On July 27, 2026, TrendForce reported that NVIDIA has shipped Spectrum-X Ethernet Photonics to some partners, and that Broadcom has also shipped small volumes of its 51.2Tbps Bailly. NVIDIA announced the product as "in production" at the end of May, and this survey reports that limited shipments have since begun. However, it remains unclear whether the shipment destinations are using the product for evaluation and verification or for commercial deployment. Broadcom, meanwhile, has been delivering CPO products to customers since 2024 and declared mass production the following year. Even so, market supply remains limited. CPO, which moves optical components right next to the switch ASIC, eases power and bandwidth density problems, but in doing so concentrates the challenges of optical engines, thermal management, and advanced packaging into a single product.
Between "In Production" and "Widely Available"
On May 31, 2026, alongside the mass-production launch of Vera Rubin, NVIDIA announced that Spectrum-X Ethernet Photonics, which uses 200Gbps SerDes, was also in production. It named CoreWeave, Lambda, and Oracle Cloud Infrastructure as early partners and adopters. The official product page indicates availability in the second half of 2026, and the shipments to some partners confirmed by TrendForce fall within that window. Shipment volumes, yield rates, and the start date of general availability have not been disclosed.
Broadcom's earlier trajectory illustrates well how broad the term "mass production" can be. The company delivered Bailly, which pairs Tomahawk 5 with an optical engine, to customers in March 2024, and in May 2025 described it as its "first mass-production CPO solution." Delta Electronics and Micas Networks have also announced production of switch systems. Yet what TrendForce observed as of July 2026 was still small-volume shipments.
Technical validation has advanced considerably. In testing at Meta disclosed by Broadcom in October 2025, cumulative operation of 400G-equivalent ports in a high-temperature lab environment reached one million port-device hours with zero link flaps (momentary outages). This is a characterization result under specific conditions, and does not guarantee full-scale deployment in production networks or fault-free operation in the future. Both companies' Ethernet CPO offerings have moved past the "can be built" stage, but have not yet reached the stage of "can be supplied in the quantities needed."
409.6Tbps Is Built from Four ASICs
TrendForce listed the processing capacity of Spectrum-X CPO as up to 400Tbps. According to NVIDIA's materials, the upper limit for the product line is 409.6Tbps. However, this figure does not represent the performance of a single massive switch ASIC.
| Product | Switch Configuration | Total Bandwidth | Port/Lane Configuration |
|---|---|---|---|
| NVIDIA SN6810 | Spectrum-6 ASIC×1 | 102.4Tbps | 800Gbps×128 |
| NVIDIA SN6800 | Spectrum-6 ASIC×4 | 409.6Tbps | 800Gbps×512 |
| Broadcom TH5-Bailly | Tomahawk 5×1 | 51.2Tbps | 400Gbps×128 |
| Broadcom TH6-Davisson | Tomahawk 6×1 | 102.4Tbps | 200Gbps/lane |
NVIDIA's 102.4Tbps-class package carries 32 silicon photonics optical engines rated at 3.2Tbps each, accommodating 512 electrical lanes running at 200Gbps. The SN6800 uses four ASICs of this class, bundled together into 409.6Tbps via a fiber shuffle within the chassis. Reading the rounded figure of 400Tbps as the capability of a single ASIC leads to a misunderstanding of the package scale and the number of components required.
Broadcom is also advancing generations. Bailly combines eight 6.4Tbps optical engines to achieve 51.2Tbps. The next-generation Davisson, announced in October 2025, doubles the number of same-bandwidth optical engines to 16, reaching 102.4Tbps with 200Gbps per lane. However, at the time of announcement, availability was limited to sample shipments for early-access customers and partners. The existence of a more advanced generation does not mean Bailly has already achieved widespread adoption.
The Cost of Moving Light Closer to the ASIC
In conventional pluggable optical transceivers, the high-speed electrical signals output by the switch ASIC pass through long on-board traces and connectors before being converted to light at a front-panel optical module. As frequency increases, signal degradation worsens, requiring compensation via DSP. In NVIDIA's comparison, electrical loss for a 200Gbps channel can reach up to 22dB, and power consumption per interface often reaches 30W.
CPO moves the optical-electrical conversion close to the ASIC package, shortening the electrical path. In the same comparison, NVIDIA states that loss can be reduced to about 4dB and power to a minimum of 9W, citing a 3.5x power efficiency improvement for Spectrum-X. Broadcom likewise claims that Bailly's optical interconnect power consumption is more than 70% lower than a standard pluggable configuration. Both figures are based on each company's own configurations, not a side-by-side performance comparison under identical test conditions.
The challenge that integration sacrifices replaceability is addressed by keeping the laser external. NVIDIA's optical engine places the modulator and photodetector inside the package, while separating the light source into a front-panel-replaceable External Laser Source (ELS). The single-ASIC version uses 16 ELS units, while the four-ASIC version uses 64. Broadcom's Bailly likewise supports a replaceable Remote Laser Module. Thus, the general description that "optical engines integrate even the laser" does not apply to all of today's major products.
The Triple Constraint of Yield, Heat, and Advanced Packaging
While CPO reduces the number of components, it does not make manufacturing the remaining components any easier. Optical ICs and electronic ICs must be bonded together, optical fibers must be aligned with high precision, and electrical and optical characteristics must then be inspected as an integrated whole. Furthermore, the tiny modulators must operate stably near the high-heat switch ASIC. A single defect can affect the yield of the entire expensive composite package.
Looking at NVIDIA's division of labor reveals where the supply constraints lie. TSMC stacks the optical IC and electronic IC using its COUPE process, while SPIL handles package assembly and testing. TFC Communication modularizes the laser dies and verifies their reliability, and Foxconn assembles the switch systems. TSMC has indicated plans to begin CPO production using the COUPE on substrate method in 2026. TrendForce views this process as competing with AI chips and HPC processors for 2.5D/3D packaging resources.
NVIDIA has also solidified its supply chain through contracts. In March 2026, it invested $2 billion each in Lumentum and Coherent—$4 billion in total—and signed multi-year purchase agreements with both companies. The agreements include rights to access future production capacity, though neither is exclusive. Rather than fully vertically integrating optical component manufacturing, NVIDIA has linked design and procurement to secure early access to the production capacity it needs.
The Path from 0.5% to 35%
Even as CPO enters mass production, pluggable optics will not disappear immediately. According to estimates presented by TrendForce in March 2026, CPO is expected to account for about 0.5% of AI data center optical communication modules in 2026, potentially reaching 35% by 2030. Broadcom predicts that copper wiring, which offers cost and low-power advantages, will remain the mainstream choice for short-distance in-rack connections through 2028.
Even within optics, paths diverge. NPO brings the optical engine closer to the ASIC on the same substrate while remaining a separate package, preserving replaceability and multi-vendor sourcing. TrendForce expects many cloud providers to prioritize NPO as a near-to-medium-term transition strategy. The denser CPO approach is better suited to configurations where switches and systems are supplied as an integrated package, as with NVIDIA.
According to the company's June forecast, the combined CPO and NPO market will exceed $39 billion by 2030, while pluggable optical modules will maintain roughly $26 billion. The period of ramped-up production for CPO switches is expected to be 2027–2028, while the period when optical adoption in scale-up connections accelerates the overall market is expected to be 2028–2029. What needs to be confirmed in the second half of 2026 is whether NVIDIA can expand shipments to some partners into general availability, and to what extent optical engine yield and COUPE packaging capacity can be demonstrated with concrete numbers. Only once that happens will CPO's "mass production" become a term that actually moves the market.