INDUSTRY ANALYSIS
NVIDIA Vera Rubin’s “Cable-Free” Design: Three Interconnect Layers and a Changing Supply Chain
Claims that “Vera Rubin computers do not use optical cables” conflate cable-free modular compute trays with the rack’s NVLink 6 copper backplane. This article separates intra-tray, intra-rack, and inter-rack connections to examine Vera Rubin NVL72’s engineering changes. It maps the competitive positions of advanced CCL / PCB materials, high-speed copper cables, and CPO silicon photonics, identifying potential shifts in value while defining the limits of public evidence.
How “cable-free” created confusion about copper and optical connections
In 2026, NVIDIA advanced the Vera Rubin NVL72 platform into volume production and demonstrated redesigned compute trays at CES, GTC, and the OCP Global Summit. Official emphasis on a modular, cable-free compute tray, alongside an intra-rack NVLink backplane containing more than 5,000 copper cables, led some market commentary to simplify the message into “Vera Rubin is abandoning optical cables.”[1][2][3]
That interpretation confuses physical layers within a server system. Connections inside one tray, within one rack (scale-up), and across racks (scale-out) face different distances, attenuation, and power constraints. Removing flyover cables inside a tray does not mean eliminating copper cables from the rack. Keeping copper inside the rack does not eliminate the need for fiber in inter-rack networks.[2][4]
Although the broad principle of copper within the rack and optics across racks remains consistent with Blackwell, engineering implementations are changing at all three levels. Architecture changes, technical qualification, and actual order shares are different levels of evidence. This article focuses on engineering trade-offs and potential value redistribution across segments in the Vera Rubin generation, without estimating undisclosed company market shares or revenue.
Distinguishing intra-tray, intra-rack, and inter-rack connections
Before assessing the supply chain, hardware scope and transmission media must be separated.
Inside a tray, earlier designs often used dozens of internal high-speed flyover cables to avoid losses in long motherboard traces. Vera Rubin compute trays instead use high-density PCB midplanes and blind-mate connectors, removing internal hoses and cable assemblies and substantially shortening assembly and maintenance time.[1][2][3]
Within the rack, Vera Rubin NVL72 connects 18 compute trays—72 Rubin GPUs and 36 Vera CPUs—through nine NVLink 6 switch trays. Bandwidth rises to 3.6 TB/s per GPU and 260 TB/s across the rack, while retaining a passive copper-cable backplane.[2][3] Inter-rack networks use ConnectX-9 adapters with Quantum-X800 InfiniBand or Spectrum-X Ethernet switches for optical transmission, with co-packaged optics (CPO) switch options formally introduced.[2][4]
| Architecture level | Hardware scope | Main Vera Rubin medium | Conclusions that should not be extrapolated |
|---|---|---|---|
| Intra-tray | CPU, GPU, and peripheral connections within one compute tray | Advanced PCB midplanes and blind-mate connectors, replacing internal flyover cables | A cable-free tray does not mean an entire rack without copper or optical cables |
| Intra-rack scale-up | Connections among 72 GPUs and NVLink 6 switches in one rack | Passive NVLink copper backplane, with over 5,000 copper cables per rack | Copper remaining at the 224G generation does not mean future Kyber / Feynman systems will never adopt optics |
| Inter-rack scale-out | East-west networks linking racks into AI factories with tens of thousands of accelerators | 1.6T pluggable optical modules, CPO silicon-photonics switches, and external fiber | CPO entering production does not mean immediate wholesale replacement of pluggable modules |
The core change is therefore not the disappearance of optical cables. It is advanced boards and blind-mate interfaces replacing internal cable assemblies, 224G copper extending the intra-rack backplane’s role, and CPO silicon photonics commercializing on top of inter-rack fiber infrastructure.[1][2][4]
How changes at the three layers affect upstream and downstream processes
The following table connects engineering changes with public supplier capabilities and competitive priorities. Suppliers within a segment compete for qualification and share. Different approaches—such as M8 / M9 laminates versus PTFE / hydrocarbon materials, or pluggables versus CPO—may also coexist across customer configurations.
| Supply-chain segment | Corresponding layer | Engineering changes and requirements | Representative companies, mapped by public capabilities | Unresolved variables |
|---|---|---|---|---|
| Advanced CCL and specialty substrates | Intra-tray / switch boards | Signals move from flyover cables into midplanes and advanced substrates, increasing demand for very-low-loss materials including M8 / M9 and glass-fabric-free PTFE / hydrocarbon formulations | Elite Material, Shengyi Technology, TUC, ITEQ, Taiflex, Panasonic, Doosan | Final stackups, hybrid-lamination proportions, and individual production shares |
| High-layer-count PCB / HDI manufacturing | Intra-tray / switch boards | Orthogonal midplanes and advanced switch boards require more layers, higher hole aspect ratios, and greater backdrilling precision | Unimicron, Gold Circuit Electronics, WUS, Victory Giant Technology, TTM, Shennan Circuits, Kinwong | New-material processing yields and allocation across tray boards |
| Blind-mate connectors, slide rails, and liquid-cooling quick disconnects | Intra-tray / rack integration | Modular cable-free trays must connect power, signals, and fully liquid-cooled fluid paths through blind-mate interfaces | Amphenol, BizLink, LOTES, King Slide, AVC, Auras, Fositek | Qualified rack-interface suppliers and progress of additional sources |
| High-speed copper backplane cables / DAC | Intra-rack scale-up | NVLink 6 moves per-lane signaling into the 200G / 224G generation, tightening shielding and loss requirements for cable cartridges | Amphenol, TE Connectivity, Molex, BizLink, FIT Hon Teng, JPC Connectivity, Woer | Bare-wire and assembly outsourcing shares, and timing of future optical backplanes |
| CPO silicon-photonics packaging and optical components | Inter-rack scale-out | Co-packaging switch ASICs and optical engines requires external laser sources (ELS), CW lasers, and dense fiber array units (FAUs) | TSMC, ASE Technology Holding, Lumentum, Coherent, LandMark Optoelectronics, Browave, FOCI, ezconn | Initial CPO switch shipment scale, packaging yield, and cloud-provider adoption |
| 800G / 1.6T pluggable optical modules | Inter-rack scale-out | Mainstream links upgrade from 800G to 1.6T while evolving alongside CPO switches | InnoLight, Eoptolink, Coherent, TFC Communication | Procurement mix between 1.6T pluggables and CPO at major cloud providers |
Within the tray, eliminating 43 internal jumpers in favor of a dense midplane and blind-mate structure moves value from fine internal cable assembly toward low-loss CCL, high-layer-count PCBs, and precision blind-mate mechanical components.[1][3]
At rack level, leading connector suppliers such as Amphenol and BizLink may lose some small internal cable demand while participating in blind-mate interfaces, power busbars, and higher-specification 224G copper backplanes. The main exposure from copper losing internal tray connections falls on companies focused purely on internal harness assembly, rather than leading suppliers with rack-backplane and blind-mate capabilities.[2][3]
Which suppliers face opportunities or structural pressure?
The following assessment concerns relative competitive opportunities from Vera Rubin’s three interconnect changes. “Positive” means the direction raises potential content per rack or technical barriers in that segment. “Relatively negative” reflects removal of an existing process or competition from a new architecture, not an inevitable decline in an entire company’s revenue.
| Supplier / segment | Direction | Basis and potential effect | Limits |
|---|---|---|---|
| Advanced CCL suppliers, including Elite Material, Shengyi, TUC, ITEQ, and Taiflex | Positive | Cable-free trays and switch-board upgrades require boards and midplanes to carry higher-frequency signals, increasing demand for M8 / M9 and PTFE / hydrocarbon low-loss materials | Compute boards, midplanes, and switch boards use different materials; qualification and share competition remain |
| High-layer-count PCB manufacturers, including Gold Circuit, Unimicron, WUS, VGT, and TTM | Positive | Replacing flyovers with board traces increases multilayer, advanced-HDI, and midplane complexity and PCB value per rack | Benefits depend on difficult-board yields and customer allocation |
| Blind-mate, mechanical, and cooling suppliers, including King Slide, LOTES, AVC, and Fositek | Positive | Drawer-style cable-free trays need precision rails, blind-mate connectors, and hose-free cooling quick disconnects for rapid replacement | Adoption must be tracked separately in reference MGX and cloud-customized racks |
| Leading backplane-cable and connector suppliers, including Amphenol, BizLink, and FIT | Positive | NVLink 6 retains a copper backplane with 260 TB/s rack bandwidth, raising barriers for high-frequency cable cartridges and blind-mate structures | Lost small internal assemblies are offset by higher-value rack backplanes and interfaces; the actual balance needs tracking |
| Companies focused on internal cable assembly | Relatively negative | Removing compute-tray flyovers substantially reduces processes dependent on internal PCIe / MCIO jumper assembly | A move into external AECs or high-speed harnesses could mitigate the impact |
| Optical suppliers relying on comprehensive intra-rack scale-up fiber adoption | Conditionally negative | NVL72 retains more than 5,000 copper cables, delaying the intra-rack optical-backplane opportunity | Does not remove inter-rack scale-out opportunities |
| CPO silicon-photonics foundry and packaging / testing suppliers, including TSMC and ASE | Positive | Spectrum-X and Quantum-X photonics switches enter production, bringing optical engines into semiconductor-level packaging | Initial volumes remain constrained by advanced-packaging capacity and optical-engine yields |
| CW lasers, FAUs, and fiber connections, including Lumentum, LandMark, FOCI, Browave, and ezconn | Positive | External lasers and dense single-mode MPO / FAU coupling increase demand for upstream optical components | Long qualification cycles require separating co-development / sampling from large-volume revenue contributions |
| Leading 1.6T pluggable suppliers: InnoLight, Eoptolink, and Coherent | Neutral to positive | The 1.6T upgrade supports initial selling prices; pluggables remain a main configuration before broad CPO adoption | Monitor CPO penetration in large AI clusters over the medium and long term |
| Smaller conventional module assemblers without silicon-photonics or 1.6T capabilities | Relatively negative | Switch-side CPO moves optoelectronic packaging upstream toward semiconductor suppliers, potentially bypassing assembly-only roles | Applies to advanced AI-cluster switch ports; general enterprise networking retains conventional pluggables |
CPO production and pluggables will coexist, setting the pace of inter-rack value shifts
Another common extreme interpretation is that CPO will immediately eliminate pluggable optical modules.
NVIDIA’s product information describes Quantum-X800 InfiniBand and Spectrum-X Ethernet silicon-photonics switches built around 200G SerDes and 1.6T CPO technology, claiming 3.5–5 times the energy efficiency of conventional pluggable transceivers, with lower latency and fewer components.[4][6] Recent supply-chain and third-party analyses note that “in production” marks an important engineering milestone, but optical-engine yield, advanced-packaging capacity, and customers’ operating and maintenance practices remain challenges. CPO has not yet replaced conventional pluggables comprehensively.[5]
Inter-rack optics in the Vera Rubin generation will therefore follow two coexisting approaches:
| Inter-rack deployment scenario | Upstream silicon-photonics, laser, and FAU suppliers | Conventional pluggable-module suppliers |
|---|---|---|
| CPO switches scale first in selected hyperscale clusters | TSMC silicon-photonics platforms, external CW laser sources, and FAU coupling suppliers benefit directly | Conventional pluggable transceivers are no longer needed at switch ports adopting CPO |
| Most customers initially retain 800G / 1.6T pluggables | CW lasers and silicon-photonics chips can still be supplied for 1.6T modules | Leading 1.6T suppliers benefit from upgrades, while smaller suppliers face technical barriers |
| External fiber cabling and dense MPO connections | Both CPO and pluggables require physical inter-rack fiber and high-density connectors | No direct substitution conflict with optical-cable suppliers |
IndustryAtlas’s assessment is that fiber cabling and upstream CW laser / silicon-photonics components occupy favorable positions in both scenarios, while downstream optoelectronic interfaces diverge between semiconductor packaging through CPO and conventional pluggable-module assembly.[4][5][6]
Four signals to track next
1. Final material stackups for midplanes and switch boards. Production choices among M8 / M9 glass-fabric-reinforced CCL, glass-fabric-free PTFE, and glass-fabric-free hydrocarbon materials will directly affect material value per rack for laminate, glass-cloth, and copper-foil suppliers.[7] 2. The balance between blind-mate interfaces and 224G copper backplanes. Track leading connector suppliers’ total content per rack to assess whether those products fully offset removal of internal flyover cables. 3. Actual CPO switch penetration and yields. Monitor the procurement mix of CPO photonics switches and conventional pluggable switches as early adopters such as CoreWeave, Meta, Microsoft, and Oracle deploy Vera Rubin clusters.[4][5] 4. Backplane roadmaps for higher-density generations such as Kyber / Feynman. As per-lane speeds move toward 400G around 2028, or rack density increases further, whether intra-rack scale-up reaches copper’s physical limits and adopts optical backplanes will be the next major transition point.
Vera Rubin does not make optical cables disappear. It redraws technical requirements through more board-based interconnection inside trays, more demanding copper links within racks, and semiconductor-integrated optics across racks. Understanding the engineering requirements at these three hardware levels is more useful for assessing competitive gains and losses than simply debating copper versus light.
Sources & notes
- [1] NVIDIA official release, January 5, 2026: NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI Supercomputer. Supports the platform announcement and modular, cable-free tray design.
- [2] NVIDIA official technical documentation, October 7, 2026: Hardware — NVIDIA DGX Vera Rubin NVL72 User Guide. Primary hardware specifications for the rack configuration: 18 compute trays, nine NVLink 6 switch trays, passive copper backplane, and ConnectX-9 / BlueField-4 networking.
- [3] BigGo Finance, compilation of international reporting, July 21, 2026: NVIDIA Vera Rubin NVL72 reaches full production: tenfold token output. Supports the description of over 5,000 NVLink copper cables totaling more than two miles per rack, 3.6 TB/s per GPU, 260 TB/s rack bandwidth, and hose-free, fanless compute trays.
- [4] NVIDIA official product page, 2026: Silicon Photonics Networking for Agentic AI (Spectrum-X & Quantum-X Photonics). Supports 200G SerDes, improved energy efficiency, operation without DSP retimers, and named early adopters.
- [5] Remio AI industry analysis, August 15, 2026: NVIDIA CPO Switches Enter Production as AI Data Centers Embrace Silicon Photonics. Discusses what entering production establishes, optical-engine yield and advanced-packaging constraints, and continued coexistence with conventional pluggables.
- [6] NADDOD technical overview, March 19, 2025: NVIDIA GTC: Vera Rubin & 1.6T Co-Packaged Optics Switch. Describes Quantum-X single-mode MPO connectors and external light sources containing CW-DFB lasers.
- [7] NVIDIA official blog, October 12, 2026: Opening the Vera Rubin architecture at OCP Global Summit; Economic Daily News, October 14, 2026: NVIDIA’s NVLink PTFE display draws attention. Supports open-architecture progress and competition in next-generation board materials. Individual suppliers’ production shares still require shipment evidence.