid maker publishes units sold and revenue per unit.
• Figure's description of BotQ and Agility's on RoboFab, for what a stated humanoid factory capacity normally looks like.

XPeng's ticker page is at /stocks/xpev/. Our other piece today, on China's August car exports and XPeng's overseas run-rate, covers the vehicle business the robot unit is being carved out of.

Every share count, subscription, warrant term, valuation and shareholding percentage here is disclosed in XPeng's 24 August 2026 announcement of the Dogotix subscription, redemption rights and equity incentive plan, linked above; that announcement prints its own cap table and the figures in this piece are read off it. The production-line claims, the automation percentage, IRON's specifications and He Xiaopeng's quote are from XPeng's 7 September release; his production, capacity and pricing remarks are from the second-quarter 2026 call of 24 August, and are company expectations rather than guidance; so is his statement that he has run the robotics business since June 2026. Derived here: the $2.026852 price per share and every dollar figure computed from it, including the incentive pool, the shares transferred out of XPeng's own holding, the paid-for subtotal and the retained stake. UBTech's units and revenue are from its first-half 2026 interim report and Unitree's from its 2025 disclosures; the 19,100-unit world figure is a research-firm estimate; the Tesla, Figure and Agility numbers are those companies' stated design capacities, not output. The $10.95 close is the 4 September 2026 NYSE session and the $11.15 close the 24 August one.