top of page
Search

Non-Stick Opportunities

11 minutes ago
6 min read

Nvidia's prototype Teflon-based switch board for Rubin Ultra hands suppliers an exotic process, a short product life, and an outsized portion of the risk.


TL;DR

  • Nvidia is shifting unfunded development risk onto its PCB supply base by qualifying seven vendors for an immature, low-yield PTFE switch board with a small cumulative market and multiple architectural escape hatches toward M10/M9 laminates or all-optical interconnects.

  • PCB fabricators need to price themselves as expertise providers through NRE billing, co-investment, and yield-indexed pricing, though competitive pressure and buyer leverage will likely prevent them from enforcing those terms until co-packaged optics erodes demand for copper switch boards and forces industry restructuring.


A few weeks ago, at an overcrowded 2026 Open Compute Project (OCP) Summit, engineers and investment analysts swarmed around the Nvidia display set up across the hall from the expo. Over the past few years, the summit has evolved from a showcase for open hardware into one of the main venues for AI supply chain coordination. So, it wasn’t surprising that Nvidia was previewing the prototype NVLink switch motherboard slated for the redesigned Rubin Ultra rack. It was surprising that Nvidia also listed its printed circuit board (PCB) and laminate suppliers next to the prototype.

The PCB reportedly features a complex 16+16 hybrid construction incorporating up to eight polytetrafluoroethylene (PTFE) layers (Teflon is the commercial name for PTFE). The two halves are bonded using a transient liquid phase sintering paste that joins their copper pads. These are exotic PCB processes far from the mainstream products shipping today.

RCD Advisors estimates the total PTFE based PCB motherboard opportunity at roughly 55k m2 of top-side board area over three years, representing a cumulative market of about $1Bn. That demand calculation comes from two products, the Rubin Ultra switch board and an eventual Kyber-style backplane for a possible subsequent Feynman architecture.

PTFE is not a new PCB material. Its use in supercomputer backplanes dates back to the early 1990s, when Tera Computer relied on a massive Teflon backplane for high-bandwidth interconnects between compute nodes. The boards were manufactured by IBM at its facility in Endicott, New York. Subsequent supercomputers designed by Cray also incorporated the exotic material in its processor modules.

Those systems belonged to the era of vertically integrated "big iron," built in limited, custom volumes for a handful of national labs and research centers. Modern AI infrastructure is shipping in much higher volumes, demanding that merchant suppliers ramp specialized, complex processes quickly.

Scaling a high-layer-count PTFE board for commercial AI deployment introduces major manufacturing hurdles. While PTFE delivers better signal integrity at high frequencies, its physical properties are problematic for standard PCB fabrication processes. Because the material is naturally soft and chemically non-reactive, it smears rather than drills cleanly. Hole walls require aggressive plasma or sodium etching simply to allow copper plating to adhere. Furthermore, PTFE is dimensionally unstable and demands specialized high-temperature lamination presses, making layer-to-layer registration across hybrid boards difficult. With current yields reportedly sitting below 30%, the path to commercial volume production in 2027 remains narrow.

Supply chain checks confirm that the final Rubin Ultra PCB design is undecided. Nvidia has options if there is no path to PTFE yield improvement. The company could pivot to a next-generation M10 laminate using a glass-cloth-free hydrocarbon resin, likely a blend with benzocyclobutene (BCB), or fall back on an M9 laminate engineered with additional isolation layers. The decision will most likely depend on which design demonstrates a credible yield ramp.

The most drastic contingency is abandoning the copper-based scale-up switch altogether and accelerating the transition to an all-optical interconnect for Feynman. That possibility is not as far-fetched as it sounds. Rubin Ultra has already been scaled back significantly since its initial announcement two years ago, and it may no longer offer enough differentiation over its predecessor, Vera Rubin, to warrant going to production.

Nvidia posted the names of seven qualified suppliers, two laminate providers (Shengyi Technology and Taiflex) and five PCB fabricators (Victory Giant, TTM, Shennan Circuits, Kinwong, and Flexium). Qualifying five fabricators and two laminate suppliers gives Nvidia leverage. Nvidia is deliberately building a multi-source vendor base to prevent lock-in and ensure that no single circuit board bottleneck can hold up future rack deliveries.

But that leverage also creates an adversarial supply base. If Nvidia commits to PTFE, these fabricators will enter into a prisoner's dilemma, where each firm acting in its own self-interest produces a worse outcome for the overall health of the supply chain. Whether they like it or not, all of the listed suppliers were competing for allocation.

Over the past eight months, Nvidia has deployed capital across its supply chain to lock in capacity and technology access. The company has long-term development commitments with optical component makers, substrate providers, and testing houses. By that logic, the PCB suppliers listed on the display at OCP should stand next in line for capital support, since high-layer-count boards are an enabling technology for scale-up networking.

It is going to be hard to win that support with that many suppliers in the mix. Nvidia won't immediately see the need for upstream support unless all suppliers independently see the downstream substitution risks.

RCD Advisors' supply chain sovereignty analysis draws a distinction between sectors defined by scarce tacit knowledge and competitive sectors where suppliers hold little bargaining leverage. PTFE-based PCB fabrication sits halfway in between. The manufacturing process remains underdeveloped, yet several capable suppliers are already competing for the business. That structure gives a dominant buyer like Nvidia the leverage to extract technology development rents without funding them upfront. In effect, Nvidia gets its suppliers to pay for the development work that protects its roadmap, and leaves the risk with them.

PCB fabricators faced a similar situation ten years ago when they commercialized modified semi-additive process (mSAP) technology to meet the component density requirements of Apple’s iPhone. The difference was that the roadmap for smartphones continued to evolve with mSAP technology playing a critical, multi-year role, and the massive unit volume from smartphone production ensured that PCB fabricators could recoup their capital investments. By comparison, the AI opportunity is orders of magnitude smaller, with a technology roadmap that clearly moves toward optical interconnects.

The PTFE example highlights a consistent theme from RCD Advisors' research. Suppliers need to charge for their expertise as well as for materials, components, and processes. Doing so requires decision makers to understand their full capital structure, including the inherent knowledge portion of their assets. When evaluating those assets alongside the roadmap risk of transitional technologies, decision makers can properly gauge their exposure and set an appropriate return on invested capital.

Ultimately, this requires the discipline to do three things in the face of competitor pressure:

  • Bill explicitly for non-recurring engineering (NRE). If a buyer wants access to scarce tacit knowledge (such as hybrid PTFE lamination), that expertise should transacted separately.

  • Demand co-investment for dedicated equipment, or any specialized assets that cannot be easily repurposed if the buyer pivots to other solutions.

  • Establish yield-indexed pricing tiers rather than fixed price-down schedules to ensure that risk is spread evenly until the manufacturing process stabilizes.

Similar terms are standard practice in the defense electronics supply chain today, even when multiple suppliers are competing for business. To their credit, most PCB fabricators already use some of these terms, though inconsistent execution and intense buyer leverage routinely blunt their effect. It took the defense electronics supply base decades to enforce those commercial boundaries, only after multiple rounds of painful restructuring and consolidation. The AI PCB motherboard supply base is likely headed down that same road, and the beginning of that transition may arrive when co-packaged optics (CPO) reaches commercialization and the demand for advanced signal integrity in copper switch boards falls away.

There is irony in choosing the OCP summit to showcase the PTFE developments. OCP originated as an internal Facebook effort to cut data center costs and escape the design lock-in of OEM server makers. Facebook published its designs openly and called its servers “vanity-free,” an implied contrast with branded OEM boxes built around proprietary features. That mission changed over the years as efforts to push white-label designs deeper into the supply chain met resistance. Now Nvidia, whose NVLink-based racks rank among the most proprietary architectures in the data center, used the OCP venue to push its own version of “vanity-free” deeper into its supply chain.

If you find these posts insightful, subscribe above to receive them by email. If you would like to learn more about our consulting practice and how we assist organizations in the tech hardware supply chain, please get in touch with us at info@rcdadvisors.com.

bottom of page