TSMC’s Q3 2026 Beat: 2nm Revenue Hits 8%, HPC Up 22%
Taiwan Semiconductor Manufacturing Company (TSMC) reported third-quarter 2026 results that exceeded market expectations, with 2-nanometer process revenue accounting for 8% of total wafer revenue — a faster ramp than many analysts anticipated. The company’s high-performance computing (HPC) segment grew 22% year-over-year, underscoring sustained demand for advanced silicon from AI and data center customers.
The results carry direct implications for Nvidia (NVDA) and Advanced Micro Devices (AMD), both of which rely on TSMC’s leading-edge nodes and advanced packaging for their AI accelerators.
Nvidia: Demand Confirmed, Packaging Still the Wild Card
Nvidia’s data center GPUs, including the Blackwell and Rubin families, are fabricated on TSMC’s 4nm and 3nm processes and packaged using CoWoS (Chip-on-Wafer-on-Substrate) technology. The 22% HPC growth is a strong proxy for Nvidia’s AI demand pipeline, suggesting hyperscaler capex remains robust despite broader macro uncertainty.
However, CoWoS capacity remains the key bottleneck. TSMC has been expanding CoWoS lines, but the ramp has not kept pace with demand for high-bandwidth memory (HBM)-integrated packages. Any shortfall could cap Nvidia’s shipment growth even as wafer supply improves.
AMD: MI450 on Track as 2nm Ramps
For AMD, the 2nm ramp is a more direct catalyst. The company’s MI450 accelerator, expected to launch in 2027, is slated to use TSMC’s 2nm process. An earlier-than-expected ramp improves AMD’s odds of hitting volume production targets and closing the performance-per-watt gap with Nvidia.
AMD’s challenge is not just silicon — it is software. The ROCm ecosystem still lags Nvidia’s CUDA in developer adoption, and enterprise customers remain cautious about multi-vendor AI stacks. But a reliable 2nm supply chain removes a critical execution risk.
Market Skepticism on AI ROI Persists
Despite the strong TSMC print, both NVDA and AMD shares have been under pressure as investors question the return on massive AI infrastructure spending. The concern: if AI monetization slows, hyperscalers could cut capex, hitting GPU orders with a lag.
TSMC’s results argue against an imminent slowdown — but they do not resolve the ROI debate. For now, the foundry data suggests the AI buildout is intact. The next test will be whether CoWoS capacity and 2nm yields can scale without margin erosion.
