〔Day Digest〕Embodied AI Crosses Mass Production: XPeng Locks $900M+ at $6.3B, Wujie Ships First France Batch, While Nvidia's $95B/$108B Print Hangs on Rubin's 12% Mix - China Export Charges as the Tail

0. Daily Arc

Tomorrow's Nvidia print is the week's gravity well: Jefferies reads $95B for July-quarter revenue (consensus $91.9B) and $108B October-quarter guide (consensus $103.7B), with Rubin at ~12% of F3Q mix rising past 40% by F4Q as the OpenAI 12GW (extendable to 16GW) anchor gets tested ^[1]. Yet the cost-curve and embodied-AI stories are running in parallel - XPeng's IRON robot unit locked $900M+ at a $6.3B post-money ^[2], Wujie Power shipped its first batch to France ^[3], and OpenAI cut GPT-5.6 Sol API by 20%+ on 8/21 with Alibaba's AI Cloud and Compute Services at ¥48.4B (+45% YoY) ^[4]. In the risk column, Nvidia and Supermicro employees face AI-server export charges ^[5], Australia's ARIA banned primarily-AI songs this week ^[6], and Counterpoint projects 2026 global FPD shipments -4.7% YoY on memory spillover ^[7]. Net: a hardware story, an embodied-AI industrialisation story, and a sovereignty story converging on the same tape.

1. Nvidia's 8/26 Print and the Rubin Question

• **[NEW] Jefferies pre-print (single sell-side view):** $95B July-quarter revenue (vs. $91.9B consensus, +~16% q/q), $108B October-quarter guide (vs. $103.7B consensus, +~4%) ^[1]. Rubin (VR/R200) at ~12% of F3Q GPU revenue, ~40%+ in F4Q, with F1Q28 the Blackwell-to-Rubin dominance crossover at ~2M units shipped ^[1]. The two-tier call items: (1) Rubin ramp cadence, (2) the OpenAI 12GW (extendable to 16GW) commitment accounting treatment ^[1].
• **[NEW] The 8/17 SB Energy PORTS-Pike site in Ohio:** Nvidia is the sole compute supplier; OpenAI holds a 20-year lease; Nvidia has invested $1.5B in SB Energy ^[1]. Mechanism: Nvidia has become a "banker to the AI boom," not just a chip vendor ^[8].
• **[NEW] Moomoo frames the print as a 3-point test (single source / unverified):** beat-and-raise is no longer sufficient; mix, margins, and forward guidance carry the tape ^[9].
• **[ESCALATED] Supply-chain spillover:** Counterpoint projects 2026 global FPD shipments -4.7% YoY as AI-driven memory prices pass through to consumer devices ^[7]. Read-across: any commentary on data-center memory tightness on tomorrow's call gets magnified.

2. Embodied AI Goes Industrial

• **[NEW] XPeng's IRON robot:** mass production from year-end 2026, in-store/factory rollout in 2026, retail launch 2027, monthly capacity scalable to "thousands" next year ^[2]. Chairman He Xiaopeng now dual-CEO of auto and robot business, pulling the Turing AI chip and auto supply chain into the robotics unit ^[2]. Funding: $900M+ first round at $6.3B post-money; IDG Capital leads, Gaorong, Tencent, Alibaba subscribed $600M, XPeng and insiders $300M ^[2]. He frames the technical bar as "at least 20x smart cars" per unit in lifecycle revenue and gross margin ^[2].
• **[NEW] NIO spinout-within:** Ren Shaoqing (head of NIO smart driving) founded an independent company focused on physical-AI foundation models and embodied intelligence; NIO is strategic investor and partner; Ren retains full NIO smart-driving remit and role ^[10]. Mechanism: a defection that isn't - the parent stays anchor customer and minority investor, while world-model and reinforcement-learning research crosses from cars to robots ^[10].
• **[NEW] Wujie Power:** first batch of humanoid robots shipped to France; co-founder Wang Yingyun confirms Korea as the next target ^[3]. Customers are European auto-parts makers and international energy firms; ~¥700M in global orders to date, ¥500M from an Envision French battery-plant project ^[3]. The core team is largely ex-autonomous-driving ^[3].
• **[ESCALATED] UBTech:** industrial humanoids deployed in stamping, machine-tending, depalletising, and e-commerce sortation at the 2026 World Robot Conference; new ¥50M Changzhou JV (Changzhou High-tech Shucai, UBTech, Daozhisen) to scale delivery and after-sales ^[11][12]. UBTech CGO Tan Min: the industry is now competing on "the brain and on systematic capability, not motion control" ^[11].
• **[NEW] Alte:** hexapod special-purpose robot concept released at NVIDIA's Physical AI & Robotics Partner Day as a Solution Advisor: Consultant; July JV with Foxconn on a Turn-Key platform for robot OEMs ^[13].
• **[ONGOING] Sports-as-stress-test:** the second World Humanoid Robot Games doubles as a public reliability test; Shoucheng Holdings capital-markets head Kang Yu reads race-day failures as engineering to-do lists, not binary wins ^[14]. Shoucheng's robotics book: Unitree, Galbot, Accelerated Evolution, Stardust Intelligence, Robotera, Songyan Dynamics ^[14].

3. Regulation, Sovereignty, and the Export Tail

• **[NEW] US export-control enforcement (Al Jazeera):** Nvidia and Supermicro employees charged for AI server exports to China ^[5]. Read this against Nvidia's banker-role pivot and the OpenAI 12GW anchor ^[8][5][1].
• **[NEW] Australia's ARIA:** from this week, the singles chart bans songs "primarily AI-created"; fully AI-generated tracks are removed, AI-as-tool remains eligible; ARIA Awards follow ^[6]. Sweden moved first; others are studying similar rules ^[6]. A University of Sydney expert calls the move "a necessary step," with cleaner IP frameworks the unfinished business ^[6].
• **[NEW] Agent-payment self-regulation (Caixin):** new industry pact requires KYA (Know-Your-Agent) and explicit user authorisation for AI agent payments ^[15]. Mechanism: governance catching up to agent autonomy.
• **[NEW] The 1.25B-euro Porsche-TCS deal (5 years):** TCS runs a dedicated AI centre for Porsche; Porsche transfers its IT consultancy arm (~€320M) to TCS ^[16]. Enterprise AI is now a balance-sheet line, not a SaaS line item.

4. What Would Falsify It

• A soft Nvidia October guide (vs. Jefferies' $108B) or a sub-12% Rubin mix in F3Q collapses the ramp thesis and the OpenAI 12GW accounting premium ^[1]. Watch the F4Q Rubin commentary, not the headline beat ^[9][1].
• An embodied-AI production slip: any 2026 XPeng IRON timeline push-back, a Wujie Power delivery miss, or a UBTech after-sales miss resets the demo-to-production arc ^[2][3][11].
• Source quality control: the Nvidia banker-role framing is single-source WSJ ^[8]; the Rubin mix percentages are a single sell-side read ^[1]; the OpenAI ~20M Codex user figure, the +60% three-month reasoning-speed gain, the "Ultra Fast" 14x mode, and the Luna -80% inference-cost claim are a single podcast by Codex head Tibo Sottiaux to Matthew Berman ^[17]; Sam Altman's "two errors" admission is a single Founders podcast transcript with David Senra ^[18]; the ARIA expert quote is a single Chinese-media relay ^[6]. Treat the line-up as a chorus, not a consensus.