UBS Semiconductor Forum: CPO trends are "particularly clear" this year, Broadcom states that "power consumption will be fully optimized in 28 years, replacing copper connections."

UBS recently discovered that in the context of soaring power consumption in AI data centers and increasingly apparent bottlenecks in copper cable transmission, Co-Packaged Optics (CPO) is becoming a key path for the next stage of AI infrastructure upgrades after attending a semiconductor forum. On September 13, according to news from the Chasing Wind Trading Desk, UBS stated in its latest research report that the trend of Co-Packaged Optics (CPO) technology has become "particularly clear" this year, with industry experts predicting that power consumption optimization will reach a critical point by 2028. The report pointed out that the speeches of three experts from Broadcom, Ayar Labs, and Lumentum at the semiconductor forum indicated that CPO is gradually breaking through power consumption and integration bottlenecks, and is expected to enable large-scale applications of AI server interconnects in 2028-2029. Broadcom stated that the key obstacle preventing CPO from fully replacing copper cables is that its energy consumption is still above 10 picojoules per bit (pJ/bit), but it is expected that mid-2028 solutions will drop below 10 pJ/bit, and advanced solutions in 2029 could be as low as 5 pJ/bit. At that time, CPO will truly have the capability for large-scale deployment. Ayar Labs emphasized that the sharp rise in power consumption in AI data centers is driving the demand for CPO, planning to limit the rack power consumption of a 576 GPU system to within 100kW through optical I/O solutions. Lumentum stated that, leveraging its technological advantages in compound semiconductor materials and ultra-high power lasers, it is paving the way for the next generation of 3.2T optical interfaces. Broadcom: CPO energy consumption will drop below 5 pJ/bit by 2029, copper cables are at a critical point Broadcom's Chief Technology Expert Tzu Hao Chow stated that the "horizontal scaling" of AI server networks has widely adopted pluggable optical modules and is gradually transitioning to CPO. However, "vertical scaling" still relies on copper cables but faces significant physical limits—when the single-channel rate increases from 100G to 200G, the effective transmission distance of copper cables shrinks from 4 meters to 2 meters, and even with the addition of relay chips, the improvement is very limited. Chow pointed out that the key obstacle preventing CPO from fully replacing copper cables is that its energy consumption is still above 10 picojoules per bit (pJ/bit), but it is expected that mid-2028 solutions will drop below 10, and advanced solutions in 2029 could be as low as 5 pJ/bit. At that time, CPO will truly have the capability for large-scale deployment. He also compared two optical module technologies developed by Broadcom: Silicon Photonics (SiPh) and VCSEL. SiPh has advantages in bandwidth, transmission distance (up to 2 kilometers), and power consumption, making it suitable for data center interconnects; while VCSEL still holds a place in terms of cost and short-distance transmission (within 30 meters). Ayar Labs: AI data center power consumption is skyrocketing, CPO is the only path Ayar Labs founder Mark Wade pointed out the pain point: “Meta's new Hyperion data center will consume 2GW of electricity by 2030, and will rise to 5GW by 2035, equivalent to 1.5 times the electricity consumption of all residents in Louisiana.” He stated that the exponential growth of AI training tasks has made the power consumption of data center racks a bottleneck for expansion. The CPO solution from Ayar Labs, including TeraPHY optical I/O chips and SuperNova multi-wavelength light sources, will enable: Power consumption of a 576 GPU system in a single rack to be controlled within 100kW (2027) Even when expanded to 2048 GPUs, the power consumption of a single rack can remain stable below 100kW (2029) In addition, Wade believes that AI hardware is transitioning from the "narrative" phase to the "plumbing" phase, meaning the infrastructure of underlying interconnects and power systems. He emphasized that optical I/O technology will enhance both AI computing efficiency (tokens/sec) and cost ratio (throughput/$), truly achieving a win-win in performance and profitability. Currently, Ayar has received investments from AMD, NVIDIA, Intel, and TSMC, and is collaborating with Alchip to promote optical integration on the TSMC COUPE platform. Lumentum: VCSEL Technology Matures, Paving the Way for 3.2T Optical Interfaces At a closer level to the device, Lumentum's Chief Photonics Expert Matt Sysak emphasized that the company has a deep accumulation of compound semiconductor materials from Gallium Arsenide (GaAs) to Indium Phosphide (InP), and can provide core devices such as MEMS optical switches and high-speed lasers. NVIDIA is using Lumentum's ultra-high power lasers as the optical engine for its Spectrum-X CPO switch, marking the full introduction of optical technology into network infrastructure by traditional GPU giants. Lumentum's latest PAM4 modulation lasers (400G/448G/450G) and distributed feedback modulators (DFB-MZI) are laying the foundation for the next generation of 3.2T optical interfaces. At the same time, its VCSEL lasers, which have shipped over 100 billion units in FaceID scenarios, also demonstrate high reliability and low cost advantages in server clusters, making them suitable for short-distance large-scale interconnects