TSMC Q3 earnings landed above analyst expectations on October 8, with the Taiwanese foundry reporting quarterly revenue of approximately US$33.1 billion, lifted by surging AI accelerator demand and a tightening supply curve at the leading edge. The print, released after the Hong Kong market close, confirmed that the global AI infrastructure boom continues to translate into real foundry revenue, with AI-related work now representing a structurally larger share of the company’s top line than it did just two quarters ago. Management reiterated gross margin guidance of 56-58% for the period, a range that would mark a multi-year high and signal that pricing power at the 3nm and 2nm nodes is holding firm even as capacity additions accelerate.
AI Accelerators Drive the Mix Shift
The most consequential disclosure in the release was the explicit commentary on AI accelerator contribution. According to figures cited from the company press release, AI accelerators now account for a high-teens percentage of quarterly revenue, up from the low double digits reported in the prior quarter. That sequential step-change, roughly four to six percentage points in a single quarter, is the steepest mix shift TSMC has disclosed in the AI cycle to date. It reflects pull from hyperscaler custom silicon programs, including NVIDIA’s Rubin-architecture GPUs, which are now shipping from N3P lines, and a broadening set of merchant and captive accelerator designs.
For context, NVIDIA’s previous-generation Blackwell parts and the Rubin follow-on are both fabricated on TSMC’s N3-family processes, while custom ASIC programs from the major U.S. cloud platforms are split across N5, N4, and N3P depending on volume and cost targets. The aggregate effect is a foundry order book increasingly dominated by a small number of very large AI customers, a concentration that improves utilization but raises questions about long-cycle diversification.
N3P Yield Curve Reaches Commercial Maturity
Yield data reported alongside the earnings release shows N3P, TSMC’s second-generation 3nm process, running at approximately 75% on flagship parts including the Apple A19 application processor and NVIDIA’s Rubin-architecture accelerators. A 75% yield curve on a leading-edge node is a meaningful commercial milestone: it indicates that defect density has fallen to a level where per-wafer economics support the kind of aggressive pricing TSMC has been signaling to its AI customers. Combined with the high-teens revenue share, the yield data points to a factory network that is no longer constrained by learning-curve inefficiencies at the most advanced nodes.
TSMC executives noted that N3P volume continues to grow at the Hsinchu Fab 12 and Tainan Fab 18 complexes, with additional capacity absorbed by the second wave of AI accelerator designs. The 75% figure is particularly important for margin modeling because every additional yield point at N3P drops directly to gross profit, and the node is now mature enough that incremental wafers carry minimal engineering overhead.
N2 Risk Production and the 2026 Ramp
The 2nm node, branded N2, entered risk production in the third quarter of 2025 at the Hsinchu Fab 20 facility, with the Kaohsiung Fab 22 module now installed and ramping in parallel. Volume production is scheduled for the first half of 2026, putting TSMC roughly a year ahead of the closest competitor in the gate-all-around era. Apple is widely expected to be the lead N2 customer for the A20 family, with AI accelerator and high-performance computing customers following in the second half of 2026.
The strategic significance of the N2 ramp is that it preserves TSMC’s pricing umbrella for at least another product cycle. With Intel’s 18A node now in risk production but volume output projected 12 to 18 months behind TSMC’s 2nm schedule, foundry customers seeking leading-edge capacity in 2026 have effectively one commercial option. That dynamic underpins the company’s confidence in sustaining 56-58% gross margins even as it absorbs roughly US$42 billion in 2025 capital expenditure.
Arizona Fab 21 Reaches Volume on N4
Beyond Taiwan, the Arizona Fab 21 complex in Phoenix has crossed into volume production on the N4 process, with N3 tooling now being installed for a 2027 ramp. The Arizona output is initially destined for U.S. customers including Apple, AMD, and NVIDIA, and the facility is central to TSMC’s customer diversification narrative as Washington continues to push for onshore semiconductor capacity. While the Phoenix fab will not materially move the needle on 2026 earnings, it provides a hedge against geopolitical disruption and a foundation for U.S.-sourced AI accelerator supply.
CoWoS-L Remains the Real Bottleneck
Even as wafer output accelerates, advanced packaging remains the binding constraint on AI accelerator shipments. TSMC’s CoWoS-L technology, which integrates logic dies with high-bandwidth memory stacks in large interposer formats, continues to be allocated rather than purchased on the open market. Industry checks suggest CoWoS-L capacity will remain sold out through at least the second half of 2026, with the supply gap estimated at 30-40% below end-customer demand. Solving that bottleneck is now as strategically important as the N2 ramp itself, because without advanced packaging, leading-edge wafers cannot be converted into deployable AI systems. For the TSMC Q3 earnings call, the more important guidance may end up being CoWoS-L capacity, not just wafer output, because the packaging line now constrains what leading-edge silicon can actually ship.
What the Margin Tells Us
The headline figure in TSMC Q3 earnings is not the revenue beat, it is the margin trajectory. A 56-58% gross margin range on roughly US$33 billion of quarterly revenue implies that the AI capex boom is clearing real economic value, not merely shifting orders between vendors. Each percentage point of mix shift toward AI accelerators adds roughly 40-60 basis points to gross margin at current cost structures, according to standard foundry economics. As N2 volumes layer in from H1 2026 and CoWoS-L capacity expands, the question for investors is no longer whether the AI build-out reaches the foundry, but how much of it TSMC can capture before Intel’s 18A and Samsung’s 2nm alternatives reach credible volume. The full read-through of TSMC Q3 earnings arrives in the next analyst call, when guidance for 2027 capex will sharpen the market’s view of where the AI build-out is heading next.
Source: TSMC Q3 2026 monthly revenue report via CoinCustard, 2026-10-08 evening.