This Round of Sharp Decline: Technology Giants Clearly Oversold
The article argues that the recent market selloff—triggered by Democratic wins in the midterms and new state governors in data-center hubs like Texas and Ohio pledging to pause approvals—has conflated two very different things: companies whose revenue is tightly bound to new data-center construction, and companies holding scarce assets or whose demand doesn't depend on that path. The former deserve repricing; the latter are being wrongly sold off. The author applies three criteria—scarcity, order visibility, and reasonable or undervalued pre-crash valuation—to identify the truly mispriced names.
TSMC is the clearest case, meeting all three: it is the only scalable 3nm/2nm pure-play foundry with ~90% AI chip foundry share, its advanced-node capacity is booked past 2028, and only about 30% of revenue ties to AI data centers, with HPC at 68% but smartphones, auto, and IoT providing a buffer. Its ADR trades near 24x PE after the drop, low for semiconductors. Nvidia fell ~10% overnight and 25% in a month; its 17x forward PE is a three-year low and order visibility is strong (management cites chip supply, not demand, as the constraint), but with data-center revenue over 90% of the total, it is a high-beta bet rather than a defensive mispricing. Apple is defensive—capex is only ~2.5% of sales and on-device AI runs on already-sold chips—but its 36x forward PE sits at the 99th percentile of its own history, so it fails the valuation test.
Google and Amazon have huge backlogs ($514B and $496B respectively) but their conversion depends on compute supply, so they aren't fully decoupled. Meta, at 18–19x forward PE near a three-year low, is partly mispriced: its three flagship US projects (Ohio, Louisiana, Texas) are already past permitting, though new pipelines will slow. Broadcom, AMD, and Intel are falling for real exposure reasons, not mispricing. In China, SMIC best fits the criteria, with 89.6% China revenue tied to domestic fabless demand rather than US cloud orders; North China Innovation has strong visibility but a fuller valuation; ChangXin Memory shows an emotional discount versus fundamentals. Among model makers, US pure-plays are hit hardest, while Chinese firms' structural cost advantages are amplified—Zhipu, running GLM-5.3 on AWS Bedrock with revenue sharing, benefits from existing cloud capacity and may be somewhat mispriced. The conclusion: only TSMC truly satisfies all three criteria; the broad "anything AI rises" phase is over, and differentiation will favor companies with real orders, revenue, and compute reserves.
marsbit21m ago