ChinaBiz Briefing | Huawei's AI Architecture Bet, CATL's Valuation Reset, XPeng's Robotaxi Launch, Manus's $500M Comeback
China's technology and capital markets delivered a dense slate of structural signals this week: Huawei staked a systems-level claim to AI infrastructure leadership; J.P. Morgan made its most forceful case yet that CATL's selloff is a mispricing, not a verdict; XPeng crossed the threshold from R&D to public robotaxi operations; a landmark IP deal between Huawei and Qualcomm formalized a new phase of patent-era competition; and Manus emerged from a Meta detour with the largest single check ever written into China's AI application layer. Taken together, the week's developments map the contours of a market in which supply constraints, brand equity, and software architecture are displacing raw chip specs as the decisive competitive variables.
• Huawei Declares Single-Chip Benchmarking Obsolete — and Claims China's AI Accelerator Crown
At Huawei Connect 2026, Rotating Chairman Xu Zhijun made the company's most pointed market-share assertion to date: Ascend has surpassed Nvidia in China's AI accelerator market on measurable data. The centerpiece announcements were the Peerium computing architecture and the UnifiedBus (UB)interconnect protocol — designed to eliminate the bandwidth cliff that currently separates intra-rack and inter-rack GPU clusters. Where Nvidia's NVLink-to-InfiniBand handoff drops bandwidth from 1.8 TB/s to roughly 0.2 TB/s at the rack boundary, UB traverses that boundary in approximately 150 nanoseconds with inter-rack bandwidth reaching 800 GB/s. More than 1,000 units of the 384-card 910C supernode configuration have already shipped.
Why it matters: China's frontier AI labs are preparing to train models in the 10-to-40-trillion parameter range — workloads requiring clusters of 200,000 to 256,000 cards. Whoever can deliver reliable, high-MFU supernodes at that scale owns a structurally enormous addressable market. Huawei's decision to publish Peerium as an academic paper and open UB to the industry — the largest download cohort post-release came from Silicon Valley — signals it is pursuing ecosystem adoption, not just domestic substitution. The binding constraint Xu identified is not technology but supply: "The key going forward is not how hard we push, but how much we can supply." The 960-series chip, incorporating HBM, is next on the roadmap. "Without progress in HBM, there is no 960" — a rare public acknowledgment of the supply dependency that sits beneath Huawei's training ambitions.
• J.P. Morgan Calls CATL's 19% Collapse a Structural Mispricing — With 49–79% Upside
CATL's A- and H-shares shed 19% in September 2026, dramatically underperforming tier-2 rivals Sunwoda (+28%) and Gotion (+11%), as the market priced in a "de-CATL-ization" thesis: OEM multi-sourcing and in-house battery programs would permanently erode dominance. J.P. Morgan's Rebecca Wen published a 60-page rebuttal on October 8, maintaining Overweight on both share classes with price targets of HK$725 (H-share, ~49% upside from HK$486.40) and RMB 520 (A-share, ~79% upside from RMB 291.11) by June 2027. The data contradicts the bear case: CATL's China passenger EV battery share rose three percentage points year-on-year to 44% in the first eight months of 2026 — and to 56% excluding BYD. Customer-level reversals at Nio, Geely, and GAC show the same pattern: tier-2 suppliers gain share on new nominations, then lose it back.
Why it matters: Three structural advantages are compounding simultaneously. First, OEM self-sufficiency is proving harder than advertised — Tesla's 4680 cells remain below 4% of its own consumption despite a 2020 target of 100 GWh by 2022. Second, consumer brand pull is measurable: 37% of Chinese EV buyers say they would reconsider a purchase if CATL batteries were absent, and CATL-equipped used vehicles command a 5–10% resale premium. Third, CATL's accounting conservatism — 100% R&D expensing, three-to-five year equipment depreciation, 3%+ warranty provisioning — means reported earnings materially understate economic earnings power; J.P. Morgan estimates reported net profit would be approximately 30% higher under tier-2-equivalent accounting. At 11.5x 2027 estimated P/E — below the January 2024 trough — with 20%+ annual EPS growth, 25%+ ROE, and 8–10% FCF yield, the stock is being valued as a cyclical manufacturer at a moment its fundamentals increasingly resemble a cash-return compounder.
• XPeng Opens YOYO Robotaxi to Public Users — Without HD Maps or LiDAR
XPeng formally branded its robotaxi service XPENG YOYO and launched a public WeChat mini-program, marking the transition from closed-loop R&D to open commercial operations. The milestone timeline is compressed: production vehicles rolled off the line in May 2026; full-service-chain employee testing launched in July; Guangzhou driverless road-test qualification followed in August; a token-fare internal pilot (RMB 1 per ride) stress-tested payment infrastructure; public onboarding — gated by invitation codes — is now live. Each production vehicle runs four proprietary Turing AI chips delivering 3,000 TOPS of on-vehicle compute, operating on a map-free, LiDAR-free architecture built on XPeng's second-generation VLA large model.
Why it matters: The map-free architecture is the strategic differentiator. HD-map-dependent competitors must conduct labor-intensive mapping campaigns before each new city deployment; XPeng's model-generalization approach eliminates that dependency, enabling faster multi-city and cross-border scaling. CEO He Xiaopeng has positioned the company as a technology and platform provider — monetizing through hardware sales, software licensing, and commission-sharing with fleet operators — rather than a capital-intensive fleet owner. The model's validity hinges on two unproven variables: whether operating partners can achieve profitable unit economics on XPeng's platform, and whether commission revenue can sustain R&D investment at scale. Watch for: conversion rate of invitation-code users to regular riders, geographic corridor expansion in Guangzhou, and any announcement of a first-wave global fleet operating partner, particularly in Southeast Asia or the Middle East.
• Huawei-Qualcomm Cross-License Formalizes China's IP Graduation Moment
On October 5, Huawei and Qualcomm announced a mutual cross-licensing agreement covering 5G, AI, networking, and computing patents, plus a discrete acquisition by Qualcomm of select non-cellular U.S.-registered Huawei patents in AI and computing. The deal pushes Huawei's cumulative patent licensing revenue toward a projected $6.9 billion threshold. Regulatory clearance — potentially including CFIUS review — is required before full execution; a 12-to-18 month timeline to completion is a reasonable base case.
Why it matters: The dominant market misreading — that the deal restores Qualcomm chip supply to Huawei's handset division — is analytically unsound. Export control restrictions remain governed by BIS Entity List designations, a policy instrument entirely separate from private IP contracts. HiSilicon's Kirin series has re-established competitive parity in premium Android; reintroducing U.S. silicon dependency would reverse five years of deliberate supply-chain hardening. The structurally significant story is Huawei's IP monetization engine reaching escape velocity: a company locked out of Western semiconductor supply chains in 2020 is, by 2026, collecting royalties from the ecosystem that sought to isolate it. The August 2026 Wi-Fi 7 agreement with HP — $0.50 per device — illustrated the same model at smaller scale. Huawei now holds approximately 165,000 active authorized patents globally and signed 34 new licensing agreements in 2025 alone. The operative investor question is how quickly royalty income scales relative to hardware revenue, and whether that mix shift alters the risk profile of Chinese technology equities under sustained geopolitical pressure.
• Manus Raises $500M-Plus — China's Largest-Ever AI Application Investment — After Escaping Meta
Butterfly Effect, parent of AI agent platform Manus, closed a funding round exceeding $500 million on October 8, co-led by Boyu Capital and IDG Capital, with Tencent, Sequoia China, and Zhen Fund participating. Bloomberg had previously reported an implied valuation of approximately $4 billion — roughly double the price at which the founding team repurchased equity from Meta following a failed acquisition that began in late December 2025 and unraveled by September 1, 2026. Twenty-seven days after regaining independence, the company shipped Manus 2.0 and Cue, a new personal agent application. The $500 million-plus check arrived ten days later.
Why it matters: The round arrives as China's agent category has been rapidly consolidated by platform incumbents. Tencent's WorkBuddy, Alibaba's Qianwen Office, and ByteDance's Doubao Work — each integrated into existing super-app distribution channels — recorded more than 20 million combined monthly PC visits by June 2026. Manus's differentiation thesis rests on product philosophy and agent-reasoning depth rather than distribution leverage — attributes platform-centric organizations replicate slowly. The forthcoming China-specific product, currently in accelerated development with domestic foundation model partnerships in place, will constitute the real-world stress test. A $4 billion implied valuation for an application-layer startup — not a foundation model provider — directly challenges the mid-2026 compression narrative that had cooled venture appetite across the AI application stack. For the broader ecosystem, 17 open roles and a $500 million-plus deployment budget represent a meaningful capital injection into a segment that had grown quiet.
What to Watch Next
The week's developments converge on a single underlying dynamic: the competition axis in Chinese technology has shifted from capability announcements to execution at scale. Huawei's constraint is supply chain, not chip design; CATL's constraint is narrative, not fundamentals; XPeng's constraint is partner economics, not technical architecture; Manus's constraint is retention metrics, not capital. The next 90 days will test all four simultaneously — Huawei's 950DT training ramp, CATL's Q3 earnings and share recovery data, XPeng YOYO's first external user cohort metrics, and Manus's China product launch. Regulatory developments — CFIUS review of the Huawei-Qualcomm IP transfer and the EU's Industrial Action Plan localization framework — represent the principal exogenous variables.
Related Coverage:
Manus Parent Butterfly Effect Raises $500M+, in China’s Largest AI App Deal
Huawei-Qualcomm Patent Pact Signals China's Rising Royalty Power
XPENG Launches YOYO Robotaxi Brand in China’s Shift to Commercial Autonomy