China’s AI industry is moving from models to agents, a state report says

China Telecom Research Institute expects AI inference to account for 80% of China’s compute market by 2029, signaling a shift in the country’s AI industry from foundation-model competition toward agent deployment. Europe is pursuing seven AI gigafactories under a EUR 30bn budget, with bids closing in November and systems targeted to begin operating by mid-2028. The initiatives highlight escalating regional investment in AI compute infrastructure and strategic AI competitiveness.

Analysis

The investable implication of an inference-heavy AI cycle is a shift from episodic training-cluster purchases toward recurring spend on memory bandwidth, networking, power delivery and low-cost accelerator capacity. NVIDIA remains a beneficiary, but its premium is most vulnerable where hyperscalers substitute custom ASICs; Broadcom (AVGO) and Marvell (MRVL) have greater upside torque if inference workloads become sufficiently standardized for proprietary silicon. China-specific export restrictions increase the probability that local deployment favors domestic accelerators and foundry capacity, creating share-loss risk for U.S. GPU vendors that is not captured by headline AI demand growth.

For Chinese cloud platforms, agent adoption is not unambiguously bullish: inference is a variable cost that can rise faster than revenue when customers expect low-priced AI features. Alibaba (BABA/9988 HK), Baidu (BIDU/9888 HK) and Tencent (0700 HK) need evidence of enterprise ARPU expansion or reduced service costs; otherwise AI adoption can compress cloud margins through price competition. The first useful confirmation will be management disclosure on inference utilization, AI-cloud revenue mix and gross-margin progression over the next 1-3 quarters.

European compute infrastructure should be viewed as a multi-year procurement option rather than near-term earnings. The higher-probability beneficiaries after binding awards are electrical-distribution and thermal-management suppliers such as Eaton (ETN), Vertiv (VRT), Schneider Electric (SU FP) and Legrand (LR FP), but project financing, grid interconnection and sovereignty requirements can push revenue recognition well beyond equipment announcements. Contrarian risk is that publicly supported capacity is underutilized or procured from incumbent U.S. vendors, limiting the expected European technology-sovereignty premium.

Key Decisions for Investors

• Initiate a 1-3 month relative-value position long AVGO / short NVDA in equal dollar amounts only if upcoming hyperscaler commentary confirms accelerating custom-ASIC inference deployments; target 10-15% relative upside with a 5% relative stop. Falsifier: NVIDIA reports sustained data-center gross-margin expansion alongside evidence that inference demand is absorbing, rather than displacing, training GPU demand.
• Maintain a watch, not a position, on BABA, BIDU and 0700 HK into the next two earnings cycles. Upgrade selectively only where AI-cloud revenue growth exceeds total cloud growth without a gross-margin decline; absent that evidence, agent adoption is more likely a cost-pressure narrative than an earnings catalyst.
• Build a staged long basket in ETN and VRT after European facility awards become contractually funded, rather than ahead of them; use a 6-18 month horizon and limit initial sizing because grid and financing delays can defer orders. A material reduction in backlog growth or evidence of customer capex deferrals would invalidate the equipment-demand thesis.
• Avoid treating European compute commitments as a direct long in regional semiconductor champions. Until procurement rules, chip sourcing and utilization commitments are disclosed, the cleaner exposure is power and cooling infrastructure, whose content per deployed MW rises regardless of accelerator vendor.