SK Hynix CPO roadmap targets 100 Tb/s as Nvidia scales optics

SK Hynix and Nvidia are pushing co-packaged optics into the mainstream of AI data centers, targeting 100 Tb/s per-node bandwidth as copper interconnects hit a physical ceiling that threatens to stall compute scaling.

"Even if computing chips become more powerful, overall system performance cannot improve unless data movement between chips keeps pace," Kyusang Lee, a professor in the Department of Electrical and Computer Engineering at the University of Virginia, said.

SK Hynix published its CPO roadmap in Nature Electronics on Aug. 20 with researchers from MIT, the University of Illinois Urbana-Champaign, Nanyang Technological University and Yonsei University. The paper sets technical targets for future AI infrastructure: more than 100 Tb/s of bandwidth per node, energy consumption below 1 pJ/bit and chip-to-chip latency under 10 nanoseconds. It comes as compute throughput triples every two years while interconnect bandwidth grows only 1.4-fold over the same period, a gap the authors call the "bandwidth wall."

The stakes extend beyond SK Hynix. Nvidia's SpectrumX Ethernet Photonics switch entered mass production Aug. 14, the world's first 200G/lane CPO Ethernet switch, claiming 5x network power efficiency and 10x mean time between failures versus pluggable modules. SK Hynix shares surged 12.73% to 1.691 million won on the CPO paper and a 40 trillion won buyback plan.

CPO moves optics inside the package

CPO integrates optical transceivers into the same package as the processor, compressing high-speed electrical signal paths from centimeters to millimeters. Traditional pluggable modules — the industry standard for 800G and 1.6T links — still route electrical signals across a circuit board to a faceplate module, where retimers and equalizers burn power compensating for signal loss. At the scale of hundreds of thousands of GPUs, that overhead becomes a drag on the entire cluster.

The paper's longer-term vision extends optical interconnects to the memory interface itself. An "optics-centric" architecture would use a photonic interposer to connect an XPU pool directly to a shared memory pool, letting multiple AI accelerators draw on the same large-capacity memory rather than each carrying dedicated HBM stacks. SK Hynix, the dominant supplier of high-bandwidth memory to Nvidia, is using the roadmap to move beyond component supply toward system-level architecture.

Supply chain shifts as pluggable module makers watch

Nvidia's SpectrumX supplier list shows how the value chain is moving: TSMC handles silicon photonics fabrication, SPIL does wafer-level packaging, Lumentum supplies laser chips, Tianfu Communication assembles laser modules and Foxconn completes system integration. Notably absent are the pluggable module assemblers that dominate today's market, including Zhongji Innolight and Eoptolink.

That does not mean pluggable optics disappear quickly. TrendForce puts CPO penetration at roughly 0.5% of the AI data center optical interconnect market, and IDC expects mass commercialization only after 2027-2028. Cloud providers still plan pluggable modules as the workhorse for 2026-2027 builds, partly because a failed CPO optical engine can take down hundreds of ports at once, whereas a pluggable module swaps out in minutes.

Broadcom is also in the race, having shipped its Tomahawk 6 Davisson switch with 102.4 Tbps of switching capacity in October 2025, and its Bailly CPO switch integrates eight 6.4 Tbps silicon photonic engines with 70% lower interconnect power than pluggable alternatives.

The CPO push favors upstream silicon photonics, advanced packaging and laser chip makers while pressuring module assemblers to move up the stack. SK Hynix, trading at a premium on its HBM franchise, is using the roadmap to defend its position as AI memory demand shifts from per-chip stacks to pooled architectures. Nvidia's 5x power efficiency claim, if it holds in production, strengthens its grip on AI networking against Broadcom and Marvell. For module makers, the 2-3 year runway before CPO scales gives them time to pivot into optical engines and near-package optics, but the direction of travel is clear.

This article is for informational purposes only and does not constitute investment advice.