Prepared for the Guangzhou HQ visit · 17 Sep 2026
Every vehicle XPeng sells, every physical-AI programme it has put into production, and the autonomy stack underneath all of it — with the prices, the volumes and the dates as they actually stand. Prices converted at USD/CNY 6.70.
01 — Passenger vehicles on sale or launching in 2026
Eleven nameplates across three tiers: MONA at RMB 120–160k, the G/P core at RMB 175–300k, and the GX / X9 flagships above RMB 270k. Four of these are brand-new in 2026 and are what management is counting on to get from 39k a month to 60k. Prices are current China MSRP before the optional Ultra SE (+RMB 12,000) and Ultra (+RMB 20,000) ADAS packs.
The car that saved the company. China's best-selling A-class BEV sedan for 19 consecutive months, 100,000 deliveries in its first eight months, and >85% of buyers take the Max or Ultra SE trim — i.e. they pay up for the Turing chip. This is He Xiaopeng's "technology equality" proof point.
Record order book for any XPeng model and the single biggest swing factor in Q4. Launched RMB 20,000 below its own pre-sale price. Currently supply-constrained — on double shifts, with management blaming extreme weather and supply chain. Deutsche Bank models ~12,500/month. Overseas deliveries start Q4. First car anywhere with the Google Maps Auto SDK in APAC.
The third MONA, and the last new nameplate before the 60k/month target has to be hit. Two further MONA models are reported to follow, taking the budget brand to at least five vehicles including a station wagon.
Core European volume model alongside the G9 — both are built locally at Magna Steyr in Graz, Austria for the EU market.
30,000 firm orders on launch night; 50,000th unit built five months later. Ranked #1 pure-electric sedan in the RMB 150–200k segment. First model to standardise the AI Hawkeye pure-vision ADAS across every trim — no LiDAR, no HD maps.
The first XPeng shipped with the in-house Turing SoC. In its first four weeks the Ultra trim was over 50% of sales — the clearest evidence that buyers will pay for compute. This is the car that made the "AI is the differentiator" argument commercially, not rhetorically.
10,000 firm orders in the first seven minutes. Priced RMB 80,000 below the RMB 300k point management had been signalling. Top-3 BEV sedan in the RMB 200–300k band and the car that pushed monthly deliveries above 40,000 from September 2025.
The premium anchor in Europe, built at Graz. Norway was XPeng's first overseas market (first store, end-2020) and the Nordics remain the strongest per-capita territory.
The China launch event is 17 September 2026 — the same day as the HQ visit. Pre-sales opened 11 Aug with six versions across BEV and EREV. Aimed squarely at the Li L7 and the Huawei-backed AITO M7. Brian Gu said on 10 Sep he had spent two hours that morning testing the latest autonomy build in a G9L.
China's first factory-fitted robotaxi platform from an OEM, and the most important car in the story — it is both the flagship and the autonomy hardware reference. >80% of initial orders took the RMB 350k+ Ultra flagship trim; >7,000 domestic deliveries in July; top-3 in China's RMB 300k+ NEV SUV segment. Note it launched RMB 130,000 below its own pre-sale price.
The Kunpeng Super EREV X9 opened the "one vehicle, dual energy" era in Nov 2025. Pre-orders came in at nearly 3× the previous X9 on a like-for-like basis, concentrated in northern and inland China — exactly the BEV-resistant buyers the EREV strategy was built for.
02 — Deliveries, and the gap that has to close
2025 was the best year in XPeng's history. 2026 has been a step backwards on any year-over-year measure, and the fourth quarter carries an unusually large amount of weight.
Units · Q1 2024 – Q4 2026 · Q3 is company guidance midpoint, Q4 is 3 × the stated 60k monthly target
Units · September bar is the midpoint of what Q3 guidance requires after July and August
03 — The four programmes that share one stack
Same Turing silicon, same VLA model family, same data and training pipeline, largely the same supply chain. That reuse argument is the whole thesis — and the reason the robot option is cheaper for XPeng than for a standalone robotics company.
Humanoid robot
A general-purpose humanoid built to automotive-grade standards, with everything from the SoC to the dexterous hand designed in-house. Production lines in Guangzhou were commissioned on 8 September 2026 with over 80% core-process automation, and IRON walked off the line under its own control. Retail and service roles first — store sales assistant, tour guide, museum docent, supermarket checkout — then industrial, then household.
Autonomous mobility
Built on the GX platform with six-way full redundancy, no HD maps and no LiDAR — which is the entire generalisation-and-cost argument against Waymo, Baidu and Pony. XPeng self-operates in Guangzhou only, as a demonstration; everywhere else it sells the vehicle and stack to operating partners and takes a share of the revenue.
Capital structure
Announced 24 August 2026, the same day as Q2 results. XPeng calls it the largest single private round in China's embodied-AI industry. The unit remains 100% consolidated — a separation window exists but Brian Gu says the financials stay inside the group either way.
Silicon
The part that makes the reuse argument real: one chip family in the cars, the robotaxi and the robot. Volkswagen is the first external customer, which is the only third-party validation of the AI stack that comes with cash attached.
Ecosystem, not a segment
Formerly XPeng AeroHT. XPeng incubated it and is now a minority shareholder; it is not consolidated, so it belongs in the sum-of-the-parts, not the P&L. The "Land Aircraft Carrier" is a six-wheel mother ship carrying a detachable two-seat eVTOL.
04 — VLA 2.0, and where it actually stands against FSD
XPeng and Tesla are the only two companies shipping a single end-to-end model driving a production car. That is the comparison management invites, so it is the one worth making carefully — including where XPeng's own marketing overstates it.
Sensor stack
XPeng markets "pure vision" and it is broadly true for the perception model — the AI Hawkeye stack has run without LiDAR and without HD maps since the P7+ in November 2024, and VLA 2.0 dropped LiDAR entirely. But "pure vision" is a claim about the driving model, not about the bill of materials.
| Dimension | XPeng VLA 2.0 | Tesla FSD |
|---|---|---|
| Architecture | Single end-to-end vision-language-action model plus a VLM for cockpit/language. From v6.3.0 (late Aug 2026) the driving and cockpit models are merged. | Single end-to-end neural network. The only other production car shipping this architecture. |
| On-device model | v6.3.0 raises parameters 3.5×; target ~20bn by end-2026, up from ~1bn | Not disclosed |
| Sensors | Cameras only for perception; radar and ultrasonics still fitted for other functions. No LiDAR, no HD maps. | Cameras only. No radar, no LiDAR, no HD maps. |
| Silicon | In-house Turing, ~700–750 TOPS/chip, 3 chips on Ultra. TSMC-fabbed. | In-house AI4/AI5. Samsung and TSMC. |
| Independent verdict | InsideEVs, April 2026: smooth and confident, but still trails FSD overall — leans too hard on the lead car, reacts late to lane changes. Forbes was more positive on China-specific driving. | Still the reference. He Xiaopeng's stated target was to match FSD v14.2 in China by 30 Aug 2026. |
| Claimed edge | He Xiaopeng: better on narrow roads, urban-village lanes, parking structures and unmapped terrain. Says FSD does emergency stops where VLA flows. | Larger real-world fleet and a far longer data history; stronger on highway and structured urban. |
| Generalisation test | The single most valuable claim in the story: a model trained essentially only on Chinese data performed nearly as well on European roads, tested in Germany's Black Forest, with no European post-training. | Trained globally; China deployment has been repeatedly constrained by regulation and data localisation. |
| Monetisation | Currently free — bundled in the Ultra trim, no separate charge. Management has said it may charge later; the model is explicitly undecided. | Paid — one-time purchase or subscription. |
| Europe | Testing now; regulatory approval targeted H1 2027 under the DCAS framework. Country order not disclosed. | Long-delayed European approval, also pending under DCAS. |
| Third-party licensing | Volkswagen is the first external customer for both Turing and VLA 2.0. A group-level BD team was created in 2026 to pursue more OEMs and Tier 1s. | Repeatedly discussed, never signed at scale. |
| AI spend | ~$500m/year on AI training (per XPeng at CVPR 2026); RMB 7bn of physical-AI R&D budgeted for 2026, all expensed | Substantially larger |
05 — Before the 30-minute NGP test drive
Long-form POV drives are far more informative than any clipped demo. The Guangzhou city-streets run is the closest analogue to what the IR team will show you; the uncut European drive is the best independent test of the generalisation claim.