$Himax(HIMX)$ is the exclusive supplier of critical wafer-level optics (WLO) components inside TSMC’s COUPE (Compact Universal Photonic Engine) platform, the core optical engine for its Co-Packaged Optics (CPO) architecture. This directly replaces power-hungry copper SerDes links with integrated light-based data transfer between chips/switches.
TSMC COUPE CPO Architecture — What It Actually IsCOUPE is TSMC’s production-ready silicon-photonics integration platform (developed with NVIDIA). It heterogeneously bonds a Photonic IC (PIC on SOI N65) with an Electronic IC (EIC on N6/65nm) using SoIC-X hybrid bonding, then co-packages the entire engine with the host ASIC/switch via CoWoS. Light comes from external CW laser source modules (detachable OSFP-style for serviceability), couples through a Fiber Array Unit (FAU/ReLFACon) into the PIC (using micro-ring modulators at 200 Gbps PAM4 per lane), and delivers 1.6 Tbps+ per engine (scaling to 3.2–12.8 Tbps+ per subassembly).
Key innovation vs. copper/pluggables: sub-millimeter electrical traces (vs. cm/inches) slash power (e.g., 30 W → 9 W per 1.6T link, or ~73% savings), latency, and heat while boosting density. NVIDIA is already shipping: Quantum-X InfiniBand CPO switches (H2 2025) and Spectrum-X Photonics Ethernet switches (H2 2026). Early volumes are ramping; full AI-cluster adoption accelerates in 2026–2028.
Himax’s Precise (and Locked-In) PositionHimax does not fabricate the PIC, EIC, or lasers. Instead, it owns the precision optical-coupling layer inside the FAU:Proprietary WLO nano-imprint lithography on glass wafers produces microlens arrays (MLA), 45° prisms, and V-groove baseplates.
These bend/align/focus external laser light into the COUPE PIC with sub-micron accuracy.
FOCI (Fiber Optic Communications) assembles the final FAU and integrates it with the engine. Himax holds a 5.3% equity stake in FOCI and co-designs the optics block.
Exclusivity is ironclad for early generations:
Ming-Chi Kuo (latest supply-chain checks, updated through 2025–2026): Himax is the sole MLA supplier for COUPE Gen 1 and Gen 2 FAUs. FOCI is the sole FAU provider. A second source is “unlikely before Gen 3 at the earliest.”
Himax’s own Jan 5, 2026 press release: “Himax is an important collaboration and strategic partner of FOCI… in the field of Co-Packaged Optics (CPO).” Gen 1 validation ongoing; mass-production readiness targeted for 2026. “All ongoing collaborations… remain active and unchanged.”
Recent independent confirmation (March 2026 analysis): Himax/FOCI process is already mass-production capable; NVIDIA Rubin Ultra CPO racks projected at thousands of units in 2027–2028, driving FAU demand into the billions.
Contribution path: Technical qualification/trial production (modest 2024–2025) → rapid ramp in 2026. Kuo models Himax revenue at $1.16B (2026), $1.42B (2027), $2.4B (2028) with EPS $1.00 / $1.60 / $3.40 — well above consensus — purely from COUPE optics (assuming 30–40% of FAU BOM at 45–50% gross margin).
(Note: TSMC has demonstrated metalens/metasurface coupling in lab papers for future efficiency gains. This is not displacing Himax-style traditional MLA/WLO in Gen 1–2 production; metalens is positioned for Gen 3+ evolution.)
tspasemiconductor.substack.com
If (When) This Scales: Exact Reshaping of the Photonics Stack
CPO/COUPE is already the biggest architectural shift since pluggables replaced direct-attach copper. The traditional “discrete module” chain compresses dramatically:Lasers ($Lumentum(LITE)$ , $COHERENT(COHR)$ ): Neutral to positive. External high-power CW lasers remain mandatory (detachable for field replacement; ~70% of engine power budget). Nvidia’s partner list explicitly includes Lumentum and similar suppliers; higher-volume AI deployments drive demand even as per-system laser count consolidates. On-chip lasers are avoided due to heat/failure risks.
Transceivers ($Applied Optoelectronics(AAOI)$ , $Fabrinet(FN)$ ): Most disrupted. Pluggable modules (the entire AAOI optics + FN contract-manufacturing model) are bypassed for intra-rack and many inter-rack links. CPO eliminates the DSP (previously ~50% of transceiver power/cost) because electrical distances shrink to millimeters. Hyperscalers and Nvidia are moving volume to integrated engines — classic pluggable TAM shrinks as CPO scales
@Tiger_comments @TigerStars @TigerPM @Daily_Discussion @TigerObserver
Comments