NVIDIA CPO Switches Enter Mass Production; Pluggable Optics to Coexist Long-Term
Taylor Wilson
Nvidia's photonic switches have begun shipping to partners and will scale in H2 this year — the first concrete timeline for CPO moving from concept to volume, shifting the optical supply chain's key question from 'whether to use light' to 'where to use it first.'
What problem do CPO switches actually solve?
The core logic is power and reliability. Integrating the silicon-photonics engine — a chip module that moves signals with light instead of electricity — inside the switch package means electrical signals no longer travel out to an external module for optical conversion.
This means → transmission paths shrink, power loss and signal degradation both drop, and the total component count falls.
In plain terms = signals used to "leave the building, loop around, and come back." Now they stay inside — shorter path, less power, fewer parts that can break.
CPO vs. pluggable optics — who replaces whom?
Nvidia networking SVP Gilad Shainer stated clearly: the two will coexist long-term; this is not a replacement story.
Pluggable optics will keep growing in four scenarios: long-haul transmission, front-end networking, brownfield upgrades, and segments where CPO lacks a cost advantage.
This means → the supply chain will not see a "CPO-takes-all" outcome. Pluggable-module makers are not facing extinction — their market is being re-bounded.
Why is copper still the first choice inside racks?
Shainer's words: "Copper is almost free in terms of power." Inside a rack, copper comes first; optics kick in only when distance forces the switch.
Optical integration lands first in the scale-out layer — the network tier connecting different racks — because that layer sees the steepest growth in port and transceiver counts.
This reflects a layered strategy, not a blanket shift to optics: copper up close, optics at distance, CPO in between.
What does NVLink Fusion's open ecosystem mean?
NVLink Fusion lets third-party accelerators, CPUs, and custom chips plug into the NVLink fabric — no longer limited to Nvidia's own GPUs.
This means → data-center infrastructure planning is decoupled from specific chip selection. A customer no longer gets locked into a single roadmap by a chip decision made two years ago.
In plain terms = building a data center used to mean "pick the engine, then build the car." Now Nvidia has standardized the chassis — the engine is swappable.
Why is "context memory" being framed as infrastructure?
Long-context models recompute key-value cache — temporary data the model uses to "remember" earlier conversation — on every turn, then discard it. Massive compute is wasted on repetition.
Shainer's view: this is fundamentally an infrastructure problem, not an intelligence problem. The model is smart enough; what is missing is a place to store and reuse "memory."
The fix: insert a pod-level shared memory tier between existing layers (HBM → system memory → local flash → network storage), accessible at speed by every GPU. BlueField, DOCA, and Dynamo will underpin this tier together.
What should the supply chain watch most closely?
Near-term anchor: the pace of CPO switch ramp-up in H2 — this sets order visibility for silicon-photonics engines, light sources, and advanced packaging.
Long-term signal: whether the context-memory tier becomes the next strategic infrastructure node after HBM, reshaping demand for storage and networking equipment.
This means → "certainty" in the optical supply chain is rising, but the real incremental opportunity may lie not in optical modules alone, but in the storage and interconnect rebuild around inference architecture.
Content is for reference only, not financial advice.