---
title: "CPO Mass Production Ignites AI Compute Infrastructure Revolution: Are Silicon Photonics and Interconnects the Biggest Winners?"
type: "News"
locale: "en"
url: "https://longbridge.com/en/news/283214955.md"
description: "Tianfeng Securities points out that CPO technology has officially entered the mass production phase. NVIDIA's first global CPO Ethernet switch is in full-scale production and integrated into the Vera Rubin platform, while Broadcom continues to iterate from the 51.2T Bailly to the 102.4T Davisson. Both giants have selected TSMC's COUPE silicon photonics 3D stacking solution, which can reduce system energy consumption to below 2 pJ/bit, improving energy efficiency by up to five times compared to traditional pluggable solutions"
datetime: "2026-04-18T07:47:52.000Z"
locales:
  - [zh-CN](https://longbridge.com/zh-CN/news/283214955.md)
  - [en](https://longbridge.com/en/news/283214955.md)
  - [zh-HK](https://longbridge.com/zh-HK/news/283214955.md)
---

# CPO Mass Production Ignites AI Compute Infrastructure Revolution: Are Silicon Photonics and Interconnects the Biggest Winners?

The demand for bandwidth and energy efficiency in AI data centers continues to climb, driving optical interconnection technology to a critical juncture.

An industry deep-dive report titled "In-depth Analysis of CPO Communication Devices," released by Tianfeng Securities on April 16, indicates that CPO (Co-Packaged Optics) switches have officially entered the mass production phase. Both NVIDIA and Broadcom have launched commercial products, with 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," requiring signals to travel a long detour; CPO, however, directly packages the optical engine next to the switch chip, shrinking the signal path from dozens of centimeters to just a few millimeters. Citing NVIDIA data referenced in the Tianfeng Securities report, this change reduces insertion loss significantly from 22dB to approximately 4dB, improves signal integrity by 63 times, boosts system optical power efficiency by up to 5 times, and enhances network resilience by 10 times.

## **NVIDIA Dual-Platform Launch: Spectrum-X First CPO Ethernet Switch Enters Full Mass Production**

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

**Regarding Spectrum-X**, its CPO switching chips have entered full mass production. As the world's first fully integrated 512-lane 200G-capable CPO Ethernet switch system, it has been integrated into the Spectrum-6 SPX network rack within the Vera Rubin platform, utilizing a 102.4 Tb/s switching chip.

Energy efficiency is the core selling point of this solution. The report cites data stating that traditional 1.6Tbps pluggable optical modules consume approximately 30W, with over half consumed by the DSP. NVIDIA's CPO solution achieves a maximum **5x improvement in optical power efficiency** and a **10x improvement in network resilience** at the system level by integrating silicon photonics devices directly inside the switch package and eliminating the DSP structure; simultaneously reducing the number of lasers by approximately four times, lowering operational costs and the risk of network interruptions.

**Regarding Quantum-X**, the liquid-cooled switch system Q3450 is equipped with four Quantum-X ASICs, totaling 144 MPO physical ports, with a total full-duplex bandwidth of **115.2 Tbps** and a single ASIC throughput capacity of 28.8 Tbps. Each ASIC integrates 18 optical engines, each capable of providing 1.6 Tbps bandwidth.

## **Broadcom: From Bailly to Davisson, Energy Efficiency Increases Over 3.5 Times**

**Broadcom was one of the first manufacturers to achieve support for CPO systems and, before NVIDIA joined, was the only one to release a production system equipped with CPO.**

Its Tomahawk 5 – Bailly (TH5-Bailly) is the **industry's first mass-produced CPO solution**, offering a total machine bandwidth of 51.2 Tbps, integrating 8 optical engines, each with a bandwidth of 6.4 Tbps, corresponding to 64 channels of 100 Gbps.

In terms of generational upgrades, Broadcom has launched the Tomahawk 6 – Davisson (TH6-Davisson) 102.4 Tbps CPO switching platform, with a single modulator rate reaching 200 Gbps, doubling that of Bailly. The report states that compared to traditional pluggable optical module solutions, this platform **reduces optical interconnection power consumption by approximately 70% and improves system energy efficiency by more than 3.5 times**.

It is worth noting that Broadcom had previously iterated multiple times using the Fan-Out Wafer-Level Packaging (FOWLP) solution developed by SPIL, but due to high parasitic capacitance, single-channel rates were difficult to expand beyond 100 Gbps. Consequently, Broadcom shifted to a packaging architecture based on TSMC's COUPE technology to further reduce signal conditioning requirements and minimize trace losses and reflections.

## **TSMC COUPE Silicon Photonics: The Technical Foundation Backed by Both Giants**

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

The report indicates that COUPE (Compact Universal Optical Engine) is a silicon photonics platform built by TSMC based on SoIC 3D hybrid bonding technology. Its core lies in vertically stacking at the wafer level to bond the Electronic Integrated Circuit (EIC) directly onto the Photonic Integrated Circuit (PIC).

A simple analogy can be made: In traditional solutions, the electrical signal connection between PIC and EIC is like shouting across a hallway, resulting in high energy loss and slow speed; whereas COUPE's 3D bonding is equivalent to two chips speaking directly face-to-face, with an extremely short path and minimal loss.

Citing TSMC data referenced in the report, compared to other bonding methods, SoIC hybrid bonding increases the PIC-EIC interface combination density by **at least 16 times**, reduces parasitic capacitance by approximately **85%**, and achieves approximately **40% energy reduction or up to 170% speed improvement** under the same power conditions; 3 dB bandwidth simulation data shows it can exceed 100 GHz.

From a system energy efficiency perspective, traditional copper interconnect systems typically consume over 30 pJ/bit, and traditional pluggable optical modules are also above 10 pJ/bit; whereas fully integrated optical engines based on COUPE can reduce unit energy consumption to **below 2 pJ/bit**, while reducing latency by over 95%.

The report also notes that during the mass production implementation phase, as CPO packaging evolves toward larger sizes and multi-optical engine integration, OSAT vendors with system-level advanced packaging capabilities may participate in downstream integration. Tianfeng Securities stated, "ASE has demonstrated the ability to integrate multiple optical engines and ASICs in packaging exceeding 75mm × 75mm, achieving <5 pJ/bit power consumption and increased bandwidth. Therefore, there is a possibility that CPO packaging under the COUPE path will be undertaken by vendors such as ASE."

## **High-Density Interconnect Devices: FAU Demand Doubles, MPO/MPC Break Through Space Limits**

CPO brings not only changes at the chip level but also an often overlooked challenge: how to route thousands of fibers.

**Inside a CPO switch, over 1,000 fibers need routing and management**, significantly increasing the importance of fiber coupling and high-density interconnect devices. Taking the NVIDIA X800-Q3450 as an example, the Tianfeng Securities report points out that over 1,000 fibers are led out from the optical engines, "creating a major organizational challenge." This directly drives demand for a batch of optical array devices.

The Tianfeng Securities report indicates that Fiber Array Units (FAU) are widely used to assist the crucial fiber coupling process in CPO, coupling light from silicon lenses into fibers with low insertion loss. Citing data referenced by Elecfan, in CPO application scenarios, **the number of FAUs required per CPO switch can be 3 to 5 times that of traditional solutions**. Broadcom has already integrated precision fiber arrays (FAU) provided by Corning into its Bailly CPO switch platform.

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

**MPO Connectors**: Capable of parallel transmission of multi-core fibers simultaneously, greatly increasing wiring density. SENKO's MPO EZ-Way connectors adopt a lower profile design; compared to traditional MPO connectors, they can increase the number of installable MPO connectors on optical module interfaces by **one fold**.

**MPC Connectors**: A fiber-to-chip direct connection solution designed specifically for CPO and high-speed data center applications, coupling light directly to the chip via a micro-mirror array. Connector height can be reduced to 0.6mm, minimizing insertion loss.

Additionally, the report mentions that OIF is advancing the consistency of optical interfaces in multi-plane interconnections and the modularization of external laser sources by developing an implementation protocol for 3.2T CPO modules, aiming to build an open CPO ecosystem.

## **External Laser Sources: Swappable Modules Solve Reliability Pain Points**

In traditional data center optical systems, lasers repeatedly undergoing thermal cycles are one of the main causes of failures and network interruptions. The CPO solution is to "invite the lasers out" and place them centrally in an independent temperature-controlled environment.

The Tianfeng Securities report introduces that NVIDIA Quantum-X Photonics is equipped with 18 External Laser Sources (ELS), with each ELS module containing 8 independent laser transmitters, totaling 144 for the entire machine. Each laser supports 4×200Gbps channels, with a total bandwidth of 800Gbps per MPO connector.

The key advantage is: **after centralized deployment of lasers, "the total number of lasers within the data center is reduced by four times" compared to traditional designs**, and each ELS module supports on-site replacement, "without affecting surrounding switch infrastructure." Citing NVIDIA data referenced in the Tianfeng Securities report, this increases network resilience by 10 times.

## **BOM Cost Breakdown: $70,000 Per Unit, Optical Engines and Splice Boxes Are Largest Cost Items**

How much is a CPO switch actually worth? The Tianfeng Securities report provides a detailed breakdown.

Taking the NVIDIA X800-Q3450 as an example:

-   **Optical Engine**: The cost for complete assembly of each unit (including fiber connection units) is approximately $1,000; the total material cost for optical engines alone reaches **$35,000 to $40,000** (for the 3.2T optical engine version);
    
-   **Splice Box**: Used to organize over 1,000 fibers, handling thousands of fibers in the X800-Q3450, the procurement cost for splice boxes will exceed **$3,000**;
    
-   **Total BOM Cost**: Including 2,000 meters of fiber and other miscellaneous components, the total is approximately **$70,640**;
    
-   **Estimated Terminal Selling Price**: Assuming a gross margin of 60%, the selling price is approximately $176,600. Adding $28,256 allocated for three years of service and warranty, the **total selling price including services is approximately $204,856**;
    
-   **Total Power Consumption**: 3,548 Watts.
    

The report notes that the above cost estimates are based on current production scales and may improve as output expands.

### Related Stocks

- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)
- [TSM.US](https://longbridge.com/en/quote/TSM.US.md)
- [AVGO.US](https://longbridge.com/en/quote/AVGO.US.md)
- [NVDU.US](https://longbridge.com/en/quote/NVDU.US.md)
- [AVGU.US](https://longbridge.com/en/quote/AVGU.US.md)
- [AVGX.US](https://longbridge.com/en/quote/AVGX.US.md)
- [XSD.US](https://longbridge.com/en/quote/XSD.US.md)
- [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md)
- [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md)
- [SOXX.US](https://longbridge.com/en/quote/SOXX.US.md)
- [AVGG.US](https://longbridge.com/en/quote/AVGG.US.md)
- [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md)
- [SMH.US](https://longbridge.com/en/quote/SMH.US.md)
- [PSI.US](https://longbridge.com/en/quote/PSI.US.md)
- [AVL.US](https://longbridge.com/en/quote/AVL.US.md)
- [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md)
- [601162.CN](https://longbridge.com/en/quote/601162.CN.md)
- [ASX.US](https://longbridge.com/en/quote/ASX.US.md)
- [GLW.US](https://longbridge.com/en/quote/GLW.US.md)
- [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md)

## Related News & Research

- [AI Chip Dark Horse Challenges Nvidia. Cerebras Refiles IPO Application Could Affect Nvidia’s Pricing Power?](https://longbridge.com/en/news/283209877.md)
- [5,137 Shares in NVIDIA Corporation $NVDA Bought by Chatterton & Associates Inc.](https://longbridge.com/en/news/283106425.md)
- [Wall Street Financial Group Inc. Acquires 2,961 Shares of NVIDIA Corporation $NVDA](https://longbridge.com/en/news/282537372.md)
- [Korean AI chip startup DEEPX, Hyundai work on robots powered by generative AI](https://longbridge.com/en/news/282774224.md)
- [Victory Giant joins IPO push by China’s circuit-board makers amid AI frenzy](https://longbridge.com/en/news/282545568.md)