Co‑packaged optics starts 2026

Co-packaged optics is moving from lab demos to factory output in 2026, as chip and network vendors push optical links closer to the switch silicon. (mitsui.com) In a standard data-center switch, electrical signals travel from the switching chip to a pluggable optical module at the front panel, burning power and losing signal along the way. Co-packaged optics puts the optical engine on the same package substrate, or right next to the chip, to shorten that path. (broadcom.com) Broadcom said on October 8, 2025 that it had started shipping Tomahawk 6 Davisson, a 102.4-terabit-per-second Ethernet switch with co-packaged optics. The company said the product doubles the bandwidth of any co-packaged optics switch it had previously offered. (broadcom.com) NVIDIA said on March 18, 2025 that its Spectrum-X Photonics and Quantum-X Photonics switches would use silicon photonics integrated into the switch and deliver 3.5x more power efficiency, 10x better resiliency, and 1.6-terabit-per-second ports. NVIDIA also said the systems were built with partners including TSMC, Coherent, Corning, Foxconn, Lumentum and SENKO. (nvidianews.nvidia.com) The pressure point is AI networking. Marvell said AI bandwidth demand is doubling every two to three years, while copper links get shorter as speeds rise, making it harder to scale large clusters with electrical connections alone. (marvell.com) Intel described the same bottleneck at OFC 2024, saying electrical input-output works well only over about 1 meter or less at these speeds, while pluggable optics adds reach but at power and cost levels that are hard to sustain for future AI clusters. (intel.com) That is shifting attention from just switch chips and optical modules to the packaging stack underneath them. NVIDIA said its co-packaged optics platform depends on foundry, laser, fiber-to-chip and packaging partners, and Broadcom says the technology combines silicon photonics, switch silicon, packaging and test on one substrate. (developer.nvidia.com (broadcom.com)) Analysts and industry researchers are not all using the same timeline. Mitsui’s Global Strategic Studies Institute wrote on April 1, 2026 that 2026 marks “full-scale mass production and practical roll-out,” while Cignal AI said in February 2025 that small amounts may appear in 2026 but large-scale deployment is more likely in 2027 or 2028 or later. (mitsui.com) (cignal.ai) Broadcom said at OFC 2026 that it was also working with partners on open optical standards for AI “scale-up” links beyond copper, a sign that vendors are trying to turn early products into a broader supply chain. (broadcom.com) The near-term question is no longer whether co-packaged optics can be built. It is how fast packaging, testing and fiber-attach capacity can ramp as AI clusters move from thousands of accelerators toward much larger fabrics. (developer.nvidia.com) (marvell.com)