NVIDIA explores Teflon for next-gen NVLink switch boards to reduce AI server signal loss
As AI chips push transmission speeds toward 200Gbps per lane, the physical limitations of circuit boards have become a critical bottleneck, where even minor signal interference can degrade system performance. To combat this, NVIDIA is reportedly exploring the use of polytetrafluoroethylene—better known by the brand name Teflon—to reduce signal loss in its next-generation NVLink switch boards.
This shift toward PTFE is driven by the material's ability to waste far less signal energy than the resins currently used in high-end printed circuit boards (PCBs). According to industry sources, NVIDIA is considering the material for use in Rubin Ultra switch boards or next-generation orthogonal mid-planes, which would allow high-speed switching to occur directly on the PCB rather than relying on complex copper cabling. The urgency is underscored by the fact that current boards are reaching 70-plus layers, a point where even top-grade M9 materials are reportedly falling short.
The practical application of this technology was highlighted at the OCP Global Summit on October 14, 2026, where NVIDIA displayed a switch-board sample made with PTFE. The presentation identified seven supply partners involved in the effort, including Flexium Interconnect, TTM Technologies, and Kinwong. However, moving from a sample to mass production presents significant engineering hurdles. Because of its non-stick properties, PTFE bonds poorly with other materials, making it difficult to create the multi-layered boards required for AI infrastructure. It is also softer and thinner than standard epoxy or hydrocarbon resins, which complicates drilling and plating processes.
Manufacturing and Adoption Hurdles
These technical difficulties have led to significant skepticism within the industry. Some reports indicate that overall yields for PTFE boards are currently below 5 percent, leaving the technology in the sample stage. An industry chairman has reportedly questioned the viability of the material, noting that it has not been used in circuit boards for 30 years and suggesting that NVIDIA's interest may be a reaction to shortages in high-end laminates. Despite these concerns, several U.S., Taiwanese, and Chinese vendors have already secured equipment capable of processing PTFE. While some analysts believe the processing bottleneck could be resolved within two years, a definitive outcome may not be clear until 2028 to 2030.
While NVIDIA works to solve the physical layer of signal loss, it is simultaneously evolving the architectural ecosystem surrounding its interconnects. At Computex 2025, CEO Jensen Huang introduced "NVLink Fusion," a program designed to open the previously closed NVLink technology to non-NVIDIA CPUs and GPUs. This move allows customers to build semi-custom AI infrastructure by combining NVIDIA GPUs with third-party processors and application-specific integrated circuits (ASICs), effectively positioning NVIDIA as the hub of the data center even when the system is not composed entirely of its own silicon.
This strategic opening of the ecosystem has already attracted a wide array of partners. Companies such as MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence are involved in the program, while customers like Qualcomm and Fujitsu are working to connect their own CPUs to NVIDIA GPUs. A significant milestone occurred at the Supercomputing '25 conference, where Arm announced it had joined the NVLink Fusion ecosystem. By integrating NVLink IP into its Neoverse CPUs, Arm enables its licensees—including hyperscalers like Microsoft, Meta, and Google—to develop custom processors that plug natively into NVIDIA's AI accelerator fabric rather than being limited to the slower PCIe interface.
Strategic Implications
The transition toward NVLink Fusion appears to be both a market expansion and a defensive maneuver. While NVIDIA dominates the GPU market, its data center networking revenue has recently lagged behind its semiconductor sales, partly due to the rise of open-source alternatives like UALink and the Ultra Ethernet Consortium. By allowing third-party CPUs and ASICs into the NVLink fold, NVIDIA reduces the incentive for cloud providers to switch to competing accelerators from AMD or Broadcom.
Ultimately, these developments suggest a two-pronged approach to maintaining dominance in the AI era. By pursuing PTFE to solve the physical limits of signal transmission and implementing NVLink Fusion to dissolve architectural barriers, NVIDIA is attempting to ensure that its technology remains the central nervous system of the AI factory. The success of this strategy depends on whether the company can overcome the steep manufacturing yields of PTFE and successfully integrate a diverse array of third-party silicon without compromising the performance of the NVLink fabric.
