Direct:
Anton Shilov's roadmap piece at Tom's Hardware (Premium) lays out where TSMC, Intel, and Samsung are headed after all three crossed into 2nm-class production over the last twelve months. His framing is clean: TSMC is running on execution discipline and segmentation, Intel is making the most architecturally ambitious bet with GAA and backside power delivery shipping together, and Samsung is in a yield-fix posture rather than chasing breakthroughs. That framing holds up, but a few of the secondary implications deserve a harder look.
The strongest part of the piece is the segmentation read on TSMC. Splitting N2P (frontside power) from A16 (with BSPDN), keeping a parallel cost-optimized track in N4C, N3C, and eventually N2C, and running automotive a generation or two behind on N7A, N5A, N3A is the kind of move only a foundry with TSMC's customer base can pull off. Shilov notes the unusual two-fab simultaneous N2 ramp but doesn't quantify the yield foundation underneath it. TSMC has publicly described N2 as ramping with good yield, and third-party reporting around the January 2026 HVM milestone put early N2 yields in the mid-60s to mid-70s percent range.
Intel's section is where the framing gets thinner. The piece argues Intel admits world-class yields won't arrive until 2027, which “likely makes Intel's 18A node significantly less attractive to demanding customers.” Defensible, but it skips the curve. John Pitzer's commentary in late 2025 put 18A on the industry-average 7 percent per month yield improvement track, with Panther Lake as the proving vehicle. The interesting question isn't whether 18A is competitive against N2 today, because it isn't. It's whether 14A's High-NA EUV bet pays back the risk in 2027 to 2028. Musk's Terafab apparently pursuing a licensing-style 14A arrangement is the tell the article mentions but doesn't pressure. A licensing posture is a hedge that lets the customer walk if yields don't show.
The Samsung framing is where I'd push back hardest. Shilov suggests SF1.4 in 2027 could let Samsung “formally leave Intel and TSMC behind.” That's a generous reading. SF2 yields sat around 50 percent at Exynos 2600 launch, and Samsung's own guidance to its Design Solution Partners has shifted toward promoting SF2P rather than SF2. The pellicle question Shilov raises is the real structural one: at sub-2nm, stochastic mask-borne defects dominate the yield budget, and Samsung's continued reluctance to deploy EUV pellicles is a drag, not a roadmap footnote. Tom's own reporting suggests the Taylor, Texas fab may be Samsung's first qualified EUV-pellicle flow, though that hasn't been officially confirmed. Without a qualified pellicle flow, SF1.4 in 2027 reads more like a roadmap claim than a bankable customer platform.
The fractured-futures conclusion is right. The interesting consequence the piece leaves on the table is that “leading edge” is no longer a single ladder. Customers now pick a foundry by which kind of risk they want to absorb: TSMC's pricing and capacity allocation, Intel's execution variance, or Samsung's yield ceiling.
Sources:
- Primary: Anton Shilov, “Leading-edge foundry roadmaps for TSMC, Intel and Samsung,” Tom's Hardware Premium, May 14, 2026.
- TSMC N2 ramp and customer mix: Tom's Hardware reporting on KLA's Ahmad Khan at Goldman Sachs Communacopia (Sept 2025).
- TSMC N2 early yield ranges: Third-party HVM coverage (FinancialContent, Jan 2026), treated as directional given third-party-accuracy caveats.
- Intel 18A yield curve: John Pitzer commentary reported by OC3D (Nov 2025); David Zinsner remarks via Tom's Hardware (Aug 2025).
- Samsung SF2 yields and SF2P pivot: ZDNet reporting summarized by Wccftech (Jan 2026); TweakTown coverage of Samsung Q3 2025 earnings (Nov 2025); Fudzilla on Exynos 2600 yield trajectory (Feb 2026).
- Samsung Taylor pellicle context: Tom's Hardware reporting on Samsung's Taylor, Texas fab production flows.
Reddit: https://ift.tt/bZuvCm8