Supply
TSMC Quarterly HPC Revenue
TSMC High Performance Computing platform revenue in USD billions, converted from disclosed NT$ totals at each quarter’s TSMC average FX rate. A vendor-agnostic proxy for leading-edge compute silicon in production—not TSMC AI revenue. Calendar quarters: 1Q is January–March.
TSMC Quarterly HPC Revenue
- TSMC HPC
API access · Pro coming soon
As of 2026-04-01, TSMC HPC platform revenue was $26.3bn in 2Q26, up 45% from 2Q25.
Methodology
Values start from TSMC-disclosed HPC platform revenue in New Taiwan dollars from quarterly and annual financial statements, then convert to USD billions at that quarter’s average USD/NTD rate from TSMC’s management reports. They are platform totals, not the rounded percentage-of-revenue figures on earnings slides. TSMC’s 2Q26 six-month statements report 2Q26 HPC of NT$830.369bn exactly (US$26.278bn at 31.60), versus NT$561.155bn / US$18.073bn in 2Q25; 1H26 HPC was about US$48.3bn.
Where TSMC publishes a full-year platform total but not a standalone fourth-quarter figure, 4Q is derived as full-year minus the first nine months. Those derived fourth quarters are flagged in the underlying dataset notes.
TSMC defines HPC broadly: servers, PCs, tablets, game consoles, base stations, and other high-performance computing products. AI accelerators, CPUs, GPUs, NPUs, and related ASICs are important drivers inside that mix. HPC is therefore not synonymous with AI. A better description is a proxy for the dollar value of high-performance compute silicon being manufactured, with AI and data-center silicon increasingly the dominant growth driver.
The series sits in the semiconductor production layer. For an AI accelerator or custom ASIC to generate TSMC revenue, the chip has been designed, taped out, and is actually being manufactured. That is earlier in the chain than NVIDIA Data Center shipments, Broadcom AI semiconductor revenue, or data-center installation, and later than hyperscaler capex guidance (which measures intention).
HPC is preferred to total TSMC revenue because smartphones, automotive, IoT, and consumer electronics can obscure the compute cycle. HPC was roughly 41% of TSMC revenue in 2022, 43% in 2023, 51% in 2024, 58% in 2025, and a rounded 66% in 2Q26.
TSMC reports calendar quarters (1Q January–March). NVIDIA’s fiscal year ends in late January and Broadcom’s in late October, so matching 1Q to Q1 across vendors misaligns the periods. Compare by calendar date, not by quarter number.
Frequently asked questions
Is TSMC HPC the same as AI semiconductor revenue?
No. TSMC HPC revenue is not AI semiconductor revenue. The HPC platform also includes ordinary PCs, tablets, game consoles, conventional servers, and base-station silicon. A useful decomposition is AI compute + conventional data-center compute + PC/console and other HPC. The reason the series has become a good AI proxy is that incremental growth since 2023 appears overwhelmingly linked to AI-related compute—not that every dollar of HPC is an AI chip.
Why is this useful as an AI indicator if it is not AI revenue?
Many AI indicators measure intention: hyperscaler capex guidance, GPU orders, data-center construction plans, power commitments. TSMC revenue is closer to actual physical production. For a chip to show up here, it has progressed through design and tape-out and is being manufactured. TSMC has said AI chips require leading-edge processes, large die sizes, and advanced packaging—the areas where its manufacturing position is strongest. Conceptually: AI demand → designs/orders → TSMC wafer and packaging production → accelerator and system shipments → data-center installation → compute online. HPC sits fairly early in that chain, as an upstream signal of the physical AI infrastructure buildout.
How is this different from NVIDIA Data Center or Broadcom AI revenue?
NVIDIA Data Center revenue is the best public proxy for merchant accelerated computing, mostly GPUs plus NVIDIA networking. Broadcom AI semiconductor revenue captures custom XPUs and a large share of AI cluster networking. TSMC HPC is vendor-agnostic foundry production: AI accelerator silicon + custom AI ASICs + data-center CPUs + networking/compute silicon + non-AI HPC. TSMC said in its 2Q26 call that agentic AI is increasing the role of CPUs in AI data centers in addition to accelerators, and that AI-related demand from cloud service providers remains extremely robust. If workloads migrate toward custom ASICs or a more CPU-heavy architecture, a GPU-vendor series can understate total AI silicon growth while TSMC still captures a large portion of the leading-edge wafers.
Why use HPC instead of total TSMC revenue?
Total TSMC revenue includes smartphones, automotive, IoT, and consumer electronics. Those cycles can hide what is happening in AI compute. HPC strips a large portion of that noise away. The mix shift is dramatic: HPC went from about 41% of TSMC revenue in 2022 to a rounded 66% in 2Q26. TSMC has increasingly become a high-end compute company from a revenue-mix perspective. TSMC has also said AI accelerators are expected to be the largest contributor to incremental revenue growth over its long-term forecast, with AI-accelerator revenue previously guided toward a mid-to-high-50% CAGR from 2024 through 2029, and 2Q26 commentary that the signals were becoming even stronger.
Why does year-over-year growth matter more than the $26bn level?
The absolute HPC figure mixes AI silicon with conventional HPC. The trajectory and YoY rates are the AI-cycle signal. HPC was essentially stagnant through 2023 (2Q22 $7.8bn → 2Q23 $7.0bn), then accelerated: 2Q24 $10.9bn, 2Q25 $18.1bn, 2Q26 $26.3bn—roughly 3.4× in four years, or about a 36% CAGR from 2Q22 to 2Q26. The major inflection starts in 2024, when hyperscaler AI infrastructure and accelerator production ramped at scale. A sequence such as +70% → +60% → +40% → +20% would be an early warning of AI-chip production deceleration. The recent path from +30.9% in 4Q25 to +46.6% in 1Q26 and +45.4% in 2Q26 is a sign that the AI silicon production cycle remains strong rather than rolling over.
What does TSMC HPC miss as an AI hardware indicator?
It is not a proxy for the entire AI capex ecosystem. It does not fully capture HBM memory, where SK Hynix, Samsung, and Micron are central; data-center racks and server assembly; networking equipment outside TSMC-manufactured silicon; power generation and grid equipment; cooling; buildings and land; or actual software and inference utilization. A stronger hardware dashboard combines TSMC HPC with HBM revenue or bit shipments, NVIDIA Data Center revenue, Broadcom AI semiconductor revenue, hyperscaler capex, CoWoS capacity, and data-center power demand. Among those, TSMC HPC is particularly valuable because it is vendor-agnostic: whether the compute winner is a merchant GPU, a hyperscaler ASIC, or a more CPU-heavy architecture, a large portion of that leading-edge silicon can still show up at TSMC.
Why are quarters labeled 1Q22 instead of Q1 FY22?
TSMC reports calendar quarters and uses 1Q22 in its own filings. 1Q is January–March, 2Q April–June, 3Q July–September, and 4Q October–December. NVIDIA’s fiscal year ends in late January (Q1 is February–April) and Broadcom’s ends in late October (Q1 is November–January). Matching 1Q to Q1 across those charts misaligns the periods. Compare TSMC with NVIDIA and Broadcom by calendar date, not by quarter number.