Foundry Capex Broadens, Teradyne’s Non-GPU Pipeline Underappreciated

Core Conclusion

Taiwan checks confirm resurgent Intel and Samsung foundry efforts are broadening equipment demand well beyond consensus, while test capacity remains a bottleneck. Teradyne’s expanding order book outside NVDA AI GPU final test—Mellanox networking, RTX, Vera CPU, LPU/Groq, and Rubin chip probing—strengthens its revenue visibility. The market’s fixation on a static GPU test share obscures this diversification and the capex tailwind from foundry competition.

What the Market May Be Missing

Consensus underestimates two forces: the spending pull from Intel and Samsung’s foundry push, and the scale of Teradyne’s non-GPU test demand. Investors over-penalize TER because it does not capture incremental AI GPU test share, overlooking that networking and custom ASIC test volumes are absorbing all available capacity.

Evidence Chain

Foundry competition fuels capex. Intel 18A yield hovers near 50%, with Apple expected to use it for A-series HVM in 2029. EMIB demand from MediaTek, Trainium, and others is extremely strong—far ahead of Intel’s foundry customer conversion. Samsung is active; Google will likely use Samsung for CPU HVM in 2028. Nearly all U.S. customers want to direct orders to Samsung’s Taylor fab, though capacity is still under development. This broadening race lifts equipment spend beyond current models.

Test bottleneck intensifies, Teradyne diversifies. Test supply remains tight, with demand outpacing capacity. Teradyne has won initial NVDA orders for Rubin chip probing and now handles a majority of NVDA applications outside GPU final test: Mellanox, RTX, Vera CPU, and LPU/Groq. Mellanox-related demand alone keeps Teradyne extremely busy. As GPU final test times climb from ~850 seconds (Blackwell) to ~1,200 seconds (Rubin), Teradyne’s absence from GPU final test is irrelevant.

NVDA GPU share static, but a non-issue. With Rubin Ultra or Feynman, Teradyne’s AI GPU test share is unlikely to move materially. Incremental gains in RTX, LPU, and probing fill the gap, making static share a stale narrative.

CPO test optionality. The landscape remains fluid. Insertion 2 skews toward Teradyne but is not settled due to technical difficulty. Insertion 4 for NVDA applications is controlled by Teradyne, while Advantest serves AMD and Amazon. Test times and share allocations are undecided, so 2027 demand projections are speculative. CPO provides an option, not a base-case driver.

Key Risks

• Intel’s foundry customer acquisition remains a work in progress; volume orders beyond Apple are uncertain.
• Samsung’s Taylor fab capacity development may lag, slowing U.S. customer migration.
• CPO test volumes and ATE share are speculative pre-2027; overreliance generates disappointment risk.
• Persistent negative sentiment if investors fixate on static GPU test share, discounting other growth engines.

Valuation and Trade Implications

Teradyne’s expanding networking and custom ASIC test exposure, combined with structurally tight capacity, supports above-consensus revenue visibility. The foundry capex upcycle provides additional tailwinds. We view TER as a primary beneficiary, with further upside from eventual CPO test adoption. Advanced packaging equipment names linked to the foundry race also warrant attention.

Appendix: Condensed Data Tables

CPO Test Insertion Ownership

GPU Final Test Time