Foundry Recovery Outside TSMC, Strong Teradyne Testing Demand

Overview

This report is based on Morgan Stanley's research meeting minutes from Taiwan, focusing on the latest developments in the semiconductor capital equipment industry. The core conclusion is that foundry spending is expanding and no longer relying solely on TSMC; Intel and Samsung's recovery brings new vitality. Meanwhile, the testing segment continues to be in short supply. Teradyne benefits significantly as it holds an advantage in NVIDIA's next-generation chips and networking products, maintaining strong demand. The report maintains its "In-Line" rating for the semiconductor capital equipment industry, while holding a more positive outlook on specific sub-segments such as network testing.

Core views

The dynamic changes in the foundry competitive landscape see Intel and Samsung regaining vitality. Surveys indicate that Intel's EMIB (Embedded Multi-Die Interconnect Bridge) technology demand far exceeds its own foundry business, with MediaTek, Trainium and other clients having extremely strong demand for TSMC alternatives. However, on the foundry side, Apple seems to be the only high-profile customer, expected to mass-produce A-series chips in 2029; Tesla may place AI6 orders before the TeraFab goes online. Currently, Intel's 18A process yield rate is about 50%, still striving to improve, with slow progress in customer acquisition. On the other hand, Samsung's foundry business also performs well, with Google expected to mass-produce CPUs manufactured by Samsung in the second half of 2028, but the challenge is that almost all US customers prefer to shift capacity to the Taylor factory which is still under construction. The testing segment continues to be a bottleneck, benefiting Teradyne significantly. Overall test environment demand far exceeds supply. Teradyne has received initial probe test orders for NVIDIA Rubin chips and is testing most applications for NVIDIA outside of GPU final test (FT), including Mellanox, RTX, Vera CPU, and LPU/Groq. As Teradyne gradually obtains RTX, LPU, and chip probe testing work, its share in NVIDIA is unlikely to change significantly under the Rubin Ultra or Feynman architecture. Notably, GPU final test time may increase from about 850 seconds in Blackwell to 1200 seconds in Rubin, but this does not make Teradyne indispensable in the GPU FT position because testing demand related to Mellanox is sufficient to keep Teradyne busy. CPO test landscape remains unclear, demand is strong in other fields. Regarding Co-Packaged Optics (CPO), the situation is fluid rather than settled. ATE share in insertion steps has not been finally determined, and test times are uncertain, making 2027 demand signals more speculative. Currently understood is that Insertion Steps 1 and 2 fall under TSMC jurisdiction, where Insertion Step 1 (EIC/PIC) is led by Advantest, and Insertion Step 2 favors Teradyne, but results remain open. Insertion Step 4 seems to belong to Teradyne, while Advantest is contacting AMD and Amazon. Besides NVIDIA and CPO, China and Tesla are two fields where demand remains strong.

Analysis framework

The institution verifies Taiwan industry chain upstream and downstream through field surveys, adopting a bottom-up verification method. First, by comparing client feedback and yield data of different Intel business lines (EMIB vs Foundry), unravel their true competitiveness; second, track the testing process changes of NVIDIA's new generation chips (Rubin), quantify the impact of increased test duration on equipment vendors, and distinguish equipment vendor shares across different testing stages (probe, final test, network testing); finally, combine the CPO technology evolution roadmap to analyze the equipment competition landscape in each insertion step, thereby deriving the structural impact on equipment vendors like Teradyne and Advantest.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply Chain Upstream-Midstream-Downstream Transmission

  Transmission of Chip Design Changes to Test Equipment Demand

  The report analyzes the evolution of NVIDIA chip architecture from Blackwell to Rubin, especially the increase in test duration from 850 seconds to 1200 seconds, derives its specific impact on test equipment utilization and vendor share, demonstrating how upstream design changes transmit to midstream equipment demand.

• Competition and Strategy FrameworkMoat / competitive advantage

  Technology Binding and Client Stickiness Analysis

  The report points out that Teradyne's deep embedding in NVIDIA's non-GPU applications (such as Mellanox) allows it to maintain high workload even without increasing its GPU final test share, revealing the competitive barrier constructed by equipment vendors binding top clients through multi-product lines.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Teradyne (Teradyne)

  Beneficiary

  Strengths

  Received NVIDIA Rubin chip probe test orders; covers non-GPU applications such as Mellanox, RTX; testing demand continues to exceed supply.

  Weaknesses

  Share in NVIDIA GPU final test (FT) may not increase significantly; CPO test share is undetermined.

  Comparison

  Compared to Advantest, occupies advantages in non-GPU network and probe testing areas; Advantest has more opportunities in insertion step 1 and AMD/Amazon.

  Risks

  CPO test landscape changes; 2027 demand signals are speculative.

• Advantest (Advantest)

  Structural Beneficiary

  Strengths

  Dominates Insertion Step 1 (EIC/PIC); contacting AMD and Amazon.

  Weaknesses

  Faces competition from Teradyne in Insertion Step 2; CPO subsequent step share undetermined.

  Comparison

  Competitive in specific test insertion steps and specific clients (AMD/Amazon), forming differentiated competition with Teradyne.

  Risks

  CPO test landscape uncertainty; potential penetration by Teradyne in Insertion Step 2.

• Intel (Intel)

  Differentiated

  Strengths

  EMIB demand is extremely strong, customers include MediaTek, Trainium, etc.; Expected to obtain Tesla AI6 orders.

  Weaknesses

  Foundry business customer acquisition progress is slow, only Apple is high-profile client; 18A yield rate is only 50%.

  Comparison

  EMIB business performance far exceeds foundry business, uneven internal development.

  Risks

  Yield rate improvement falls short of expectations; loss of key clients or delayed mass production.

• Samsung (Samsung)

  Beneficiary

  Strengths

  Foundry business status is good; Expected to obtain Google 2028 CPU orders.

  Weaknesses

  US customers tend to transfer to Taylor factory, but plant capacity is still under construction.

  Comparison

  As a major alternative choice outside TSMC, benefits from client diversification strategy.

  Risks

  Taylor factory capacity construction schedule; changes in client preferences due to geopolitical issues.

Key data

• Intel 18A Yield Rate~50%Still working on improvement
• Apple Foundry Commitment2029Expected mass production of A-series chips
• Google CPU FoundryLate 2028Expected mass production using Samsung foundry
• GPU Final Test Duration Change850 seconds -> 1200 secondsIncrease from Blackwell architecture to Rubin architecture

Impact & implications

The report considers that diversified foundry spending is beneficial for alleviating reliance on single suppliers; Intel and Samsung's recovery provides new growth points for equipment vendors. For Teradyne, even if its share in NVIDIA GPU final test does not increase significantly, its strong demand in network chips and other application areas is enough to support performance. The uncertainty of the CPO testing landscape means low earnings visibility for related equipment vendors in the short term, but in the long run, increased testing complexity will benefit leading manufacturers with comprehensive solution capabilities.

Risks

• CPO test landscape has not been定型，2027 demand signals are highly speculative.
• Intel foundry yield rate improvement and customer acquisition progress fall short of expectations.
• Samsung Taylor factory capacity construction delays affect US client order landing.

What to watch

• Further improvement of Intel 18A yield rate and new customer (such as Tesla) order landing.
• Final determination of equipment vendor shares in CPO test steps (Insertion 1-4).
• NVIDIA Rubin chip mass production progress and actual impact on test duration.
• Sustainability of semiconductor test demand in China and Tesla fields.