Silicon Nexus

NVIDIA CEO Jensen Huang announced full-scale production of the next-gen Vera Rubin AI platform, succeeding Blackwell. The article surveys Taiwanese supply chain beneficiaries spanning foundry, CoWoS advanced packaging, ABF substrates, networking and cooling, framing Vera Rubin as the next leg of AI infra capex.

Why it matters: Sector-wide AI infra theme tied to NVIDIA's Vera Rubin roadmap and Taiwanese supply chain beneficiaries, not a fresh earnings-grade catalyst.

Ayar Labs is joining NVIDIA's NVLink Fusion ecosystem, aligning its co-packaged optics (CPO) and SerDes technology with NVIDIA platforms so hyperscalers can build optical-interconnect-centric rack-scale AI infrastructure including custom silicon. The tie-up validates CPO as a foundational building block for next-gen NVIDIA AI factories, with implications for advanced packaging and AI-server supply chains as bandwidth, latency and power become binding constraints over copper interconnects.

Why it matters: Ecosystem/roadmap announcement validating CPO in NVIDIA's rack-scale stack — supply-chain relevant for TW advanced packaging and AI server ODMs but not an immediate order or capex catalyst.

TrendForce projects combined EML and CW-DFB LD monthly capacity will more than triple to 50.7M units in 2026, driven by NVIDIA, Google, and Meta strategically locking in supply for 1.6 Tbps+ optical transport. Broadcom, Lumentum, and Sumitomo Electric lead with 55% combined share; Lumentum demoed 400 Gbps/lane EML at OFC 2026 targeting 3.2 Tbps, while Coherent is ramping 6-inch InP wafers for SiPh CPO.

Why it matters: Sector roadmap and capacity forecast for AI optical components — relevant supply-chain context but dominated by US/Japanese suppliers (Broadcom, Lumentum, Sumitomo, Coherent, Mitsubishi) with no named TW/KR beneficiary in the tracked universe.

Photonic computing firm Lightmatter is joining NVIDIA's NVLink Fusion ecosystem, providing CPO and NPO products compatible with NVIDIA's optical and SerDes technology to enable semi-custom XPUs to connect with NVIDIA switch silicon while cutting fiber and connector needs by 50%. The partnership validates CPO as a maturing technology for hyperscale AI clusters and broadens addressable market for photonic interconnect suppliers, with implications for the optical transceiver and CPO supply chain including TSMC's advanced packaging.

Why it matters: Ecosystem partnership announcement validating CPO roadmap for AI infrastructure; meaningful for advanced packaging supply chain but no direct capex or contract figures disclosed.

Chenbro (8210-TW) CEO Chen Ya-nan said at Computex 2026 that 70% of capacity at its under-construction US and Malaysia plants will be dedicated to server racks, with a new China rack line completing in June and additional rack facilities planned for Taichung and Vietnam. The chassis maker is extending from enclosures into IT, switch and CDU racks, expects H2 to be lifted by ASIC ramp at a large CSP customer, and unveiled a 1U MGX Vera Rubin chassis as an Nvidia partner.

Why it matters: Capacity allocation and rack expansion plans are meaningful for Chenbro's roadmap but the company is not in the tracked ticker universe, limiting direct PM relevance to a sector read-through on AI rack supply chain.

The TAIEX jumped 1.98% (+901.85 pts) to a record 46,459.16 as COMPUTEX-driven enthusiasm lifted heavy-electric and optical communications names, with TSMC (2330) closing at a record NT$2,425 (+1.89%). Jensen Huang's endorsement of Marvell and AI power demand drove limit-up moves in heavy electrics (1503, 1504, 1514, 1519, 7788) and optical comms peers, while ABF substrate maker Unimicron (3037) held above NT$1,000 and Nan Ya PCB (8046) advanced; Kinsus (3189) lagged. Memory names underperformed on rotation concerns.

Why it matters: Broad market wrap with sector rotation signals (heavy electrics, optical comms strength; memory weakness) and a fresh TSMC closing high, but no single stock-specific catalyst beyond the COMPUTEX/Huang commentary already in the tape.

Taiwan's Environment Ministry collected NT$4.97B (~US$155M) in its first carbon fee cycle from 461 plants, with semiconductors paying ~NT$2.2B (~US$69M, >40% of total) across 123 plants — TSMC accounting for 33. Power utilities paid NT$635M, steel ~NT$400M, cement NT$130M; the take beat the NT$4.5B estimate after 28 firms had self-reduction plans rejected or withdrawn and reverted to the NT$300/tonne base rate.

Why it matters: Sector cost-of-doing-business data point with named TSMC exposure (33 plants), but the NT$ amounts are immaterial to TSMC earnings and there's no new capex, contract, or guidance change.

Taiwan memory module maker Linghang sees the AI-driven DRAM/NAND shortage lasting through 2028, with Q3 contract prices guided +30% for DRAM and +70% for NAND as the big three prioritize HBM capacity. 2025 revenue hit NT$7.7B (+40% YoY, EPS NT$6.33) and Q1 2026 already reached NT$3.96B (EPS NT$10.76, +178% YoY); the company is raising ~NT$6B in July and flags SK Hynix as its main supplier, with CSP customers already negotiating 2028 contracts.

Why it matters: Explicit DRAM +30%/NAND +70% Q3 contract price guidance plus 2028 visibility and a named SK Hynix supplier relationship is a clear positive read-across for Korean memory makers.

Alphabet is raising $80B in equity to fund an estimated $187B of 2026 AI capex, intensifying shortages across HBM, DRAM, NAND, CPUs, servers and copper. Citi sees copper testing $15,000/ton while Micron has broken $1,000/share, with memory tightness expected to persist into 2027 — a clear tailwind for Korean memory makers and Taiwan's AI supply chain.

Why it matters: Quantified hyperscaler capex acceleration plus explicit HBM/DRAM/NAND shortage commentary through 2027 is directly stock-moving for Korean memory names and Taiwan AI supply chain.