SILICON NEXUS
Research NotesTaiwan· Sep 2, 2026· 2330· 6 min read

The 14x Reticle — The Week TSMC Made the Package the New Chip

CoWoS goes from 3.3x reticle to 14x, system power 6x. TSMC's SEMICON roadmap redrew the thermal, power, and substrate stack all at once.

TSMC CoWoS Package Size Roadmap — Reticle MultiplesTaiwan Thermal & Power Sub-Chain — Reading the 6x Power Roadmap

Summary

At SEMICON Taiwan on September 1, 2026, TSMC's advanced-packaging R&D chief laid out a roadmap that quietly ends a 30-year physical constraint in this industry: the reticle limit. CoWoS package size, which sits at ~3.3x reticle today, is guided to 5.5x by year-end, 8x by 2028, and 14x reticle by the end of the decade. In the same session, director Chen Yan-Ming forecast total AI system power will rise 6x within five years. Those two numbers, disclosed side by side, are why every layer of the thermal, power, and photonic stack in Taiwan is being repriced this week.

The supply chain moved immediately. Silicon wafers logged their first simultaneous price hike across 6"/8"/12" in over three years — advanced 12" for AI ran into double-digit increases. Asia Vital (1514) reported a record NT$13B (~US$400M) backlog with the semiconductor share now at 35%. Shuang Hong (3324) guided 2026 revenue growth of 70% with the NVIDIA VR-series liquid-cooling ramp peaking in Q4. Sunnex (5274) posted an all-time high August at NT$611M as a TSMC supply-chain vendor.

Why this week is different — the end of a 30-year grammar

For three decades, chips were defined by one number: the 858 mm² reticle field of a lithography scanner. CoWoS was the first product that broke that wall, but the escalation was gradual — B100/B200 sat at 3.3x, VR-Ultra reached 5.5x. What TSMC disclosed this week is an acceleration of that curve. A 14x-reticle package is not simply a bigger chip; it is the arrival of a System-on-Substrate era where the package, not the die, is the unit of design.

The roadmap carries three concrete consequences.

First, HBM5 mounts directly on N3 logic. Until now HBM has lived beside the package; TSMC's chart puts HBM5 in a 3D stack on top of N3 logic dies. That precisely matches SK Hynix SVP Hoshik Kim's SEMICON assertion the same week that KV cache — not model parameters — is now the dominant memory bottleneck in agentic AI: bandwidth and capacity are needed simultaneously, not sequentially.

Second, COUPE (Compact Universal Photonic Engine) enters commercialization at 4 Tbps. Optical interconnect moves from beside the GPU to inside the package. France's Soitec, holding ~95% share of the silicon-photonics SOI substrate market, has already locked a US$200M revenue floor via 10+ multi-year fixed-price supply agreements — the substrate layer is being reserved before customers even name themselves.

Third, TIM-less cooling and integrated power delivery. TSMC and AMD jointly disclosed a next-generation cold-plate roadmap that eliminates thermal interface materials entirely. Taiwan's thermal ecosystem is aiming at a mid-2027 ramp.

Heat and power are the new bottleneck — and the pricing shows it

The winners of this redesign are already visible in the tape. Wantai Technology (6190) told analysts H2 revenue and gross margins are rising in parallel, driven by PCIe 5.0/6.0 AI cable shipments alongside LEO satellite orders. Teco Electric (1504) guided all four segments — electromechanical, power energy, HVAC, and energy automation — to double-digit H2 growth on data-center power demand. Xunget/Sunnex (6438/5274) confirmed PCB-equipment orders from 4958 and 3189.

Memory has begun locking this roadmap into contracts. Samsung has committed ~70% of memory capacity through 2031 in long-term agreements, with NVIDIA, Microsoft, and Google as anchor customers. HBM spot prices have run 4–5x. The top-five enterprise SSD brands doubled Q2 revenue to US$37.6B (+103.6% QoQ). DDR5 16Gb spot at $54.25 (today) remains in an uptrend consistent with these contracts. This is the demand side of the 14x-reticle stack signing itself into the next five years.

Placing the NVIDIA–MediaTek deal correctly

The US$3.5B NVIDIA subscription of MediaTek's convertible bonds absorbed the week's headlines, but inside this report's frame the deal is a sales channel for the 14x-reticle roadmap. Opening NVLink Fusion, Spectrum-X switching, and the NVHBI interface to MediaTek creates a second design channel for cloud XPUs that scale out on top of TSMC 3nm + CoWoS 14x. Microsoft Azure Global Infrastructure GM Alistair Speirs, visiting for SEMICON, distilled it in one line: "No Taiwan, No Maia 200." The 3nm logic, the CoWoS packaging, the TIM-less cooling, the wafer supply, the ABF substrate, the photonic engine, and the liquid cooling all have to sit on one island for the system to close.

PM view — three axes of re-rating

Axis 1 (foundry + packaging core). TSMC (2330) captures logic margin, HBM margin, and photonic-engine margin simultaneously under 14x-reticle economics. That is a different story from the last cycle's "node margin" narrative — this is package margin.

Axis 2 (thermal + power sub-chain). Asia Vital's NT$13B backlog, Shuang Hong's 70% guide, and Teco's four-segment upgrade are not idiosyncratic prints; they are the first inventory move against a 6x-power roadmap.

Axis 3 (materials + substrate sub-stack). LCY Advanced Materials' sub-10μm wet chemistry for advanced packaging, Innolux's (3481) Chip Last / FOPLP with H2 2027 mass-production targets, and Tenhong's (6937) new Advanced Process Laboratory (opened Sept. 1 with a TSMC R&D VP in attendance) are the actual physical layer that has to exist for 14x reticle to ship.

Risks

The roadmap is not free. Micron Taiwan's union, representing ~10,000 of ~15,000 employees, has cleared 80% strike support heading into mandatory mediation over an 83-month bonus dispute — Taiwan hosts roughly 60% of Micron's global output. Qualcomm implemented double-digit chip price increases effective September 1. Huawei, Xiaomi, and Honor simultaneously raised handset prices by up to RMB 1,000 on memory cost pressure. The 14x-reticle / 6x-power roadmap will keep pushing the cost curve up; who absorbs the pass-through decides the next quarter's margin story. Taiwan's advantage here is that the same island that owns the scaling roadmap also owns the pricing power at every layer of it.

Key Sources: - TSMC Lays Out Full-Dimension Scaling Roadmap: CoWoS to 14x Reticle, HBM5 on N3, COUPE at 4 Tbps (technews, 2026-09-01) - TSMC: AI System Power to Surge 6x in Five Years, Driving Shift to Liquid Cooling and Integrated Power Delivery (cnyes, 2026-09-01) - TSMC, AMD Reveal TIM-less Cold Plate Plans; Taiwan Thermal Makers Eye Mid-2027 Ramp (cnyes, 2026-09-01) - Microsoft Azure GM: 'No Taiwan, No Maia 200' as TSMC 3nm Powers Custom AI Chip (technews, 2026-09-01) - Silicon Wafers See First All-Size Price Hike in 3+ Years on AI Demand Surge (cnyes, 2026-09-01) - Asia Vital (1514) Backlog Hits Record NT$13B; Semicon Now 35% of Orders (cnyes, 2026-09-01) - plus 24 more

If this analysis was helpful · Support Us · ✈️ Telegram