Layer 3 — Networking & interconnect
Position in the die→grid hierarchy: inside the server.
How chips, servers and sites talk to each other: the scale-up fabric inside a rack, the switch ASICs and systems of the scale-out network, the optical transceivers and copper cables between them, and the fiber and coherent optics linking buildings and campuses.
Layer 3 covers the fabric that moves data between accelerators, servers, and racks once it leaves the package: scale-up links inside a rack or pod, scale-out switching between racks, and the optics, copper, and fiber that physically carry the signal. The chain runs from switch silicon (scale-up and scale-out ASICs) through switch systems and white-box ODMs, out to transceivers, the lasers and DSPs inside them, and the copper or fiber that connects everything — with a parallel track for campus-to-campus and metro links once a build spans more than one building. Two forces are reshaping this layer at once: rising per-GPU bandwidth is pushing scale-up links from copper toward optical and co-packaged designs as reach and cable count become limiting, while the sheer number of accelerators per cluster keeps expanding the scale-out fabric that ties pods together. Choke points sit in a few places worth watching: optics assembly and testing concentrates in a handful of contract manufacturers building modules for the merchant module makers, the laser and DSP layers inside those modules have relatively few qualified suppliers per speed grade, and the shift toward co-packaged optics and higher-radix switch ASICs is a qualification-heavy transition where incumbents in switch silicon and optical components have an edge but are not guaranteed to keep it. Copper still dominates the shortest, highest-bandwidth links for cost reasons, which keeps demand for high-speed cable assemblies growing alongside optics rather than being displaced by them.
Parts in this layer
- 3.01 Scale-up fabric (GPU-to-GPU inside the rack/pod) · 6 companiesThe fabric linking GPUs to each other inside a single rack or pod, separate from the switching fabric that connects racks to each other.
- 3.02 Scale-out switch ASICs · 5 companiesThe switch ASICs that route traffic between racks and pods, distinct from the scale-up fabric inside a single rack.
- 3.03 Switch systems & white-box ODMs · 13 companiesThe physical switches — built by branded vendors and white-box ODMs — that house scale-out switch ASICs and connect racks into a cluster fabric.
- 3.04 Optical transceivers (400G / 800G / 1.6T) · 15 companiesPluggable optical modules that convert electrical signals to light for transmission over fiber between switches and servers, sold at 400G, 800G, and 1.6T speeds.
- 3.05 Optics contract manufacturing · 5 companiesContract manufacturers that assemble and test optical transceivers on behalf of module brands and switch vendors.
- 3.06 Lasers (EML, CW/DFB, VCSEL) · 8 companiesThe laser sources — EML, CW/DFB, and VCSEL types — that generate the light signal inside optical transceivers and co-packaged optics.
- 3.07 Optical DSPs, drivers, TIAs · 7 companiesThe signal-processing silicon inside optical modules — DSPs plus the drivers and TIAs that condition the electrical signal on either side of the laser and photodetector.
- 3.08 Silicon photonics & co-packaged optics (CPO) · 18 companiesAn approach that integrates optical components directly onto or next to the switch or GPU package, replacing some pluggable transceivers with light routed closer to the silicon.
- 3.09 Copper: DAC / ACC / AEC cables · 13 companiesHigh-speed copper cable assemblies — DAC, ACC, and AEC types — that link GPUs, NICs, and switches over short distances as a lower-cost alternative to optical transceivers.
- 3.10 NVLink copper spine / backplane cartridges · 5 companiesThe copper cable assemblies and backplane cartridges that physically wire the scale-up fabric together inside a rack, most visible in Nvidia's NVL72 architecture.
- 3.11 Fiber, connectors, structured cabling · 15 companiesThe fiber optic cable, connectors, and structured cabling that carry optical signals between transceivers across a data hall or campus.
- 3.12 DCI / coherent optics (campus-to-campus, metro) · 9 companiesCoherent optical systems that carry data center interconnect (DCI) traffic between buildings on a campus or between metro sites, distinct from the shorter-reach optics inside a data hall.
- 3.13 Timing & telemetry silicon · 5 companiesPrecision timing and telemetry chips that synchronize clocks and monitor signal integrity across the network fabric.
Public plays in this layer
Listed companies with the most part placements in layer 3. Placement count is breadth, not revenue exposure.
| Company | Ticker | Parts here | Price | Mkt cap |
|---|---|---|---|---|
| Broadcom | AVGO | 8 | 372 USD | 1.77T |
| Marvell | MRVL | 6 | 241 USD | 217.0B |
| Nvidia | NVDA | 5 | 228 USD | 5.51T |
| Lumentum | LITE | 4 | 956 USD | 85.8B |
| Coherent | COHR | 4 | 295 USD | 57.8B |
| Cisco | CSCO | 4 | 112 USD | 442.0B |
| TE Connectivity | TEL | 2 | 203 USD | 58.9B |
| Sumitomo Electric | 5802.T | 2 | 2223 JPY | 6.93T |
| Luxshare | 002475.SZ | 2 | 58.24 CNY | 450.7B |
| Jabil | JBL | 2 | 312 USD | 32.7B |
| Foxconn Interconnect | 6088.HK | 2 | 5.55 HKD | 39.5B |
| Credo | CRDO | 2 | 240 USD | 45.1B |
| Corning | GLW | 2 | 153 USD | 131.4B |
| Celestica | CLS | 2 | 317 USD | 40.0B |
| Bizlink | 3665.TW | 2 | 2190 TWD | 427.2B |
| Astera Labs | ALAB | 2 | 304 USD | 52.8B |
| Amphenol | APH | 2 | 161 USD | 199.0B |
| ZTE | 000063.SZ | 1 | 33.90 CNY | 162.2B |
| YOFC | 6869.HK | 1 | 180 HKD | 256.0B |
| Wistron NeWeb | 6285.TW | 1 | 249 TWD | 120.1B |
| Volex | VLX.L | 1 | 625 GBp | 114.4B |
| TSMC | 2330.TW | 1 | 2420 TWD | 62.76T |
| Tower Semiconductor | TSEM | 1 | 220 USD | 24.8B |
| TFC | 300394.SZ | 1 | 273 CNY | 297.6B |
| Spectra7 | SEV.V | 1 |