Intel
2026-07-20
Intel Foundry 18A Yields Jump to 85%+; EMIB Packaging Hits 98%, Challenging TSMC N2
Summary
Key Takeaways
According to a KeyBanc report, Intel Foundry's 18A node yields jumped from ~65% to over 85% in a single quarter, approaching TSMC N2's 90%. 18A is Intel's first high-volume RibbonFET GAA and PowerVia backside power delivery node. The 18A-P variant is in risk production with +9% performance and -18% power. 14A targets 2028 risk production, 2029 HVM.
In advanced packaging, EMIB yields reached 90%, EMIB-T 98%, Foveros 3D 80-85%. CFO Zinsner sees packaging contributing billions by 2027.
Customer signings: NVIDIA (18A/14A), AMD (18A), Apple (18A-P), Microsoft, OpenAI (18A+14A). Google/Amazon interested in EMIB. However, these are mostly secondary supply; Apple/NVIDIA still rely on TSMC. Yield numbers may be based on specific test chips, not large AI accelerators.
Why It Matters
Intel's yield PR is a defensive move against TSMC, using EMIB packaging to lock in multi-die integration. It hides that the 85% yield likely comes from small test chips, not reticle-sized AI accelerators. Customer signings are secondary, not primary. PowerVia and RibbonFET are unproven in large AI training chips regarding tail latency and thermal management. EMIB-T 98% is good but Foveros 3D at 80-85% is a bottleneck. Adopting EMIB risks toolchain lock-in and single-source capacity, less flexible than TSMC's CoWoS. Investors: Intel Foundry still losses; 2027 breakeven depends on real orders; 14A misses NVIDIA Rubin window.
PRO Decision
【Vendors】TSMC should publish direct benchmark comparisons against Intel 18A, especially large-die yields and performance, emphasizing N2's mature ecosystem and CoWoS capacity. Samsung must accelerate SF2 yield improvement to avoid third-tier status.
【Enterprises】CIOs/architects should conduct independent audits of Intel's yield data, demanding reticle-sized die data. Assess toolchain lock-in risks from EMIB and maintain multi-sourcing with TSMC/Samsung. For AI training chips, prioritize proven N2 and CoWoS to avoid delays.
【Investors】Recognize that Intel's yield improvement is catch-up, not leadership; customer signings are hedging, not commitment. Intel Foundry's profitability depends on 14A orders by 2028-2029; losses will continue. Packaging revenue from 2027 is reliant on few customers. Focus on TSMC's expansion and Intel's cash flow, not short-term yield news.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)