Goldman’s Moe Says He’s in ‘Stronger-for-Longer Camp’ on AI

Goldman Sachs APAC equity strategist Timothy Moe remains in the “stronger-for-longer” AI camp, arguing that AI-related stocks can stay attractive despite rising government bond yields. He cited hyperscaler investment of roughly $800 billion this year, rising to about $1.2 trillion in 2027, as a major demand catalyst for Asia’s AI hardware supply chain.

Analysis

The investable implication is not broad AI-beta but continued pricing power at bottlenecks: TSMC (TSM) in leading-edge foundry, SK Hynix (000660 KS) in HBM, and advanced packaging/equipment beneficiaries such as ASE Technology (ASX) and ASML (ASML). Incremental infrastructure spend has a high semiconductor-content intensity, but the economics are uneven: GPU/server assemblers face competitive pass-through risk, while scarce process capacity can sustain gross-margin upside. This favors component and manufacturing chokepoints over diversified hardware OEMs such as Quanta (2382 TT) and Wistron (3231 TT) after their initial order-book repricing.

Higher real yields are not uniformly negative for the complex. Names with near-term earnings revisions driven by contracted capacity and supply shortages should withstand multiple compression better than long-duration software or pre-revenue AI proxies; the key relative trade is AI infrastructure versus unprofitable software, not AI versus value. Over the next 1-3 months, quarterly capex guidance from Microsoft, Alphabet, Meta, Amazon and Oracle will determine whether supplier estimates keep rising; a deceleration in the rate of capex growth, rather than an outright cut, would be enough to compress the most crowded hardware valuations.

Consensus may be underestimating the second-order constraint: power availability, networking, and memory bandwidth can delay server deployment even if accelerator demand remains strong. That supports HBM and networking exposure, but creates inventory risk for lower-value server components if hyperscalers defer installations. Over 6-18 months, custom ASIC adoption by hyperscalers is the principal risk to merchant-GPU concentration; it would likely redirect, rather than eliminate, demand toward TSMC and advanced packaging while pressuring NVIDIA-linked system suppliers.

GS is not a clean expression of this theme: AI enthusiasm may support capital-markets activity and equity-financing volumes, but that is secondary to macro, trading conditions, and deal flow. There is insufficient company-specific information here to alter a GS position.

Key Decisions for Investors

• Prefer a 3-6 month long TSM / short QQQ pair: TSM has direct exposure to leading-node and advanced-packaging utilization, while the short leg hedges duration-driven technology multiple compression. Reassess if TSMC's next revenue outlook implies sub-20% year-on-year AI-related growth or if U.S. 10-year real yields sustain above 2.5%.
• Accumulate SK Hynix (000660 KS) on volatility rather than chase server OEMs; HBM supply qualification and tightness provide a clearer margin mechanism over the next 2-4 quarters. Thesis is invalidated by material HBM pricing concessions, Samsung qualification gains beyond expectations, or evidence of hyperscaler memory inventory buildup.
• Avoid adding to broad AI hardware baskets until the next hyperscaler reporting cycle confirms forward capex trajectories. Set an alert for two or more major hyperscalers guiding infrastructure capex below consensus; that would favor reducing ASX/TSM beta and rotating toward defensive semiconductors.
• For portfolios needing AI exposure with rate hedging, fund selective infrastructure longs with a short in high-multiple, cash-flow-light software via IGV or individual names with weak earnings revision breadth. Expected payoff is relative: bottleneck suppliers should retain earnings support even if sector multiples contract; cover if real yields fall materially and software revision breadth broadens.