Still Priced as a Car Company

Publisher's note. This is an excerpt from a longer research paper commissioned by an equities strategy firm, published here with permission. The commissioning firm is not named at its request. Passages that would identify positioning, the full sum-of-the-parts build, or client-specific assumptions have been removed. A black bar marks each cut, with a short note saying what was withheld. The underlying paper carries a formal rating and price target; both are withheld from this excerpt. Nothing here is a recommendation. See the notice at the foot of this page.

XPeng is still being discussed, and still largely being valued, as a Chinese electric-vehicle maker in a brutal price war. That description was fair for several years but is now incomplete. The company that shipped the Mona M03 into the heart of China's volume market, put an in-house AI chip into cars priced like appliances, and is now testing a vision-language-action driving stack on European streets, has spent eighteen months turning itself into something harder to compare: a physical-AI company that sells cars, and uses the car to pay for the body.

The listed equity does not, on the evidence in this excerpt, reflect that second business with much seriousness. In late August 2026, outside investors marked the robotics subsidiary at more than US$6.3 billion after a financing of over US$900 million. On 7 October 2026 the New York listing closed at US$9.58, a market capitalization of about US$9.2 billion, 66 percent below the November 2025 closing high of US$28.07. One of those numbers can be wrong, and both can be early, but they cannot both be a calm description of the same set of assets.

Executive summary

XPeng's nearest equivalent to Tesla's Full Self-Driving (Supervised) is not a single feature. It is a stack: XNGP, the navigation-guided pilot that reached nationwide urban coverage in China, now running on VLA 2.0, a second-generation vision-language-action model that maps camera input to driving action without a language bottleneck in the middle. The company ships it on cars, not as a perpetual beta confined to one geography's highways, and it is in open-road testing in Europe ahead of a targeted 2027 release.

The car business is what makes the claim bankable. Mona M03, launched at a starting price of RMB 119,800, delivered roughly 176,000 units in 2025 and remained the volume anchor through the first half of 2026. Around it, a broader cycle – P7+, G6, the new L03, the flagship G9L, the GX – has pushed the company up the price ladder while overseas revenue reached about a quarter of the first-half 2026 total. Second-quarter gross margin was 20.7 percent on RMB 19.74 billion of revenue. That blended figure needs a qualification, set out below. It is still a long way from the cash-burning narrative the stock sometimes trades on.

The larger catalyst is IRON. The next-generation humanoid – 76 degrees of freedom in the body, 21 in each hand, three in-house Turing chips, up to 2,250 TOPS on device – walked off a newly commissioned Guangzhou line under its own power on 8 September 2026. Management is targeting internal deployment in stores and on campuses by year-end, and commercial deliveries in China and overseas in 2027. Those are company targets rather than orders, but they have moved past the slide deck.

He Xiaopeng took the robotics chief executive role himself in June 2026 and bought into the round. That detail matters more than the gait video. A founder who already sold one company to Alibaba, who has spent eight years on this machine, and who describes humanoid robotics as roughly twenty times harder than cars, is not running a side project.

The excerpt's conclusion is narrow: the listed shares look undervalued relative to the car franchise alone, and more so once a controlled robotics subsidiary carrying a third-party mark is admitted into the frame. The bridge from that judgment to a number is in the full paper. Withheld: scenario weights, the valuation build, and the commissioning firm's position limits.

A founder who stayed on the tools

He Xiaopeng does not present as a financial engineer who found a theme. He built UCWeb, the mobile browser, and sold it to Alibaba; he then co-founded XPeng in Guangzhou and has remained chairman and chief executive through the cycle that nearly broke several of his peers. The operating signature is vertical integration of the practical kind: chip, operating system, driving model, vehicle, and now actuator and robot line, rather than a brand wrapped around a supplier stack.

In an internal letter dated 10 June 2026 he added a second job, chief executive of the robotics business, and described the shift plainly: from an intelligent-automobile company to a physical-AI company. The following quarter he and co-president Brian Gu committed personal capital alongside the outside round. Related-party participation is not proof of value, though a refusal at this stage would have said something too.

The stagecraft is better known than the organization chart. At AI Day in November 2025 the next-generation IRON walked with a gait so human that viewers decided there was a person inside. He had the lower leg cut open on stage and walked the machine with the mechanism exposed. It was good theater. It was also a useful tell: the company is sensitive to the accusation that the robot is a costume, and it has organized the subsequent year around making that accusation harder to sustain – a line, a financing, a named production cadence, and a chief executive whose other job is the car company.

The driving stack, without the slogan

Tesla's Full Self-Driving (Supervised) is the reference product in every serious conversation about urban assisted driving, and XPeng has chosen not to dodge the comparison. Its system of record is XNGP – navigation-guided pilot – extended from highways to city streets and, on the company's account, to nationwide coverage in China by the middle of 2025. The current software generation is VLA 2.0.

The technical difference is straightforward. First-generation vision-language-action models translated what the cameras saw into language-like tokens, then translated those tokens into an action. That intermediate step is legible and lossy. VLA 2.0 is built to go from vision to action more directly, which is the direction Tesla has taken with its end-to-end driving models, and the direction XPeng uses for IRON. Car and robot are different products that increasingly share one research program.

Hardware sits underneath. The in-house Turing AI chip, a 5-nanometer part the company rates at roughly 700 TOPS, is the compute XPeng intends to own rather than rent indefinitely. Cumulative shipments had passed 200,000 units by the March 2026 earnings call. On the cars that matter for the claim – the Ultra configurations of models such as the L03, and the higher trims spreading through the range – two or three Turing chips are allocated to driving and cabin, quoted at up to 2,250 TOPS combined. The 2026 Mona M03 refresh put a Turing chip into a car that still starts at RMB 119,800. That is the strategic point, more than any benchmark: high-compute assisted driving is being treated as a volume feature, not a flagship option.

What the system is not: a driverless robotaxi network. XPeng has shown robotaxi concepts. The product customers use is supervised assistance, with a person responsible for the car. Claims of parity with particular FSD versions should be read as engineering boasts until they are fleet statistics. Their specificity is what makes them testable. A ride report in April 2026 described a faultless urban run of more than forty minutes in Chinese traffic; a September comparison in Amsterdam – VLA 2.0 in an L03 prototype, then FSD on the same roads – found a system that already drove Dutch streets confidently and was still learning local custom. Europe is scheduled for public release in 2027, regulation permitting.

Stated more narrowly than the keynote, the case still holds. XPeng has a shipping urban assistance stack, on its own silicon, trained in one of the messier driving environments on earth, with a credible path into markets where Tesla's supervised system is only now being allowed onto the road. Whether it is "ahead" is the wrong argument for an equity. Whether it is real, and whether it compounds through the fleet, is the right one. On the evidence available outside the company, it is real.

The model cycle that made the stack affordable

A driving story without a car story is a demo. XPeng's last two years are a product cycle.

Mona M03 is the proof of distribution. It delivered 175,689 units in 2025, about 41 percent of the company total, at a price starting just under RMB 120,000 – roughly US$17,000. Through the first half of 2026 it was still the best-selling nameplate. The 2026 update kept the entry price and put the Turing chip into the upper trims. That is how a software and silicon advantage reaches a market that will not pay a luxury premium to try it.

Above Mona, the range has filled in rather than been replaced. The P7+ sedan and the G6 SUV remain material contributors. The L03, a new coupe-SUV, broke the company's own order record at launch. The G9L, described by the company as its next AI flagship SUV, launched in China in September 2026 and is the centerpiece of the Paris Motor Show reveal on 12 October, with a first European production trial already completed.

Two years ago the overseas mix would have been hard to underwrite. In the second quarter of 2026, deliveries outside China exceeded 20,000 units, up 81 percent year on year, at an average selling price above €40,000. Overseas markets contributed 25 percent of first-half revenue. September was the first month overseas deliveries cleared 10,000. A Chinese EV brand that can sell a €40,000 car in volume outside China is no longer only a domestic price-war combatant.

From Aberdeen to Abu Dhabi

The second demand pool already has a street presence. An XPeng is now a common sight in cities from Aberdeen to Abu Dhabi – a G6 on a North Sea commute, a G9 outside a Dubai mall, a P7+ in a showroom that did not exist eighteen months ago. A network build looks like this once it is working.

Europe is the volume proof. By early October 2026 XPENG had delivered more than 100,000 vehicles overseas in total, of which more than 60,000 were in Europe. Local manufacturing has moved past press-release intention: G6, G9 and P7+ are in production at Magna's plant in Graz, and on 7 October the G9L completed its first European trial build there, the fourth nameplate through the same facility. Order books open with the Paris premiere on 12 October, across a launch planned for 64 markets.

The Gulf deserves more attention than another European price comparison, because it is being opened in the right order. XPENG entered the UAE in October 2024 with Ali & Sons, through Gulf Star Motors. A year later it opened a flagship on Al Salam Tower in Abu Dhabi and a regional parts warehouse in Dubai with JD Logistics – the unglamorous asset that decides whether a premium EV brand survives its second summer. Qatar followed with Pioneer Motors in December 2025; Bahrain in July 2026 with the Ebrahim K. Kanoo group; Oman, signed in September 2026, is the fourth GCC market. Saudi Arabia and Kuwait are the two still to be announced.

Country-level delivery figures for the GCC are not disclosed separately, so no percentage is given here. What is disclosed is enough. Overseas deliveries in 2025 were 45,008, up 96 percent, and the 2026 plan was to double them, lift overseas revenue above 20 percent – already 25 percent at the half – and double the sales and service network to 680 stores. The GX drew more than 1,000 overseas blind orders after its Beijing debut, most from the Middle East, named the first region outside China for pre-sales.

The hiring matters as much as the hardware. XPENG has not tried to staff Europe or the Gulf with a travelling cadre from Guangzhou. It has hired regional operators and paired them with groups that already know how to register, service and finance a car in their own market: Ali & Sons in the UAE, Almana in Qatar, Kanoo in Bahrain, Emil Frey in Germany. That is the correct shape of an international expansion. The cars are good enough to be wanted. The teams being built around them will decide whether Aberdeen and Abu Dhabi are anecdotes or a channel.

Margins warrant the same plain reading. Group gross margin of 20.7 percent in the second quarter is the headline, and a genuine improvement. Vehicle margin was lower – the company reported 12.1 percent – with a higher-margin service line, including collaboration income, doing a meaningful share of the blending work. Withheld: the separation of collaboration income, vehicle contribution margin by nameplate, and the sensitivity of 2027 margins to Mona mix. None of that turns the quarter into a mirage. It does mean the constructive case should not be written as if the car business already earns a software company's gross margin. It earns an improving manufacturer's margin, with a second income stream on top.

The Volkswagen relationship belongs in this section and not in the robot section. A global incumbent choosing XPeng as a technical partner for China E/E architecture and jointly developed models is external evidence that the stack is licensable, not merely demonstrable. Withheld: partnership economics and the residual value assigned to further platform licensing.

IRON, and why it can matter more than the next model year

The robot is where the valuation argument changes character.

IRON, in the specification XPENG now repeats consistently, has 76 degrees of freedom across the body and 21 in each hand, a fully enclosed flexible lattice structure the company describes as synthetic skin over bionic structure, and three Turing chips delivering up to 2,250 TOPS. It runs the same physical-world model family as the car – vision, language, action – so a movement can be generated from what the machine sees rather than from a teleoperator's script. The November 2025 reveal quoted a higher degree-of-freedom count; the production-intent figure used since the August financing is the one above. The later figure is used throughout.

On 8 September 2026 the company commissioned the Guangzhou line and had the first machine walk off it. More than 80 percent of core processes are automated, on XPENG's account. Initial commercial use is modest and, for that reason, believable: guides and service roles in the company's own stores and campuses before the end of 2026, then an official launch and external deliveries in China and overseas in 2027. Capacity language has ranged from more than 1,000 units a month by the end of 2026 to several thousand a month in 2027, depending on demand.

The financing is the fact that does not depend on a gait video. In August 2026 the robotics business agreed a round of more than US$900 million at a post-money valuation above US$6.3 billion, led by IDG Capital, with Alibaba, Tencent and Gaorong among the investors. XPENG subscribed a further US$200 million and kept control. He Xiaopeng and Brian Gu committed US$100 million between them. The company called it the largest single-round private financing in China's embodied-AI industry. Even after the usual discounts for a mark that is not a public price, a controlled subsidiary that sophisticated investors will fund at that level is not a science project.

He has been unusually direct about difficulty. He has said building humanoid robots is on the order of twenty times harder than building cars, and that a visible step-up in capability should show from the fourth quarter of 2026 into the first half of 2027. That is a measured way to put it. The failure mode for this program is not a missed demo. It is a machine that cannot do useful work at a cost a store, a campus, or eventually a household will pay. XPENG's answer, so far, is automotive discipline applied to a new bill of materials: shared chips, shared models, a line designed by people who already run car plants, and a chief executive who has made the robot his second full-time job.

What the price is, and what it leaves out

The dislocation is now visible in the share price.

On 7 October 2026, XPeng's New York listing (NYSE: XPEV) closed at US$9.58, a market capitalization of about US$9.2 billion, sitting on a 52-week low of US$9.18 set on 2 October against a 52-week high of US$28.23. The last closing high of the cycle was US$28.07 on 11 November 2025, so the stock is 66 percent below that print. The Hong Kong line (9868.HK) was around HK$37.50 on 8 October, against a 52-week high of HK$110.80.

A 25-analyst consensus compiled by S&P Global, as of that same close, carried a Buy rating and an average target of US$18.40 – 92 percent above the New York price – with a range from about US$11.50 to about US$25. Consensus is a data point rather than a thesis. It shows the sell side has noticed the derating and has not closed the argument.

Read the chart against the operating record, not the last headline. Between the November 2025 high and this October low, the company put VLA 2.0 into customers' hands, crossed 20 percent group gross margin, took overseas revenue to a quarter of the half-year, opened the Gulf beyond the UAE, started European production at Graz, commissioned the IRON line, and raised more than US$900 million for the robot at a US$6.3 billion mark. The share price treated that sequence as a deterioration.

The numbers here are the awkward part. A robotics subsidiary marked above US$6 billion, inside a parent marked near US$9 billion, implies a thin residual for the car business, the chip program, the driving stack, the European and GCC ramp, and the Volkswagen collaboration combined. At US$9.58 the market is paying roughly 0.8 times trailing sales for the entire group. Withheld: the allocated value for the vehicle franchise, the software and silicon attach, and the discount applied to the robotics mark for control, illiquidity and related-party participation. The full paper also shows the residual if the subsidiary is valued at zero.

That residual is where the excerpt locates the opportunity, and it is mislabeled. Even a sober car-only reading has to account for a brand that delivered 429,445 vehicles in 2025, a rising export mix at European prices, an in-house chip with a path to a million units a year, and a group gross margin back through 20 percent. The robot is the option on top of that case, not the case itself. Options of this kind are usually carried at zero until a production line exists, and one now does.

What would make this wrong

A constructive reading still has to name the ways it fails.

The China price war is not over, and Mona's volume is a double-edged asset: it proves reach, and it can dilute mix if the upper range stalls. Vehicle margin at 12.1 percent has further to go. Net profitability has not been durable. Exporting the driving stack is a regulatory problem as much as a technical one; a 2027 Europe date can slip without anything on the model being false. IRON's first jobs are internal. A store guide is not a revenue line, and a 2030 volume ambition should be ignored until there is a customer who is not XPENG.

The subsidiary mark can also be wrong. Private rounds in a fashionable category overshoot. Control provisions, redemption terms, and the fact that the parent is both seller and buyer of the story all argue for a haircut. Withheld: the haircut schedule used in the paper, and the downside case in which the robotics stake is written to the parent's cost.

None of those caveats restore the old description. They bound it. XPeng can be a good car company with a promising robot and still be a disappointing stock, if the car cycle rolls over before the robot earns a yuan. Both possibilities have to be held at once.

The line we would actually quote

XPeng has earned the right to be talked about as more than an electric-vehicle ticker that had a difficult 2023. The driving stack is in customers' hands. The chip is in the volume car. The export business is large enough to move a half-year. The robot has a line, a financing, and a founder standing on it.

The market is still pricing the first of those facts and treating the last as a video. That gap is the subject of this research.