NVIDIA is replacing superior PTFE with slightly inferior hydrocarbon resin CCL in its switch trays. The trade-off exchanges electrical performance margin for improved production line yield.
0.4dB/in is the transmission loss requirement for the NVL576 switch tray. On the same platform, signal rates must reach 224Gbps or higher. With these two figures side by side, the answer emerges: the path for materials is already narrow, too narrow to rely on single-metric selection.
According to supply chain reports from October 2, NVIDIA has begun testing a glass-free copper-clad laminate (CCL) based primarily on hydrocarbon resin (HC) to replace the previously tested glass-free PTFE-based CCL, while continuing to use glass-free HC-based prepreg (PP). The test target is the NVL576 switch tray of the Rubin Ultra platform—an eight-rack interconnect solution expected to enter mass production in the second half of 2027.
Conclusion first: This is not a material downgrade.
Preliminary tests indicate that glass-free HC CCL already meets the high-frequency electrical requirements of the switch tray. Its electrical performance is indeed inferior to the PTFE solution but significantly higher than M9-grade materials, with some metrics even exceeding the published specifications of M10. In other words, it has margin above the 'sufficient' threshold.
PTFE has long been considered the preferred system for high-speed, high-frequency boards due to its low dielectric constant and low dissipation factor, which minimize signal attenuation. The cost lies in processing: PTFE is soft and has poor dimensional stability, making drilling, lamination, and pattern transfer more difficult than with conventional materials, with particularly significant yield pressure in high-layer-count boards. HC resin also has very low molecular polarity, closely approaching the ultra-low loss range, but offers much better processing adaptability.
Putting these two factors together, the intent behind NVIDIA's current testing is clear: it is confirming whether the swap is worthwhile—sacrificing some electrical performance in exchange for higher PCB manufacturing yield and production efficiency.
224Gbps
Signal rate threshold
0.4dB/in
Loss requirement
2027H2
NVL576 mass production
These three figures correspond to the signal rate target, the per-inch transmission loss limit, and the mass production timeline for the NVL576 eight-rack interconnect solution on the Rubin Ultra platform. They constrain each other: the higher the rate, the harder it is to keep loss within 0.4dB/in for the same trace length.
The answer in the part numbers
Supply chain sources have identified specific part numbers. Shengyi Technology is currently one of the primary suppliers in the material evaluation for the NVL576 switch tray, involving three materials: SG5300N is a glass-fiber-free PTFE-based CCL, SG1030N is a glass-fiber-free HC-based CCL, and SIF09 is a glass-fiber-free HC-based prepreg. The intent of this testing round is clear: replace PTFE with HC in the CCL while keeping the prepreg part numbers unchanged.
A detail easily misread: although SG1030N and SIF09 are primarily HC-based, they still contain a small amount of PTFE. Therefore, this testing round should not be interpreted as a complete move away from PTFE. A more accurate reading is that the supply chain is adjusting the resin system and material composition ratios, using a formulation dominated by HC with PTFE as a supplement to balance high-frequency electrical performance with PCB processability.
Glass-fiber-free HC CCL meets the high-frequency electrical requirements of the switch tray. While its electrical performance is inferior to the original PTFE solution, it outperforms M9-grade solutions and even exceeds the established specifications of M10.
— Per supply chain material test feedback dated October 2, 2026; not an official NVIDIA decision
Following this line further, the next verification direction emerges. To improve manufacturability, the supply chain is evaluating replacing glass-fiber-free HC-based prepreg with glass-fiber-reinforced HC-based prepreg. This approach introduces new challenges: after adding glass fiber, whether the material can still meet signal integrity requirements and whether the loss can be kept within 0.4dB/in.
This is an unavoidable contradiction in high-speed PCB materials. Glass fiber improves mechanical properties and processing characteristics, but it also introduces dielectric constant non-uniformity, affecting high-speed signal transmission. The presence or absence of glass fiber has long transcended material selection; it is directly tied to signal integrity and production line processes.
Material grades are climbing
Beyond this specific project, the application scope of HC materials is expanding. In addition to the NVL576 switch tray, M9-grade materials are expected to enter mass production starting from Q4 2026, followed by a push toward M10-grade. As AI servers, switches, and high-speed network equipment continue to raise signal rates, the penetration of HC in high-end PCBs will continue to rise.
For upstream suppliers, this represents a shift in order structure. If the NVL576 ultimately adopts HC-based materials and enters mass production, Shengyi Technology is expected to increase procurement from Chinese HC resin suppliers. Current suppliers include existing players such as Dongcai Technology and Shengquan Group, with the possibility of introducing more potential suppliers in the future. Increasing the number of suppliers will alleviate supply pressure for high-end HC materials on one hand and optimize the cost structure on the other.
I am making a falsifiable claim here: as long as the fiberglass content remains an undetermined variable, the material roadmap for high-end switch trays will not converge on a single system before 2028. The rationale is that the starting point of this round of testing is manufacturability, and once fiberglass-reinforced solutions are validated, the material combination will be reshuffled again. This judgment becomes invalid if, by the time the NVL576 is finalized within 2027, a single resin system is explicitly locked in and a complete part number list is provided.
On a single server, how much is it worth if the yield of a switch tray improves by one percentage point, multiplied by a scale of tens of thousands of units? NVIDIA is calculating this account, and material manufacturers are calculating it too, but the answers they arrive at are not necessarily the same.