A technical deep dive into NVIDIA Rubin Ultra’s orthogonal backplane material selection
A material born nearly 80 years ago and known to every household — PTFE (polytetrafluoroethylene), the “King of Plastics” — is entering its AI golden age. From non-stick cookware coatings and chemical-resistant piping to a leading candidate material for NVIDIA’s next-generation server orthogonal backplanes, a cross-industry paradigm shift is underway.
The information landscape around this topic is dense and multi-layered: what NVIDIA has officially confirmed, what sell-side research has uncovered through supply chain checks, what vendors have publicly disclosed, and what remains speculative. Without drawing these boundaries clearly, it’s easy to mistake a direction for a done deal. This article approaches the topic from a technical angle, systematically covering architecture, materials, processes, and supply chain — with every key data point tagged by evidence level.
Evidence Level Key: [Official] = NVIDIA official / IEC standard / vendor datasheet; [Supply Chain] = brokerage research / supply chain surveys; [Vendor] = official website / marketing claims.
Table of Contents
1. Current Status & Information Boundaries: What’s Officially Confirmed
2. Three Key Concept Clarifications
3. Why PTFE: The First Principles of “Dual-Low”
4. Four Real Challenges of Glass-Free Processing
5. Supply Chain Landscape (Evidence-Tagged)
6. Market Size & Open Questions
7. Conclusion & Risk Notes
1. Current Status & Information Boundaries: What’s Officially Confirmed
Let’s start with what NVIDIA has officially confirmed. As of August 2026, three things have clear official anchors:
Officially Confirmed:
1. Rubin Ultra NVL576: 576 GPUs in a unified NVLink domain (8 Kyber racks × 72 GPUs), shipping in H2 2027 [Official]
2. NVLink 7: SerDes rate of 224G PAM4 [Official]
3. Kyber high-density rack form factor and NVLink switch tray structure [Official]
The following widely circulated claims, however, all belong to supply chain research and have never been disclosed by NVIDIA:
Claim: “NVIDIA has confirmed PTFE as the primary material for the orthogonal backplane”
Status: GF Securities research note, not officially confirmed; August 2026 ODM surveys indicate the design is not yet finalized
Claim: “The orthogonal backplane reaches 78 or even 108 layers”
Status: The 78-layer figure originates from a SemiAnalysis estimate; 108 layers is further supply chain extrapolation
Claim: “Shengyi Technology as primary supplier, Taiflex as secondary — the格局 is set”
Status: Taiflex does have PTFE CCL products on its website [Vendor], but the supplier hierarchy is brokerage speculation
Claim: “Zhonghao Chenguang passed NVIDIA’s full M10 certification”
Status: From research reports; not confirmed by company announcements; “M10” is itself not an official grade
Claim: “2027 market size of ¥8B, resin demand 4,000–5,000 tons”
Status: GF Securities calculation (assumptions: ¥150K/ton, ~800g resin/sheet, ~¥2,500/sheet) — a conditional estimate, not an order guide
Two timeline variables deserve attention: First, industry media reported that in July 2026, Jensen Huang indicated the original Kyber design would be replaced by “a better solution” — if true, the backplane topology and material selection would need reevaluation. Second, the orthogonal backplane final design — originally expected to lock in Q3 2026 — showed signs of delay in August ODM surveys. Three parallel routes remain in play: pure PTFE, hybrid (PTFE + hydrocarbon), and pure hydrocarbon.
Bottom line: “PTFE entering the orthogonal backplane” is a high-confidence industry direction (multiple independent supply chain sources converge), but “confirmed adoption, design locked” is a low-confidence timing claim. Direction determines whether to invest; timing determines when.
2. Three Key Concept Clarifications
Concept 1: What Are M4/M6/M8/M10, Really?
The “M” series originated as Panasonic’s Megtron product line naming convention. Because Megtron long dominated high-speed CCL, the industry gradually adopted “M6/M7/M8/M9” as a benchmark for low-loss tiers — similar to how “3nm/5nm” evolved from process nodes to marketing labels. Two key points: ① “M8-grade” or “M9-grade” claims from different manufacturers are not strictly comparable (test frequency, glass fiber system, and resin content all affect values); ② “M9-Q” and “M10” are not official Panasonic product names. As of August 2026, Panasonic has no published M10 datasheet. “M10” is an industry consensus term for “next-generation materials for the 224G era.” When you see “M10 certified,” read it as “a vendor completed an internal validation per industry convention.”
M-Series Loss Tiers (industry benchmark, approximate):
• M4: Df ~0.005 — Server motherboards, 28G era
• M6: Df ~0.002 — 112G SerDes mainstream
• M7/M8: Df 0.0008–0.0015 — 224G switch chip boards
• M9 (low-loss hydrocarbon): Df 0.0006–0.001 — 224G high-speed layers, hybrid builds
• “M10” (unofficial grade): Df <0.0005 — Orthogonal backplane candidate: only PTFE-type materials can achieve “dual-low”
Note: Values are typical ranges from public industry reports; actual values vary by manufacturer and test frequency. For magnitude reference only.
Concept 2: PTFE’s Melting Point of 327°C ≠ Tg
327°C is PTFE’s crystalline melting point (Tm) — that’s correct. But PTFE is a semi-crystalline polymer with no well-defined glass transition temperature (Tg) (literature values range from -80°C to 130°C; ~130°C is commonly used in engineering). This has direct engineering implications: PTFE enters a rubbery state and experiences significant modulus drop well below 327°C (above ~130°C), producing cold flow / creep — the real risk source for PTH barrel reliability and long-term dimensional stability in high-layer-count backplanes. Additionally, PTFE’s sintering temperature (>360°C) is incompatible with PCB lamination lines (typically <200°C), so glass-free PTFE CCL must use a surface-coated hydrocarbon resin + low-temperature lamination route, not melt-processing like conventional thermoplastic resins. “Operating temperature -190°C to 260°C” is a valid chemical stability specification, but it does not mean mechanical dimensional stability up to 260°C.
Concept 3: 224G PAM4 and the Correct Signal Frequency Calculation
The commonly seen “337Gbps high-speed signal” conflates link bandwidth with per-lane SerDes rate. The correct chain: NVLink 7 SerDes is 224Gbps PAM4 (4-level modulation) → 2 bits per symbol → symbol rate 112 GBaud → Nyquist frequency ~56GHz (with FEC and equalization overhead, the frequency band of interest extends to 70GHz+ — which is the engineering basis for Taiflex’s PTFE CCL being marketed for “70GHz+ mmWave applications”). The material selection logic for high-speed digital is “select by sinusoidal loss parameters at the Nyquist frequency band,” not “select by the Gbps number.”
3. Why PTFE: The First Principles of “Dual-Low”
Dielectric loss approximation: αd ∝ f·√Dk·tanδ. At 56–70GHz with backplane trace lengths of 0.5–1 meter, the loss budget is fought inch by inch. But Dk is even more frequently underestimated than Df: reducing Dk from 3.3 to 2.2 means signal propagation is ~22% faster, dielectric thickness can be greater at the same impedance (allowing wider traces), and copper foil roughness-related loss proportion drops. PTFE is the only mature commercialized resin system that simultaneously achieves Dk ≈ 2.1–2.2 and Df < 0.001 — this is the first-principles reason it ranks at the top of the orthogonal backplane candidate list.
PTFE’s fundamental properties (“three extremes, one low”) — well-established material science:
• Extreme temperature resistance: Continuous use -190 to 260°C; crystalline melting point 327°C (note: not Tg; no well-defined Tg)
• Extreme chemical inertness: Virtually insoluble in any solvent; resistant to aqua regia, hydrofluoric acid (except molten alkali metals)
• Ultra-low friction / non-stick: Dynamic friction coefficient ≈ 0.04, extremely low surface energy (the principle behind non-stick pans)
• Low dielectric constant / loss: Dk ≈ 2.1, Df 0.0002–0.0009 (10GHz), nearly invariant with temperature and frequency
Cross-Resin Comparison (typical public ranges):
PTFE (with SiO₂ filler) — Dk: 2.1–2.3 | Df: 0.0002–0.0009 | Key weakness: poor copper adhesion, cold flow, difficult drilling/plating
Low-loss hydrocarbon (M9-tier) — Dk: 3.2–3.4 | Df: 0.0006–0.001 | Dk ~50% higher
PPE/PPO modified — Dk: 3.2–3.6 | Df: 0.001–0.002 | Glass cloth raises overall Dk
LCP — Dk: 2.9–3.0 | Df: 0.002–0.004 | Difficult to make rigid boards, high cost
Epoxy FR-4 — Dk: 4.2–4.6 | Df: 0.015–0.02 | Unusable at 56GHz+
Representative Product Benchmarking (verifiable parameters):
Rogers RT/duroid 5880 — Dk: 2.20±0.02 | Df: 0.0009 @10GHz | Official datasheet (thermal conductivity only ~0.20 W/m·K)
Taiflex PTFE CCL (publicly listed on website) — Dk: not disclosed | Df: marketed as “ultra-low” | Vendor website (glass-free, targeting 70GHz+)
Shengyi Technology PTFE CCL — Parameters per vendor materials | IEC 61249-3-6:2026 standard lead author [Official]
A common question: Rogers has made PTFE laminates for decades — why haven’t they captured this AI wave? Because RT/duroid 5880 is an RF/microwave laminate (small format, for antennas and mmWave circuits), while AI backplanes require large-format, ultra-thick, ultra-high-layer-count, hybridable, mass-producible PCB-grade CCL — an entirely different product form and process capability. This is exactly the opportunity for CCL manufacturers like Shengyi and Taiflex.
4. Four Real Challenges of Glass-Free Processing
“Glass-free” (removing the glass fiber cloth) is the technical core of this generation of PTFE CCL. Glass cloth has a Dk of ~4.5–6, the primary drag on laminate Dk. Removing it and compensating with spherical SiO₂ filler (40–60% loading) for stiffness and CTE is what enables Dk ~2.2. But the cost is four process challenges:
1. Copper foil adhesion: PTFE’s extremely low surface energy means direct lamination peel strength is inadequate. A hydrocarbon resin bonding layer must be applied — IEC 61249-3-6:2026 specifically covers “non-reinforced, filled” systems, confirming this is a new process category [Official]
2. Stiffness compensation: Removing glass cloth drastically reduces modulus. Rigidity and drilling quality depend entirely on the particle size distribution, surface treatment, and interfacial bonding of spherical silica flour
3. Mechanical processing: PTFE is soft yet tough. Drilling produces smear, and drill bit wear is severe (supply chain data: bit life drops 1–2 orders of magnitude vs. FR-4). Deep hole aspect ratios can reach 50–100:1
4. Hybrid lamination thermal matching: PTFE high-speed layers laminated with hydrocarbon / low-loss glass fiber layers — the two systems have different CTEs and lamination temperature windows. Warpage and registration accuracy are the primary yield killers. The 40–50% pilot yield [supply chain] is mainly bottlenecked in these areas
5. Supply Chain Landscape (Evidence-Tagged)
Electronic-grade PTFE Resin — Daikin, Chemours, Zhonghao Chenguang, Dongyue
Domestic PTFE production capacity is real (verifiable in annual reports). Claims of “5N high-purity, Df < 0.0005, certified” originate from research reports; company announcements have not directly confirmed these.
PTFE CCL — Shengyi Technology
IEC 61249-3-6:2026 standard lead author [Official]. “Primary supplier expectation” is brokerage judgment [Supply Chain].
PTFE CCL — Taiflex
PTFE CCL products publicly listed on official website [Vendor]. “Strategic partnership with Taiwan Union, certification sprint for secondary supplier” are market rumors with no official materials.
Spherical Silica Flour — Lianrui New Materials et al.
Spherical silica use in low-loss CCL is a public product fact. PTFE board supply share is from research reports.
PCB Manufacturing — Shennan Circuits, Wus Printed Circuit, et al.
High-layer-count backplane capacity buildout is public information. “Received terminal capacity assurance demands” is supply chain sourcing.
Copper Foil — Mitsui et al.
“Primarily uses imported Mitsui copper foil” is from research reports. The importance of low-profile HVLP copper foil is industry consensus.
6. Market Size & Open Questions
GF Securities estimates: 2027 Kyber platform PTFE CCL market of approximately ¥8 billion, PTFE resin demand of 4,000–5,000 tons (implied assumptions: electronic-grade resin ~¥150K/ton, ~800g resin per sheet, ~¥2,500 per sheet). The calculation is internally consistent, but three premises are all uncertain: ① pure PTFE route wins; ② Rubin Ultra ships on schedule in H2 2027; ③ Kyber design is not “replaced by a better solution.” Any premise change would significantly revise the estimate — treat it as a directional magnitude, not an order guide.
Open Questions as of August 19, 2026:
1. Orthogonal backplane final design (pure PTFE / hybrid / hydrocarbon alternative) lock-in timeline has slipped
2. Pilot yield of 40–50% [supply chain] remains far from mass production threshold
3. All “passed NVIDIA certification” claims lack official document anchors
4. M9-grade hydrocarbon is sufficient for most non-extreme-speed layers; “all-PTFE” is not inevitable
5. PTFE cold flow-induced barrel stress and high-temperature/high-humidity Dk/Df drift lack publicly available validation data for 10-year server lifecycle scenarios — this may be a more insidious risk than yield
7. Conclusion
With the evidence laid out, the structure of this PTFE narrative is clear: The direction is certain — 224G PAM4 has pushed high-speed laminate loss requirements into a range only fluoropolymer materials can cover; the IEC standard publication and Taiflex’s public product listing show the industry is making real investments. The timing is uncertain — design not locked, yields not at target, certifications unanchored. The numbers need to be discounted — ¥8B is a conditional probability estimate.
For materials professionals, the story isn’t just about PTFE itself — spherical SiO₂ filler, low-profile copper foil, and hydrocarbon hybrid bonding layers are the “pick-and-shovel” plays. In the window before the design locks, their certainty is actually higher than the end material choice.
— END —
Sources & Disclaimer: NVIDIA GTC official materials (NVL576 / NVLink 7), IEC 61249-3-6:2026 standard, Rogers official datasheet, Taiflex official website, Panasonic Megtron product literature, GF Securities overseas electronics & communications research, SemiAnalysis analysis, and July–August 2026 supply chain survey reports. Evidence grading reflects the author’s judgment and may contain biases. Companies mentioned do not constitute investment advice. Please cite source when reposting.