Amazon CEO Andy Jassy Explained Why Semiconductor Sales Could Keep Climbing for Years to Come. Here's My Favorite AI Chipmaker Right Now.

The article identifies TSMC as the preferred long-term AI semiconductor investment, citing its 73% share of third-party chip manufacturing, $143 billion in trailing-12-month revenue, and expected annual revenue growth of roughly 25%. Hyperscalers are expected to shift a larger share of data-center capex toward shorter-lived chips and equipment as facilities are reused across multiple server generations; Microsoft says about two-thirds of current capex already goes to GPUs, CPUs, and related equipment. While Nvidia and Broadcom remain major AI beneficiaries, TSMC is positioned to benefit regardless of whether customers adopt GPUs, custom ASICs, or other chip architectures, with shares trading below 25x earnings.

Analysis

The investable implication is not simply higher AI chip demand, but a mix shift from one-time site construction toward recurring compute replacement. That favors TSM because it captures wafer content across GPUs, custom ASICs, CPUs, networking and potentially edge inference, while NVDA and AVGO remain exposed to architecture choices and customer concentration. TSM's leverage is especially attractive if hyperscalers optimize total cost of ownership by diversifying accelerators rather than reducing absolute silicon intensity.

The key near-term risk is that custom-silicon adoption is interpreted as an NVDA share-loss event and drives a broad AI-semiconductor de-rating. For TSM, this would likely be a relative-entry opportunity unless leading-node utilization, CoWoS advanced-packaging capacity, or customer tape-out volumes weaken; those are the operational variables that matter more than aggregate hyperscaler capex headlines. A stronger dollar, Taiwan geopolitical risk premium, and any US export-control expansion remain the principal reasons TSM may not receive a full NVDA-like multiple.

Over 1-3 months, watch earnings commentary from AMZN, GOOG and MSFT for the ratio of short-lived equipment spending to total capex, plus TSM monthly sales and advanced-packaging lead times. Over 6-18 months, the more consequential catalyst is whether 2nm pricing and packaging scarcity sustain gross-margin expansion despite overseas-fab start-up dilution. Consensus may be underestimating TSM's ability to monetize silicon diversification: an ASIC share gain for AVGO or internal chips at AMZN/GOOG can still increase TSM content rather than displace it.

The contrarian caveat is valuation: “picks-and-shovels” ownership does not eliminate cyclicality. If hyperscalers simultaneously slow build-outs after capacity comes online, foundry utilization can fall faster than end-demand because customers work down inventory; a sequential decline in TSM's HPC revenue or a material cut to 2027 capex would falsify the structural-through-cycle thesis.

Key Decisions for Investors

• Initiate a 6-12 month long TSM / short SMH relative-value position, sized to hedge broad semiconductor beta. Thesis: TSM benefits from both merchant GPU and custom-ASIC substitution; reassess if TSM reports declining leading-edge utilization or cuts capital expenditure materially.
• Add TSM on any 10-15% AI-semicapex-led pullback rather than chase positive commentary. Target a 15-25% relative return versus SMH over 12 months if advanced-node pricing and packaging mix support earnings revisions; hard risk is a Taiwan-risk shock, which can gap the ADR independent of fundamentals.
• Maintain AVGO as the higher-beta custom-silicon expression, but pair it against NVDA only after confirmation of incremental hyperscaler ASIC programs in quarterly commentary. The trade is attractive over 3-6 months if AVGO AI revenue guidance continues to accelerate; exit on evidence that customer programs are delayed or advanced-packaging supply constrains shipments.
• Set an event alert around AMZN, GOOG and MSFT results: a reduction in equipment-capex growth without offsetting disclosure of higher utilization or cloud revenue acceleration is a warning for 2027 wafer demand and argues for trimming AI semiconductor exposure.