Results: ASML and TSMC

When the leaders play game of chicken

To start off, the guidance from ASML and TSMC are not always the best leading indicator for the industry despite their bellwether status. This is how the supply chain works:

End Customers → Fabless / IDM → Foundry → WFE

Foundry sees demand picks up only when the fabless customers are optimistic. Similar situation for WFE, which relies on foundry and IDM customers for the outlook.

ASML - Orderbook miss is just a noise

With the street expectation for Q1 order at €5-6bn, market was clearly disappointed when the number came in at €3.6bn with EUV accounting for only €656 mn (~3 EUV machines, probably 2 low-NA and 1 high-NA).

EUV orders are lumpy in nature given the high value of each equipment (€200mn for low-NA) and how customers can choose to delay order from 31st March to 1st Apr.

Following strong EUV orders from memory in Q4 2023, TSMC was the main culprit for the order miss. ASML claimed that TSMC was delaying the orders as they were trying to negotiate on the pricing of the machine.

How is this possible? It goes back to ASML’s plan to increase capacity to 90 EUV machines by 2025-26. ASML current Low-NA EUV backlog is around 60-70 and this year shipment will be in the 40s. TSMC will require 20+ EUV machines in each of 2025 and 2026 to support the ramp of its 2 N2 fabs.

Lead time for EUV is around a year and there is surplus capacity which allows ASML to prebuild inventory into 2025. Even if TSMC made the orders for the 40+ machines in Q3 2024, TSMC will receive the EUV machines in time for the ramp of N2 fab in Q4 2025. ASML knows that they are the only provider of EUV machines and TSMC is ramping N2 fab in 2025.

The other alternative theory is that TSMC will reuse some of its N3 and N5 tools for N2, just as they did for N3 using some of N5 tools. However, this is very unlikely given the difference in location and the high demand for N3.

“We can convert one technology node capacity to the next one is because of our there's physical advantage, meaning, let me give you one example, our 3-nanometer and 5-nanometer are adjacent to each other, the fabs, and they are all connected. So it's much easier for TSMC to convert from 5 to 3. And that doesn't mean that every node can do the same” CEO of TSMC, C.C. Wei at the Q1 2024 call

Both N3 (Phase 5 to 9) and N5 (Phase 1 to 4) are located at Fab 18 in Tainan, so they are beside each other. N5 demand has been relatively stagnant given customers migration to N3. For N2, TSMC is ramping Fab 20 at Baoshan, Hsinchu at the end of 2025 and another Fab 22 at Kaohsiung Science Park in 2026. Fab 20 in Hsinchu is beside the R&D fab and is far away from the N3 capacity located at Fab 18 in Tainan. For Fab 22, the plan is 3 phases for N2 and 2 phases for N1.4. This is not forgetting that N2 will be TSMC’s 1st node to transition to GAAFET.

In addition, N3 revenue will grow more than 3x yoy in 2024 led by the ramp of Mediatek’s Dimensity 9400, Qualcomm Snapdragon 8 Gen 4, Apple and Intel’s client CPU tiles (Arrow Lake and Lunar Lake). N3 is a long-lasting node with ASIC migration in 2H 2025 and potentially X100 (also possible to be N2) from Nvidia.

In order to ramp Fab 20 in Hsinchu and Fab 22 in Kaohsiung, TSMC will have to order EUV machines from ASML over the next few quarters. Demand is strong for N2 as it is the industry leading node.

“We are observing a high level of customer interest and engagement at N2 and expect the number of the new tape-outs from 2-nanometer technology in its first 2 years to be higher than both 3-nanometer and 5-nanometer in their first 2 years.” C.C. Wei

DRAM suppliers will also be increasing their EUV intensity as I wrote in this article: Memory demand for EUV. Intel will also be ramping up 18A as they are bringing Panther Lake back to IFS in 2026. Given the strong demand from customers, ASML is very likely to hit the expectation in 2025 and 2026.

TSMC

1. N7 to recover

“we expect the N7 in the next couple of years, it will pick up, the demand will pick up again.”

N7 has been weak for TSMC as smartphone clients undergo inventory digestion. Wi-Fi 7 is coming in 2024 and this will be done at N7. Wi-Fi 7 will be a slower ramp than Wi-Fi 6 but will be able to boost demand at N7.

Inventory correction is also over for smartphone customers and Mediatek will be providing Dimensity 6100 at N6 for mid-range 5G phones.

2. Capex guidance unchanged for now

Expectation was for capex guidance to be raised, but TSMC sticked to the original $28-32bn capex guidance for 2024. This creates another concern for WFE the day after ASML reported.

In Q4 2023 call, CFO Wendall Huang said “And this year, if you do the math, is going to be mid-30s. We expect, in the next several years, it will remain around mid-30 percentage capital intensity.”

Since 2009, TSMC’s capital intensity has fluctuated between 30% to 53%. Looking at consensus, 2024 revenue will be USD 84bn and 2025 revenue will be USD 100bn. This number is conservative, but let's use this number for capex calculation for now.

Assuming 35% capital intensity and the conservative consensus numbers, we will be looking at a capex of USD 30bn for 2024 and USD 35bn for 2025. With a strong ramp of N2 in 2026, capex number will grow from here. Revenue growth of TSMC is the leading indicator of WFE, not TSMC’s current capex forecast.

3. AI to be the main driver of revenue growth

“We forecast the revenue contribution from server AI processors to more than double this year and account for low-teens percent of our total revenue in 2024. For the next 5 years, we forecast it to grow at 50% CAGR and increase to higher than 20% of our revenue by 2028.”

“Almost all the AI innovators are working with TSMC to address the insatiable AI-related demand for energy-efficient computing power.”

Yes, TSMC is dominating all the AI processors from N3 to N7 with its process leadership. In my X’s post, I have listed out the AI customers for TSMC in 2024. This is based on TSMC’s definition of “GPUs, AI accelerators and CPUs performing training and inference functions.”

In 2025, next generation Google’s TPU V6, Nvidia’s X100, AMD’s MI350, Amazon’s Inferentia and Trainium 2.5 and 3 will continue to be produced at TSMC. Remains to be seen if Intel will outsource Falcon Shore to TSMC, although Gaudi 3 will be done by TSMC.