0

Intel Foundry closes the gap with TSMC on yield and packaging

In sum – what we know:

• A yield leap – Intel’s 18A yields reportedly climbed from roughly 65% to over 85% in a single quarter, closing in on TSMC’s N2.
• Big-name design wins – Nvidia, AMD, Apple, Microsoft, and OpenAI are reportedly attached to 18A/14A projects, mostly as hedges rather than primary sourcing.
• Packaging strength – EMIB yields are around 90% (EMIB-T near 98%), turning what was once a bottleneck into a selling point.

For years, the story around Intel Foundry was a question of whether the company could fix its yields fast enough to matter. That question now appears to have an answer. Intel’s 18A process node and its advanced packaging technologies have reportedly hit a turning point in both yield and capacity, and the numbers coming out of analyst channels reflect a business that has moved from “acceptable” to “strong” in the space of a few quarters.

A wave of reported design wins and foundry orders has landed at Intel’s door, with names like Nvidia, AMD, Apple, Microsoft, Google, and OpenAI now attached to 18A and 14A projects. That combination of node progress and packaging progress is reshaping expectations for Intel Foundry’s path to profitability, and it strengthens Intel’s position as a viable leading-edge alternative to both TSMC and Samsung — though it still trails TSMC on raw process yields.

Much of the credit is going to CEO Lip-Bu Tan, who has said 18A yields were “not good” when he took over. Tan pushed for 7–8% month-over-month yield improvements and brought in external optimization partners like PDF Solutions and KLA to accelerate the learning curve. That approach appears to have paid off faster than most expected.

Yield surge across 18A and advanced packaging

The headline number comes from KeyBanc Capital Markets, which reportedly found that 18A yields jumped from roughly 65% to more than 85% in a single quarter. That’s a remarkable step up for a leading-edge node, and it puts Intel in striking distance of TSMC’s N2, which currently yields around 90%. It also puts Intel comfortably ahead of Samsung, whose SF2 yields are reported at just 50–60%. In other words, Intel has arguably become the clear second-best option at the extreme leading edge — behind TSMC, but well clear of Samsung.

The technical foundation matters here. 18A is Intel’s first high-volume node to combine RibbonFET gate-all-around transistors with PowerVia backside power delivery, giving it concurrent power and performance gains that neither TSMC nor Samsung offers in quite the same configuration at this generation.

Advanced packaging is where the story gets even more interesting. Packaging was a meaningful bottleneck for Intel not long ago — and for the industry broadly, as anyone watching TSMC’s overbooked CoWoS capacity can attest. That’s changed. EMIB yields are now reported at roughly 90%, and the newer EMIB-T variant is rumored to yield near 98%, which would make the packaging layer more robust than the underlying die yields in some cases. Pilot runs across the broader portfolio, including Foveros 3D die stacking, are landing at 80–85%. For large multi-die designs, where a single packaging defect can scrap several perfectly good dies at once, those numbers dramatically reduce risk — and they’ve become a genuine selling point rather than a liability.

Historic design wins and customer roster

The customer list reads like something that would have been unthinkable a few years ago. Nvidia has reportedly engaged Intel Foundry for at least one high-end design on 18A or 14A, likely for AI accelerators or related infrastructure chips. AMD has at least one reported design win of its own, likely leveraging Intel’s advanced chiplet packaging. Watching Intel’s two longest-standing rivals become foundry customers is a striking role reversal, whatever the volumes turn out to be.

Apple is reportedly eyeing the enhanced 18A-P variant for lower-end processors used in MacBooks and iPads. OpenAI is developing custom AI accelerators and inference chips tuned to its workloads on 18A and 14A, while Microsoft, Micron, and Marvell are undertaking 18A projects for memory-adjacent and networking silicon. Google and Amazon, meanwhile, are particularly interested in EMIB and EMIB-T for complex multi-die AI systems — which suggests some hyperscalers may use Intel primarily as a packaging provider, potentially for dies fabbed elsewhere.

The caveat runs through all of it. Heavyweights like Apple and Nvidia are keeping TSMC as their primary manufacturing partner for flagship products, using Intel largely for diversification and risk mitigation. These are hedges and second-source arrangements, not migrations. But given where Intel Foundry was two years ago, hedges from the industry’s biggest names still count as real progress.

Future roadmap

The financials are starting to reflect the momentum, if slowly. Intel Foundry’s operating loss narrowed to $2.4 billion in Q1 2026 — a $72 million quarter-over-quarter improvement aided by better process yields. That’s still a substantial loss, but the direction matters. The company’s oft-cited 2027 breakeven target is increasingly viewed as realistic given rising yields and a strengthening 14A pipeline, and CFO David Zinsner has indicated that advanced packaging commitments alone could generate billions in revenue starting in 2027.

On the technology side, the performance-enhanced 18A-P variant has entered risk production, targeting 9% higher performance at the same power, or 18% lower power at equivalent performance — the characteristics that likely explain Apple’s reported interest for power-sensitive Mac and iPad silicon. 

Further out, 14A is targeted for risk production in 2028 and volume manufacturing in 2029. Tan has positioned it as the vehicle for Intel to fully match or surpass TSMC’s most advanced nodes by the end of the decade — a claim worth treating as an ambition rather than a certainty, given Intel’s roadmap history. 

Skepticism

There are real reasons to keep the champagne corked. The optimistic yield figures may be based on specific product types or internal test chips, and actual yields on massive, complex external AI accelerators could diverge significantly — a reticle-sized Nvidia-class die is a very different animal from a Panther Lake CPU tile. And that points to the more fundamental gap in the story so far. High-volume external 18A parts aren’t shipping yet; current 18A output primarily supports Intel’s own Panther Lake CPUs.

The customer roster deserves similar scrutiny. Much of the demand is described with qualifiers like “reportedly” and “assessing,” with limited public confirmation from the clients themselves. Design wins are not production wins, and the 2027 breakeven goal depends entirely on these engagements translating into sustained, high-volume manufacturing runs rather than cancelled projects. Anyone who followed Intel through the past decade knows how roadmaps can slip.

Advanced packaging carries its own risk. Capacity that looks healthy today could quickly become a bottleneck again if external AI demand scales faster than Intel’s facility expansion in Malaysia and elsewhere — especially if the company has to support both internal and external packaging loads simultaneously.

And then there’s TSMC. Even a narrowed yield gap is still a gap, and customers may continue relegating Intel to secondary or specialized sourcing roles for their most critical flagship designs. The fact is, Intel Foundry has earned its seat back at the table.