←News detail

Semiconductor Industry ObserverUpdated Nov 2 17:15

AI Chip Transmission Hits Bottleneck, Nvidia Evaluates PTFE Copper-Clad Laminate Solution

AI summary

Nvidia is evaluating the use of PTFE (Teflon) to replace traditional copper-clad laminate interlayer materials to reduce signal loss in high-speed AI chip transmission, with plans to use it in Rubin Ultra switch boards or next-generation orthogonal midplanes. Currently, the overall yield rate of PTFE boards in the industry is below 5%, and they can only be laminated by mixing with rigid materials. Dong Dingyu, chairman of Taiwan Union Technology, believes it is unlikely to become mainstream. Shengyi Technology is making the fastest progress, while Victory Giant Technology has already invested heavily in purchasing equipment. Analysts expect the processing bottleneck to be resolved within two years, but a clear conclusion on large-scale implementation will only come between 2028 and 2030.

Body · AI translation

AI chip transmission hits a bottleneck, Nvidia sets its sights on PTFE Source: Semiconductor Horizons Published: 2026-11-02 15:51 AI chip materials generate posters. Limited by the process, the current industry-wide comprehensive yield of PTFE boards is less than 5%.

AI chips are currently encountering a transmission bottleneck: single-channel rates have already reached 200Gbps, and even extremely tiny signal interference can affect performance. Sources say Nvidia has realized that better solutions are needed to ensure servers maintain optimal throughput, so it has set its sights on an unexpected material to solve a major thorny problem in the AI hardware field: PTFE, commonly known by its trade name Teflon, the coating material for non-stick pans.

PTFE has very low signal loss and can meet Nvidia's needs, but the material itself involves a complex process and is very difficult to implement. We can compare copper-clad laminate to a hamburger: the upper and lower layers of copper foil are like the buns, and the resin and glass fiber cloth in the middle are the filling. Ideally, if the entire middle layer were replaced with low-loss PTFE, transmission performance would improve significantly. But the problem is that before this "hamburger" is even put into use, its structure easily loosens and cracks.

Industry insiders say that materials commonly used in copper-clad laminates, such as epoxy resin, polyphenylene oxide (PPO), and hydrocarbon resin (HC), bond firmly after heating, and the board has high hardness after curing. Through high temperature and high pressure, they can be pressed tightly together, the walls of drilled holes are smooth, and electroplating is also smooth. By contrast, PTFE, used as a non-stick pan coating, has extremely low surface tension and is very difficult to bond with other materials, so it cannot be made into dozens of layers of AI server circuit boards on its own. At this stage, the industry can only adopt a compromise: mixing and pressing PTFE together with traditional rigid materials.

Dong Dingyu, chairman of Taiwan Union Technology, the leading copper-clad laminate maker currently occupying the high-end board market, revealed that when Nvidia came three years ago to inquire about PTFE board development, he raised a question: "Why has no one in thirty years ever used PTFE on circuit boards?" He said bluntly that Nvidia's initial search for this new material was only a temporary choice brought about by a shortage of high-end copper-clad laminates, and that PTFE is unlikely to become the mainstream solution.

After Nvidia exhibited its new-generation NVLink switch board at the OCP Summit in early October, industry rumors emerged that it is evaluating PTFE materials, and Teflon also appeared at this year's TPCA exhibition. The core purpose of using this non-stick pan material in AI hardware is to reduce signal loss. As mentioned earlier, the transmission rate of AI chips is already approaching the limit, while co-packaged optics (CPO) technology is not yet mature, so improving circuit boards has become the most direct and feasible solution at present.

According to industry sources, Nvidia plans to use it in Rubin Ultra switch boards or the next-generation orthogonal midplane backplane, allowing high-speed switching to be completed within the PCB board and no longer relying on complex copper cables. Some analysts suggest that introducing PTFE is not just a production process issue, but also an inevitable result of hardware design evolution. When PCB layer counts reach more than 70 layers, even top-grade M9 boards struggle to meet signal transmission requirements; combined with future demand for space computing power, faced with physical limits, Nvidia can only choose PTFE.

But a copper-clad laminate industry practitioner admitted that "time waits for no one." Nvidia's primary goal is to ensure that new products launch on schedule. Limited by the process, the current industry-wide comprehensive yield of PTFE boards is less than 5%; if process iteration cannot keep pace with the release rhythm of new-generation AI chips, Nvidia will never let an immature material drag down product progress. Therefore, even though the industry recognizes PTFE's performance advantages, most companies remain cautious.

Despite many doubts, most of the equipment capable of processing PTFE has already been acquired by several of Nvidia's partner suppliers. Among the PTFE partners selected by Nvidia, copper-clad laminate maker Shengyi Technology is making the fastest progress; Victory Giant Technology has invested heavily in purchasing equipment specifically to take on Nvidia orders; among Taiwan-funded companies, only flexible copper-clad laminate maker Taiflex and flexible board maker Flexium have entered, and neither is a mainstream player in the current AI circuit board track. Some analysts predict that the processing bottleneck is expected to be resolved within two years, but whether it can ultimately be implemented on a large scale will not have a clear conclusion until around 2028 to 2030.

This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our republication is solely for sharing and does not mean that we approve of or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Record-breaking! 2.75GHz gallium nitride micro-LED unveiled, targeting low-power data centers 2026-11-02 16:18

Broadcom becomes the big winner in Anthropic's IPO 2026-11-02 16:15

Backed by the blockbuster earnings reports of the two memory giants, South Korea unveils its largest-ever AI semiconductor budget 2026-11-02 16:11

Source

Semiconductor Industry Observer