TSMC's 28nm Retreat: A Calculated Bet on AI-Era Silicon
TSMC has reportedly cut 28nm output by more than 25% since early 2026, redirecting capacity toward advanced nodes as AI-driven revenue hits record levels — a pivot that concentrates both the company's upside and its geopolitical exposure.
TSMC's decision — reportedly confirmed by TrendForce — to cut mature 28nm process output by more than 25% since the start of 2026, reallocating those wafer starts toward advanced and specialty nodes, is more than a capacity reshuffle. It marks a structural inflection point at the world's most consequential foundry: the deliberate retirement of commodity silicon to concentrate resources on the AI-infrastructure workloads that now dominate demand.
The financial backdrop makes the logic clear. Revenue grew 37% year-on-year in January 2026, 31% in September 2025, and 30% again in May 2026, even as persistent concern about potential Taiwan chip export restrictions kept the stock under pressure. Advanced packaging bookings are reported to be fully subscribed through 2028, with clients including Nvidia — which overtook Apple as TSMC's largest single customer in January 2026 — competing for wafer allocation to feed AI accelerator supply chains. Nvidia's Vera Rubin GPU and Vera CPU, which taped out at TSMC in August 2025, exemplify the demand profile driving the reallocation: at the frontier, TSMC commands premium pricing and faces no credible near-term substitute. Against that backdrop, 28nm economics are becoming structurally unattractive.
TSMC's packaging roadmap reinforces the pivot. The company is accelerating its next-generation CoPoS architecture to replace CoWoS, with glass core substrates reportedly cutting costs by 30% and pushing wafer utilization above 90%. Advanced packaging has become a second competitive axis alongside process nodes: Nvidia's and AMD's AI chips are as dependent on TSMC's integration capabilities as on its lithography. MediaTek's May 2026 announcement of support for both TSMC and Intel advanced packaging technologies signals that the broader ecosystem is hedging, even as TSMC's utilization data suggests its gravitational pull remains intact. The Amkor partnership for US-based advanced packaging, disclosed in June 2026, extends that integration capability into the American supply chain — partly in anticipation of regulatory mandates, not merely client preference.
Yet the very tightness of TSMC's capacity generates its own risks. Industry reports note that the crunch is already prompting clients to evaluate Samsung as an alternative, and that competitive threat has acquired more substance over the past year: Google's December 2025 deal for Samsung to manufacture TPU chips demonstrated that hyperscalers will accept performance trade-offs to reduce single-vendor exposure, while September 2025 reports of Samsung gaining Nvidia AI chip qualification pointed in the same direction. TSMC CEO C.C. Wei's February 2026 reaffirmation of production commitments to Nvidia was read as reassurance, but it also underscored how concentrated the revenue base has become. Nvidia's ascent to TSMC's top-customer position means that any disruption — from trade policy, export controls, or alternative sourcing — carries amplified financial consequences.
The geopolitical overlay remains the most difficult variable to price. Despite 30% revenue growth in May 2026, TSMC's stock stayed under pressure, reflecting market concern that Taiwan chip export restrictions could tighten further. The Trump administration's return in late 2024 introduced policy uncertainty that has not fully dissipated, and TSMC's US fab buildout and the Amkor collaboration are partly structured responses to that environment. The 28nm capacity cut, while commercially rational, also shifts TSMC's production mix toward the advanced nodes most scrutinized by regulators — and most coveted by rivals seeking technology exposure.
Three signals warrant close attention in the coming quarters. First, whether client migration to Samsung moves from exploratory qualification into volume production — that would indicate genuine erosion of TSMC's moat rather than mere bargaining leverage. Second, the trajectory of Taiwan and US export control legislation: material tightening on advanced node output or leading-edge packaging would compress TSMC's addressable market faster than new capacity can replace it. Third, the commercial validation of glass-core CoPoS packaging at scale: if the reported 30% cost reduction and above-90% utilization figures hold in volume production, TSMC may recapture margin at the advanced packaging tier even as 28nm revenue recedes. The retreat from 28nm is a deliberate bet on where semiconductor value is accreting. Whether the transition is fast enough, and the alternatives scarce enough, to sustain TSMC's structural position is the question that will define the next phase of the AI infrastructure buildout.