Skip to content
All analysis

Sector · Semiconductors · 23 May 2026

Co-packaged optics moves from demo to AI fabric constraint

As switch ASICs and GPU clusters push past copper limits, CPO promises lower energy per bit—while yield, lasers, thermal design, and serviceability decide real adoption.

Co-packaged optics moves from demo to AI fabric constraint

AI data-center fabrics are hitting a physics wall: electrical SerDes paths are becoming too lossy and too power-hungry as link rates climb and switch ASICs scale toward higher radix. Co-packaged optics (CPO) places optical engines next to—or inside—the same package as the ASIC or accelerator, shortening the electrical path and converting to light earlier.

Industry analyses frame CPO as strategic for scale-up networking later this decade, not as an overnight replacement for every pluggable transceiver. Pluggables remain the default where serviceability and interoperability matter most. Hyperscale AI clusters are where the power-per-bit math forces a harder look.

What CPO is meant to unlock

  • Energy per bit — Shorter electrical runs cut I/O power that otherwise crowds package and board budgets.
  • Bandwidth density — Optical engines support denser fabrics as copper reaches practical limits at 112G/224G-class signaling.
  • Package real estate — Fewer high-speed electrical pins can free bumps for power delivery and advanced packaging schemes.

What still blocks broad deployment

Manufacturing yield for photonic engines, laser reliability, thermal management beside hot ASICs, automated test, and field service models remain open issues. A failed optical engine inside a co-packaged assembly is harder to swap than a front-panel pluggable. Standardization of form factors is still maturing.

Foundry packaging platforms for silicon photonics are becoming part of the competitive map. That concentrates process choices and supply-chain commitments for companies building CPO products.

What to watch next

  1. Which AI networking generations ship CPO in volume versus which stay on pluggables.
  2. Yield, known-good-die flows, and OSAT qualification for heterogeneous photonic–electronic packages.
  3. Operator trust: mean time to repair and sparing models that match hyperscale operations.

CPO is no longer a research slogan. It is an interconnect roadmap decision for AI factories—gated by manufacturing and operations, not only by bandwidth slides.

More in this sector