CPO mass production ignites a fundamental revolution in AI computing power—are silicon photonics and interconnects the biggest winners?

CPO mass production ignites a fundamental revolution in AI computing power—are silicon photonics and interconnects the biggest winners?

The demand for bandwidth and energy efficiency in AI data centers continues to rise, driving optical interconnect technology to a critical junction.

Tianfeng Securities released an industry in-depth report "CPO Communication Device Deep Analysis" on April 16, pointing out that CPO (Co-Packaged Optics) switches have officially entered the mass production stage, with both NVIDIA and Broadcom launching commercial products and TSMC’s COUPE silicon photonics platform becoming the underlying core architecture.

Why is CPO so important? Simply put, traditional optical modules are "plugged outside the switch," forcing the signal to take a long detour; CPO puts the optical engine directly next to the switch chip, shortening the signal path from tens of centimeters to a few millimeters. The Tianfeng Securities report, citing NVIDIA data, states that this change reduces insertion loss from 22dB to about 4dB, enhances signal integrity by 63 times, improves system optical power efficiency by up to 5 times, and increases network resilience by 10 times.

NVIDIA launches dual platforms, first Spectrum-X CPO Ethernet switch enters mass production

According to the Tianfeng Securities report, NVIDIA first released two CPO product lines, Quantum-X Photonics and Spectrum-X Photonics, at GTC 2025.

On the Spectrum-X side, its CPO switch chip has entered full-scale mass production. As the world’s first fully integrated 512 lane 200G-capable CPO Ethernet switch system, it has been integrated into the Vera Rubin platform’s Spectrum-6 SPX network rack, using a 102.4 Tb/s switch chip.

Energy efficiency is the core selling point of this solution. According to cited data, the traditional 1.6Tbps pluggable optical module power consumption is about 30W, with more than half consumed by DSP. NVIDIA’s CPO solution integrates silicon photonics devices into the switch package and eliminates DSP, achieving up to 5 times improvement in optical power efficiency and 10 times improvement in network resilience at the system level; reduces the number of lasers by about 4 times, lowering operating costs and network outage risk.

On the Quantum-X side, the liquid-cooled switch system Q3450 features four Quantum-X ASICs, a total of 144 MPO physical ports, with full-duplex bandwidth up to 115.2Tbps, and throughput per ASIC at 28.8Tbps. Each ASIC integrates 18 optical engines, each providing 1.6Tbps bandwidth.

Broadcom: From Bailly to Davisson, energy efficiency improved over 3.5 times

Broadcom was one of the early vendors to support CPO systems; before NVIDIA joined, it was the only company to release a CPO production system.

Its Tomahawk 5 – Bailly (TH5-Bailly) is the industry’s first mass-produced CPO solution, total bandwidth 51.2Tbps, integrates 8 optical engines, each with 6.4Tbps bandwidth, corresponding to 64 channels at 100Gbps each.

For generational upgrade, Broadcom launched Tomahawk 6 – Davisson (TH6-Davisson) 102.4Tbps CPO switch platform, with single modulator rate up to 200Gbps, doubling compared to Bailly. The report states that, compared to traditional pluggable optical module solutions, optical interconnect power consumption decreased by about 70%, and system energy efficiency increased by over 3.5 times.

It is noteworthy that Broadcom’s early iterations used fan-out wafer-level packaging (FOWLP) developed by SPIL several times, but due to high parasitic capacitance, single-channel rates could not scale beyond 100Gbps. Thus, Broadcom switched to packaging architecture based on TSMC COUPE technology to further reduce signal conditioning requirements and minimize trace loss and reflection.

TSMC COUPE silicon photonics: The technology foundation jointly chosen by the two giants

Why did NVIDIA and Broadcom both choose the same supplier’s solution?

The report points out that COUPE (Compact Universal Photonic Engine) is TSMC's silicon photonics platform built on SoIC 3D hybrid bonding technology, with the core innovation being wafer-level vertical stacking to directly bond electronic integrated circuits (EIC) to photonic integrated circuits (PIC).

A simple analogy: in traditional solutions, the electrical signal connection between PIC and EIC is like shouting across a corridor—high energy loss and slow speed; COUPE’s 3D bonding is like two chips directly talking face-to-face, with a much shorter path and lower loss.

Citing TSMC data, the report says SoIC hybrid bonding improves PIC-EIC interface density by at least 16 times compared to other bonding methods, lowers parasitic capacitance by about 85%, and achieves around 40% lower energy or up to 170% speed boost under same power consumption; 3 dB bandwidth simulations show values exceeding 100 GHz.

From a system energy efficiency perspective, traditional copper interconnect systems typically consume over 30 pJ/bit, traditional pluggable optical modules over 10 pJ/bit; with COUPE fully integrated optical engines, energy consumption per bit can drop to below 2 pJ/bit, with latency reduced by over 95%.

The report also notes that, at the mass production stage, as CPO packaging evolves to larger sizes and more integrated optical engines, OSAT vendors with advanced system-level packaging may join later-stage integration. Tianfeng Securities states: "ASE (Advanced Semiconductor Engineering) has demonstrated integrating multiple optical engines and ASICs in packages >75mm × 75mm, achieving <5 pJ/bit power and boosting bandwidth, so for COUPE-based CPO packaging, vendors like ASE may undertake the integration work."

High-density interconnect devices: FAU demand doubles, MPO/MPC break spatial limits

CPO brings not only changes at the chip level, but also an often overlooked challenge: routing thousands of optical fibers.

CPO switches have over 1000 optical fibers needing routing and management, making optical fiber coupling and high-density interconnect devices significantly more important. For example, Tianfeng Securities report notes that more than 1000 optical fibers pull out from the optical engine in NVIDIA X800-Q3450, "presenting a major organizational challenge." This directly drives demand for optical array devices.

Tianfeng Securities reports that fiber array units (FAU) are widely used to assist the critical fiber coupling process in CPO, coupling light from silicon lenses into fibers with low insertion loss. According to electronics enthusiast website data, in CPO applications, the number of FAUs needed per CPO switch can be 3 to 5 times higher than traditional solutions. Broadcom has integrated precision FAUs from Corning in its Bailly CPO switch platform.

At the connector level, the report introduces two key products:

MPO connectors: Enable parallel transmission of multi-fiber cables, greatly improving routing density. SENKO’s MPO EZ-Way connectors use lower profiles, doubling the number of MPO connectors that can be installed on optical module interfaces compared to traditional MPO.

MPC connectors: Designed for CPO and high-speed data center applications, these direct fiber-to-chip solutions use micro-mirror arrays to couple light into the chip, lowering connector height to 0.6mm and minimizing insertion loss.

Additionally, the report mentions OIF is formulating the 3.2T CPO module protocol to advance optical interface consistency and modularization of external laser sources in multi-plane interconnects, aiming to build an open CPO ecosystem.

External laser sources: Hot-swappable modules solve reliability pain points

In traditional data center optical systems, lasers repeatedly undergo thermal cycles, which is one of the main causes of failures and network interruptions. The CPO solution moves lasers "outside," centralizing them in independent temperature-controlled environments.

The Tianfeng Securities report states that NVIDIA’s Quantum-X Photonics is equipped with 18 external laser sources (ELS), each ELS module contains 8 individual laser emitters, totaling 144 in the system. Each laser supports 4×200Gbps channels, with total bandwidth per MPO connector at 800Gbps.

The key advantage is: After centralized deployment, the total number of lasers in the data center falls to one-quarter of traditional design, and each ELS module can be replaced on site, "without affecting surrounding switch infrastructure." Tianfeng Securities, citing NVIDIA data, says this improves network resilience by 10 times.

BOM cost breakdown: $70,000 per switch, optical engine and fiber organizing box are the largest expenses

How much is a CPO switch worth? Tianfeng Securities report offers a detailed breakdown.

Taking the NVIDIA X800-Q3450 as an example:

  • Optical engine: Each unit (including fiber connection unit) costs about $1000; total material cost for optical engines reaches $35,000–$40,000 (for the 3.2T optical engine version);
  • Fiber organizing box: Used for managing more than 1,000 fibers; for the X800-Q3450 handling thousands of fibers, procurement cost exceeds $3000;
  • Total BOM cost: Including 2,000 meters of optical fiber and miscellaneous components, total about $70,640;
  • Estimated end-user price: Assuming a 60% gross margin, price is about $176,600; including three years of service and warranty allocation of $28,256, total price including service is about $204,856;
  • Total power consumption: 3,548 watts.

The report notes that the cost estimate above is based on current production scale and may improve as volume increases.

Risk Disclosure and DisclaimerThe market is risky, and investment should be cautious. This article does not constitute individual investment advice, nor does it take into account specific users’ special investment objectives, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their particular circumstances. Investments based on this article are at users’ own risk.