• AI demand remains strong, with smartphone and high-performance computing businesses growing together. The wider adoption of AI agents is also creating additional demand for CPUs and networking chips, broadening TSMC's growth base.
• Capacity is expanding for both 3nm and 2nm processes, while custom memory offers another growth opportunity.
• New processes and overseas fabs bring cost pressure, but past performance supports our confidence that gross margin can stay high. Supplies of materials, equipment and substrates will help determine whether additional wafer output can translate into AI chips ready for delivery.

• Revenue growth: Third-quarter revenue reached USD46.72 billion, up 41.2% year on year.
• Rising AI demand: HPC accounted for 68% of revenue and grew 22% quarter on quarter.
• Advanced processes: The 2nm share of wafer revenue rose to 8%, while 3nm accounted for 33%.
• Profitability: Gross margin reached 66.4%, with fourth-quarter guidance of 65%-67%.
• Target price: Based on our 2028 earnings forecast and a 26x P/E, our TSMC target price is USD713, implying potential upside of around 70%.

TSMC Earnings Analysis

TSMC's earnings remained strong this quarter. As we expected, sustained demand for leading-edge processes drove revenue growth, while high capacity utilisation and cost control supported profitability. The company raised its full-year outlook again, forecasting US-dollar revenue growth of around 44%. It expects the next quarter to continue benefiting from AI-driven demand for leading-edge processes and a rapid increase in N2 output.

Table 1: TSMC's Third-Quarter 2026 Results, Company Guidance and Fourth-Quarter Outlook

Metric	3Q26 Actual	3Q26 TSMC Guidance	YoY Change	4Q26 Guidance
Revenue (USD billion)	46.72	44.6-45.8	+41.2%	Midpoint: +14% QoQ Midpoint: +58% YoY
Gross margin	66.4%	65%–67%	+6.9 pp	65%–67%
Operating margin	58.7%	56%–58%	+8.1 pp	56.5%–58.5%
Net margin	51.1%	—	+5.4 pp	—
Net profit (USD billion)	23.86	—	+57.8%	—
EPS per ADR (USD)	4.60	—	+57.8%	—
Wafers shipped (million)	4.476	—	+9.6%	—
Source: TSMC and iFAST Compilations. Data as of 19 October 2026.

TSMC's AI Demand Drivers: Growth in Smartphones and HPC, with Added Demand from AI Agents

Despite pressure on the smartphone market from higher memory prices, TSMC's smartphone and high-performance computing (HPC) businesses grew together, giving the company several sources of growth.

Table 2: TSMC's Third-Quarter Revenue Mix by Platform and Quarterly Growth

Platform	3Q26 Revenue Share	Revenue Change QoQ
High-performance computing (HPC)	68%	+22%
Smartphones	21%	+10%
Internet of Things	4%	+5%
Automotive	4%	+23%
Digital consumer electronics	1%	−12%
Others	2%	−10%
Source: TSMC and iFAST Compilations. Data as of 19 October 2026.

AI agents are also becoming a new source of chip demand. C.C. Wei said AI applications are moving from chatbots towards AI assistants, creating additional demand for CPUs and networking chips alongside accelerators.

The recently launched Grok Bot, Meta Muse and OpenAI dots show the latest direction of AI agents: remembering users' needs, retaining work progress and continuing to carry out tasks. Meta Muse has a dedicated cloud computer that can run a browser and use tools, with support for scheduled and parallel tasks. Each OpenAI dot also has its own cloud computer.

Take Muse helping to book a flight. The agent needs to consider the user's budget, preferred departure times and preference for direct flights, then search for flights and compare fares and connections. If prices or requirements change, it needs to search again and adjust the options before booking with the user's confirmation.

These steps involve repeatedly running browsers, calling tools and processing data, adding CPU work. Understanding preferences and replanning may also increase model calls, adding inference work for GPUs and other accelerators. Parallel tasks raise memory requirements; retaining preferences and work progress requires storage; and checking prices across websites increases network traffic.

In our view, wider use of these agents, more frequent activity and longer-running tasks will increase demand for several types of hardware. This could extend TSMC's foundry opportunities beyond accelerators to CPUs and networking chips.

TSMC's Advanced Processes Gain Share: N3 Leads as N2 Output Accelerates

TSMC's wafer revenue mix continues to shift towards advanced processes, with 7nm and below rising to 79% this quarter. N3 overtook N5 as the largest contributor to wafer revenue. Sustained strong customer demand also supports further capacity expansion, extending N3's growth cycle.

Table 3: Advanced Processes as a Share of TSMC's Wafer Revenue

Process	3Q26	2Q26	3Q25
2nm (N2)	8%	3%	0%
3nm (N3)	33%	30%	23%
5nm (N5)	28%	33%	37%
7nm (N7)	10%	11%	14%
Advanced processes (7nm and below)	79%	77%	74%
Source: TSMC and iFAST Compilations. Data as of 19 October 2026.

In our previous article, we noted that TSMC had begun expanding N3 capacity to support customers' multi-year demand. In the past, the company would typically stop expanding a process node once it reached its target capacity. Management confirmed this quarter that N3 had reached its original target, but strong customer demand was still prompting further expansion, changing its previous approach.

As N3 continues to expand, the most advanced 2nm (N2) process is also contributing more revenue. N2's share of wafer revenue rose 5 percentage points quarter on quarter, the largest increase among all processes, with demand coming from both smartphones and HPC/AI. AMD's next-generation EPYC server processor will use TSMC's 2nm process, MediaTek has completed the tape-out of an N2P chip, and Apple's A20 Pro and M6 also use 2nm technology. We expect smartphone and HPC/AI demand for advanced processes to keep growing. Together with improving yields, this should gradually make N2 a key revenue driver in TSMC's next phase of growth.

To meet demand, TSMC is expanding N2 and N3 capacity together. Management expects annual capacity for the N2 family to more than double in 2027. N3 supply will be increased through new fabs and the conversion of existing equipment.

Further Reading: TSMC | The Kingmaker of the AI Era: Investment Opportunities Under the New Compute Hierarchy

Table 4: TSMC's N2 and N3 Capacity Expansion Plans

Process	Location / Project	Plan	Timing / Target
N2	Hsinchu and Kaohsiung	Phased expansion and higher production efficiency	Annual capacity to more than double in 2027
N3	New Tainan fab	New N3 fab	Mass production in 1H 2027
N3	Additional Tainan fab	Additional N3 fab	Mass production in 1H 2028
N3	Second Arizona fab	N3; equipment move-in and installation in progress	Mass production in 3Q27, slightly earlier than planned
N3	Second Kumamoto fab	N3 process	Mass production in 2028
N3	Existing Tainan fabs	Convert 5nm tools to support N3; excluded from new fab count	Not separately disclosed
Source: TSMC and iFAST Compilations. Data as of 19 October 2026.

Although TSMC's third-quarter capital expenditure fell 11.8% quarter on quarter to USD13.84 billion, its full-year budget remained at USD60-64 billion. This implies fourth-quarter spending of around USD19.4-23.4 billion, roughly 40% to almost 70% more than in the third quarter. Management also reiterated that capital expenditure in 2026-2028 would be significantly higher than in the preceding three years, to support long-term demand.

TSMC Gross Margin Outlook: N2 Costs Emerge, but Fourth-Quarter Margin Is Expected to Stay High

Gross margin was 66.4% this quarter. Although it declined quarter on quarter, it remained above the midpoint of company guidance and at a historically high level, demonstrating TSMC's strong profitability.

Chief Financial Officer Wendell Huang said the decline mainly reflected the initial costs of bringing N2 into production. Unit costs should gradually fall as yields, utilisation and production scale improve. Overseas fabs, particularly those in the US, create more persistent cost pressure. The company expects overseas fabs to dilute gross margin by close to 2 percentage points in 2026. As more US fabs enter production, this impact is expected to widen to 3-4 percentage points over the coming years.

The maturing N3 process could help offset some of this pressure. In the first quarter, management had expected N3 gross margin to exceed the corporate average in the second half of 2026, supporting overall gross margin.

Looking ahead, fourth-quarter gross margin guidance is 65%-67%, which we consider slightly conservative. Figure 1 shows that actual margins in recent quarters have consistently exceeded the midpoint of guidance, with some also above the upper end. We remain confident that fourth-quarter gross margin can stay high despite the cost pressure from new capacity.

Figure 1: TSMC Quarterly Gross Margin - Actual Results, Guidance and Market Consensus

What Is NVHBM, and How Can TSMC Benefit from Custom HBM?

In Memory Supercycle: A Structural Shift Rewriting Supply and Demand Paradigm, we explained the structure of HBM and how the base die can be customised to customers' computing needs, moving some functions from a GPU or another AI chip into it. The DRAM dies above store data, while the base die handles signals and control.

NVHBM is NVIDIA's custom HBM solution. It moves the memory controller from the compute chip to the base die, freeing up area for more compute circuitry while increasing memory bandwidth and reducing HBM power consumption.

NVIDIA's common NVHBM design lets multiple memory makers supply products to the same specifications, reducing integration and validation work across suppliers. Memory partners will also offer the solution to external customers using the NVLink Fusion platform. Amazon's Annapurna Labs is the first partner and plans to use the technology in its own Trainium chips. The technology used in NVHBM will also be used in NVIDIA's future GPUs.

Further Reading: How AWS Uses Custom Chips to Build a Money-Printing Machine

According to reports, NVIDIA has asked Samsung, SK hynix and Micron to produce NVHBM base dies using TSMC's 3nm process (the N3 family), and Samsung has accepted the requirement. Samsung has its own foundry, and TSMC's N3 capacity also needs to serve other customers, yet NVIDIA has still specified TSMC for manufacturing.

In our view, this reflects the value NVIDIA places on TSMC's manufacturing capabilities and platform collaboration. TSMC already manufactures HBM4 base dies. By adding functionality and upgrading the process, NVHBM could increase TSMC's manufacturing value per HBM stack, broaden its customer base and bring additional demand for advanced logic processes.

AI Chip Supply Bottlenecks: ABF Substrates and Material Constraints

TSMC is expanding advanced-process capacity to meet strong demand. At the same time, packaging capacity and supplies of substrates and materials face constraints. Whether they can scale together is another test for AI chip delivery.

Alongside compute chips and HBM, AI packaging also relies on organic substrates and several materials:

• The organic substrate sits beneath the chips and HBM, supplying power, carrying high-speed signals, supporting the chips and connecting the package to the motherboard.
• ABF is an insulating film developed by Ajinomoto and used between wiring layers in the substrate;
• T-glass fibre cloth helps control thermal expansion and reduce warpage during high-temperature processing.

GPU makers are integrating more compute chips and HBM into a single package, increasing substrate area and material use. Higher requirements for power delivery and high-speed signals also call for more wiring layers and finer traces. This adds processing steps, extends production times and complicates yield control, making it harder for high-end substrate supply to keep up with demand.

Customers are securing supply through long-term reservations. Reports indicate that some AI chip customers and major cloud service providers have booked ABF substrate capacity through 2028, and have asked suppliers to plan expansion early to increase supply in 2029-2030.

Unimicron, a Taiwanese manufacturer of printed circuit boards and semiconductor package substrates, has raised its 2026 capital expenditure budget from NT$34 billion to over NT$53.7 billion. Around 80%-85% of the incremental investment is going to ABF. Its budget for long-lead-time equipment has also increased from NT$14.1 billion to NT$31.6 billion.

Materials still constrain expansion. Unimicron Chairman Shan-Chieh Chien said supplies of T-glass fibre cloth, copper-clad laminates (CCL) and ABF remain tight.

Equipment lead times also affect the speed of expansion. Chien said some key machines have lead times of up to 30 months. Substrate makers must also complete equipment delivery, installation and qualification. Published plans from some substrate manufacturers show that capacity will come on stream in phases, with some additions scheduled for 2027-2028.

In our view, even after substrate makers increase investment, the pace of supply improvement still depends on materials, equipment and qualification progressing together. Wafer capacity expansion needs matching packaging and substrate supply before it can translate into deliveries of complete AI chips.

TSMC's 2028 Target Price: Advanced Processes and Customer Partnerships Support Long-Term Growth

TSMC will continue to benefit from broader chip demand. AI agents are extending demand beyond accelerators to CPUs and networking chips; custom HBM, particularly NVHBM, could also increase the foundry value of its base-die business.

In our view, TSMC's advanced processes, advanced packaging and close customer relationships allow it to keep contributing to customers' new product development. We continue to find the company highly attractive as a long-term investment. Following the recent share-price pullback, its forward P/E has fallen back to near its five-year average. Based on our 2028 earnings forecast and a 26x P/E, our target price for TSMC (NYSE: TSM) is USD713, implying potential upside of around 70%.

As well as individual stocks, investors could consider the Global X Asia Semiconductor ETF (HKEX: 3119) for broader exposure to Asia's semiconductor sector.

Figure 2: TSMC ADR Price and Forward P/E

Table 5: TSMC Earnings Forecasts and Potential Upside

TSMC ADR (USD)	2025A	2026E	2027E	2028E
P/E (x)	39.4	25.2	19.0	15.3
Earnings growth YoY (%)		56.1%	32.6%	24.4%
EPS per ADR (USD)	10.65	16.62	22.04	27.42
Target price (26x fair P/E)				713
Potential upside (%)				70%
Source: Bloomberg and iFAST Compilations. Data as of 22 October 2026.