Nvidia unveils OOI CPO prototype for post-2028 AI needs

Nvidia's OOI move is not a packaging iteration—it is a re-architecture of where I/O physically resides in the silicon stack. Embedding optical engines directly on the interposer collapses the electrical-optical boundary that has governed datacenter design for over a decade. The ripple effects are structural. TSMC's CoWoS-L transitions from a wiring substrate to a photonic-electronic co-integration platform, eroding the pricing power of organic substrate houses like Ibiden and Shinko. EDA toolchains must now co-simulate photonics and electronics at the interposer level—a capability no major vendor has fully delivered. On compliance, 300mm silicon interposers sit in the grey zone of US export controls. If OOI volume production depends on specific hybrid-bonding processes, this pathway risks becoming a de facto US-exclusive technology, widening the capability gap between advanced packaging capacity in Taiwan, China and mainland China, and amplifying single-point supply-chain risk. Competitively, Broadcom's Tomahawk 5 CPO and AMD's UALink roadmap are the nearest threats. Yet Nvidia's vertical integration—GPU, NVLink, and CPO as one system—creates a moat that point-solution rivals cannot replicate. The 12-24 month tail: with 2028 as the volume target, 2027 is the packaging arms-race window. Expect TSMC to ship CPO-ready interposer SKUs by mid-2027, and optical vendors (Coherent, Lumentum) to pivot from module sales to wafer-level optical engine supply. The value chain shifts upward, irreversibly.