Nvidia targets 15x optics capacity expansion by 2027 to meet CPO demand - SDxCentral

This is not a procurement decision—it is Nvidia cementing vertical monopoly at the interconnect layer. CPO pulls optics from the network tier into the compute die. A 15x capacity ramp by 2027 means silicon-photonics wafers, EML lasers, and micro-lens arrays face a supply shock comparable to the HBM3 lock-in on SK Hynix in 2019. Whoever controls optical-engine packaging controls the bandwidth ceiling of every AI cluster. CoWoS-L and silicon-photonics foundries will face order congestion. Supply-chain risk is acute: roughly 60% of transceiver capacity sits in mainland China. Scaling 15x via outsourcing pins AI interconnect's throat to a single jurisdiction. The 2023 EDA restrictions proved any chokepoint can be weaponized overnight. Nvidia will likely pull silicon-photonics design back to US-based foundries, but the yield and cost penalty will hit P&L hard in 2026-27. Competitive read: Broadcom's Bailly and Marvell's Innocuity have staked CPO switch positions. Nvidia's implicit message is "I don't need your switch"—in-house optical engines plus NVLink direct-connect structurally ejects Broadcom from the AI networking layer. AMD's MI400, if it follows, is forced onto Coherent or Lumentum, creating asymmetric dependency. 12-to-24-month outlook: CPO yield (~70% today) is the single gating variable. Cross 90% and pluggable transceiver TAM shrinks roughly 40% by 2028. The real tail effect: module vendors morph from component suppliers into system-level players, and their valuation anchor shifts from manufacturing P/E to platform P/S.