--- title: "XPeng's US$1.3 billion R&D year is quietly turning a carmaker into a robot firm" date: 2026-11-09 category: Capital & Markets site: NeuroAI canonical: https://neuroai.site/a/na-capital-xpeng-physical-ai-capex language: en --- # XPeng's US$1.3 billion R&D year is quietly turning a carmaker into a robot firm > XPeng reported RMB 9.49 billion (US$1.36 billion) in 2025 R&D, with about RMB 4.5 billion aimed at AI; it is scaling Turing AI chips and an Iron humanoid toward mass production. Most car companies talk about self-driving. XPeng (小鹏汽车) is spending like a robotics company that happens to sell cars. In its 2025 results, the Guangzhou-based automaker reported a research bill bigger than its entire revenue was a few years ago — and pointed much of it at things that do not look like cars at all. ## The R&D number For fiscal year 2025, XPeng's **research and development expenses were RMB 9.49 billion (US$1.36 billion)**, up **47.0%** from RMB 6.46 billion the year before, according to the company's own unaudited results filed with the US SEC and posted on its investor-relations site. Of that, chairman and CEO He Xiaopeng (何小鹏) told investors that roughly **RMB 4.5 billion (≈ US$620 million)** went to AI-related development. For 2026, he said R&D aimed at "physical AI" (物理AI) would rise to about **RMB 7 billion (≈ US$965 million)**. - FY2025 R&D: RMB 9.49 billion (US$1.36 billion), +47.0% YoY - AI-related R&D in 2025: ~RMB 4.5 billion (≈ US$620 million) - Planned 2026 physical-AI R&D: ~RMB 7 billion (≈ US$965 million) ## Where the money goes: Turing chips The clearest output is the **Turing AI chip (图灵芯片)**, an in-house processor XPeng says can run in a car, a flying car and a humanoid robot. Mass production began in the third quarter of 2025. Cumulative shipments have already passed **200,000 units**, and the company has set a **2026 target of roughly 1 million units**. From the second quarter of 2026, XPeng plans to switch its entire model line to the self-developed chip. Notably, Volkswagen is the first external customer for the Turing chip and the second-generation VLA (Vision-Language-Action) driving stack — a sign XPeng is trying to turn its tech into a licensable asset, not just an internal cost. ## Where the money goes: the Iron humanoid The second bet is **IRON (IRON 人形机器人)**, XPeng's humanoid robot. He Xiaopeng has said mass production is targeted for **late 2026**, and the company began building a Guangzhou production facility in the first quarter with a **monthly output goal of more than 1,000 units by year-end**. IRON is designed to carry three Turing chips for roughly 2,250 TOPS of effective onboard compute, and XPeng says its VLA 2.0 stack already runs on the robot. The company has scoped IRON for commercial, industrial and home scenarios, starting with showroom guide and campus service roles before any consumer rollout. ## The business behind the bet This is not a cash-burning moonshot detached from the core business. XPeng's total revenue reached **RMB 76.72 billion (≈ US$10.6 billion)** in 2025, up 87.7%, and it posted its **first-ever quarterly net profit — RMB 380 million (≈ US$52 million) — in the fourth quarter**. Cash on hand at year-end was **RMB 47.66 billion (≈ US$6.57 billion)**, which He Xiaopeng cited as the cushion for sustained physical-AI investment. According to Li An, chief scientist at BrainNet (脑机网), an automaker with in-house chips and a humanoid program has a structural advantage because the same silicon and models can be reused across vehicles, robots and factories — if the company can keep shipping. The bet is coherent on paper: a carmaker with its own silicon and a humanoid program can amortize the same R&D across vehicles, robots and factories, while car profits fund the build-out. The open question is timing — humanoid mass production and the million-chip run rate are 2026 targets, not results, and the capital to hit them competes with the core auto business for every yuan. ## What readers can do now - Watch XPeng's quarterly chip-shipment figures: the jump from 200,000 to a ~1 million 2026 target is the single clearest proof point for its "physical AI" claim. - If you invest in EV or robotics names, treat Turing-chip external licensing (starting with Volkswagen) as the metric that separates a cost centre from a new revenue line. - Engineers should track whether IRON hits its late-2026 mass-production date and >1,000-unit monthly run rate before assuming humanoid scale-up is real. ## Honest limitations Core financial figures come from XPeng's official FY2025 results and SEC filings; forward-looking items (2026 R&D of RMB 7 billion, 1 million chip target, IRON late-2026 mass production) come from He Xiaopeng's earnings commentary and Chinese press coverage, not locked commitments. We have not verified independent benchmarks for the Turing chip's claimed compute, and "physical AI" is XPeng's framing, not an industry standard. Humanoid mass production at scale remains unproven across the industry, not just at XPeng. ## Key takeaways - XPeng spent RMB 9.49 billion (US$1.36 billion) on R&D in 2025, with ~RMB 4.5 billion aimed at AI and a planned RMB 7 billion (≈ US$965 million) physical-AI push in 2026. - Its Turing AI chip (图灵芯片) has shipped >200,000 units with a ~1 million 2026 target, and Volkswagen is its first external customer. - The IRON humanoid (IRON 人形机器人) is targeted for late-2026 mass production — a bet that turns an automaker into a robotics platform, if execution holds. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-capital-xpeng-physical-ai-capex Free to quote with attribution and a link to the original.