Why AI's next performance leap may come from fiber, not silicon

For years, the AI race was a chip race. In 2026 it became an interconnect race, and the named mechanism is co-packaged optics (CPO): the practice of placing laser-driven fiber links directly on the chip package so data centers can stop paying the bandwidth and power cost of copper wiring between processors. (SK hynix newsroom)

The reason this matters now is that the constraint has moved. Transistor density and clock speed still improve every year, but the cost of moving data between chips has stopped keeping up. As model sizes and cluster sizes have grown, the energy and latency of ferrying bits across a server rack has become a large share of the bill, which is why the same week produces a peer-reviewed roadmap, a packaging milestone, and a standards deadline aimed at the same problem. A Nature Electronics paper published this year lays out the tradeoff: as electrical interconnects hit a bandwidth wall, optical links win on energy per bit at longer reaches, and co-packaged optics wins by moving the optical components as close to the silicon as the package allows. The paper maps a 2D, 2.5D, and 3D roadmap for the technology through the rest of the decade.

The industry has begun treating that roadmap as a public commitment. At the OCP APAC 2026 conference, the Open Compute Project set a Q4 deadline for a co-packaged optics standard aimed at the AI compute that gets wasted on slow links. TSMC, presenting at the same conference, laid out work on CoWoS material risks and cut its 3DIC development cycle to a year, a pace the company's packaging customers are now planning against. SK hynix put a CPO technology roadmap in Nature Electronics that recasts AI competition as moving from chips to systems.

The packaging side is moving too. Chinese OSAT JCET reported record first-half 2026 revenue with profit up 79% on AI demand, and its 1.5μm through-silicon-via milestone now supports denser high-bandwidth memory and 2.5D and 3D packaging, the substrate that any CPO module has to land on. The same DIGITIMES tag coverage this month notes that TSMC's share of Taiwan chip manufacturing revenue slipped to 81.8% as memory and packaging specialists picked up share, a small but visible reshuffle consistent with the industry shift toward interconnects. An Nvidia executive, interviewed by DIGITIMES at this week's Tech Forum in Taipei, discussed the company's progress on CPO production and pluggable optics.

That last point is the honest counterweight. CPO is still early. The standards work at OCP is not done. The packaging yields are not proven at hyperscaler volume. Pluggable-optics vendors are publicly pushing back on the CPO timeline, and the EDN state-of-CPO analysis frames the technology as a bet that still has to clear packaging and standards work. Samsung, separately, is holding off on High-NA EUV until 1nm production, a reminder that not every node-level bet pays on the schedule vendors prefer.

If CPO does land on the announced cadence, the value moves. Hyperscalers have spent the recent cycle buying GPUs as the scarce input. A working CPO stack means optical component makers, laser foundries, and advanced packaging houses sit on the same critical path, which is one reason JCET's 79% profit jump matters as a signal even if it is one quarter of one company's results. It is also why pluggable-optics incumbents are pushing back: their installed base is what CPO would replace.

The shape of the next eighteen months is therefore a set of public tests, not a single launch. Watch the OCP Q4 standard, because it sets the spec the data center buyers will write RFQs against. Watch the SK hynix roadmap milestones and whether the Nature paper's 2.5D and 3D stages show up in real product on the announced cadence. Watch TSMC's CoWoS material work, since any yield miss there is a yield miss for every CPO module that has to ride on it. Watch whether the Nvidia interview's production language turns into a ship date before the next OCP summit. And watch the pluggable-optics counter-narrative, because if hyperscalers keep buying pluggable modules at the rate they have been, the industry's CPO bet will look early in hindsight rather than overdue.

The category shift is real, and the public sources now support calling it that. What is not yet public is whether the packaging, the standards, and the yields line up on the same calendar.