The value of optical interconnect isn't saving power; it's unlocking the physical boundary of the pod. Astera Labs laid out the scale-up fabric bottleneck plainly at OCP APAC, and its candid Q&A reservation on CPO serviceability is where Taiwan's supply chain has a real opening.

AI infrastructure's next bottleneck isn't compute but connecting distributed compute with optics. With the three AICC models, NTT and OCP move DCI from a supporting role to center stage. Is this a tailwind for Taiwan's optical supply chain, or still far off? Three signals to watch.

Token demand wants 10x a year while supply delivers only 2x, and the compounding 5x gap can only be closed by architecture. Ayar Labs CEO Mark Wade's key line at OCP APAC wasn't about CPO but this: the unit of adoption shrinks from a whole rack to a single chassis.

Edgecore admitted 1.6T LPO is "not yet successful," so the stopgap is gone and CPO can no longer wait a generation. Three MSAs at OFC 2026 push CPO toward swappable engines, while SONiC only now has the models it needs. Hardware decides who ships; software decides who deploys at scale.

Optical circuit switching has stayed out of scale-up not because it fails to save power, but because it is two to three orders of magnitude slower than AI communication bursts. At OCP APAC, Linque showed a 64-channel silicon photonics OCS reconfiguring in 15 µs, clearing the “switch is too slow” hurdle. The other half—cutting transceiver lock time from milliseconds to nanoseconds—is the real battleground.

PCIe 7 keeps PAM4 but doubles insertion loss per inch, turning retimers from optional to mandatory. Astera Labs laid out the reach and pJ/bit trade-offs of AEC, ODSP and LPO, and noted that PCIe-SIG deliberately leaves the optical side undefined—where supply chain positioning happens.

Broadcom's CPO session skipped the theory and laid out field failure data from three product generations, because what holds CPO back was never power, but how to fix it when it breaks. With reliability now the selling point, the OCI spec and a 2027 fourth generation are opening three orders for Taiwan's supply chain.

CPO stalled for ten years not because the technology fell short, but because it wasn't yet a real problem. At OCP 2026 APAC, NVIDIA redefined CPO from "can it be built" to "can it be mass-produced," citing 10x MTBI and Spectrum-6 CPO now shipping in volume.

AMD's OCP keynote was framed around open standards, but the line worth underlining is the CPU-to-GPU ratio converging from 1:4 toward 1:1. Agentic workflows hand orchestration, retrieval and tool execution back to the CPU — and that matters far more to Taiwan's rack-integration and scale-up optical interconnect supply chain than a slide full of open acronyms.

The optics problem at 400G/lane is essentially solved; what remains is electrical, packaging and connectors. Coherent pinpointed the bottlenecks: wire bonds collapse above 110 GHz, OSFP connectors dip at 90 GHz, and PAM6 in optics hits an SNR wall. IM/DD can stretch to 800G, but it needs electronics, not photonics.

On the surface, Microsoft's keynote showcased in-house silicon; underneath, it publicly moved the AI infrastructure bottleneck from accelerators to power, cooling and networking. For Taiwan's supply chain, Microsoft's own "70% of components come from Taiwan" is an entry ticket, not a moat: the money is flowing to system-level co-design.

Google's fungible data center lets one facility host any generation of TPU, GPU, or CPU. As OCP standardizes power, cooling, and facility specs, Taiwan's thermal and power suppliers face a new spec-in battle, and the ticket in is OCP: a higher bar than capacity.

CPO's real bottleneck isn't bandwidth; it's whether a fiber can be detached and reattached. SENKO's elastic averaging gets repeatability under 0.15 dB and interchangeability under 0.5 dB, unlocking CPO serviceability. Module-level attach and customers remain open: technically validated, not yet in volume.

The market loves to frame scale-up as a copper-vs-optical death match; SemiAnalysis calls it a false choice. The real inflection is the HBM downgrade: the savings flow into scale-up networking, memory's share of TCO drops from 41% to 28%, and scale-up network cost triples. Copper holds inside the rack, optics bridges racks, and both ramp together.

In gigawatt-scale AI factories, the variable that decides what a GPU is worth has shifted from the GPU to the network. NVIDIA is turning "open" into a moat across four custom networks — opening Ethernet, the layer hardest to monopolize, while holding on to the most profitable layers: NVLink and CPO packaging. Taiwan's supply chain should track three milestones, not pretty numbers.

More than half of an AI training cluster's compute sits waiting on the network, and beyond 448G copper only reaches 25 cm. CPO's technical risk has been retired; the real fight is now over who writes the interface spec. 19 companies, a 300-page white paper, a spec due in Q4 2026 — Taiwan's supply chain is on the list, but in which seat?

The CPO vs. NPO vs. XPO debate was never about technology; it is about paying down supply-chain risk in installments. This six-way OCP 2026 APAC panel produced one hard deadline: 409.6T switches will push pluggables to the limit in about 18 months, and all four inflection points land on Taiwan's packaging and connector supply chain.