Skip to main content
Invest AI
From market noise to cited evidence
L5Upstream

Networking

Interconnect, switching, and high-speed data-movement silicon linking compute clusters.

Stack position

Where Networking sits in the AI stack

  1. L4 · Custom Silicon

    ASICs and custom accelerator designs (in-house or merchant silicon) for specific AI workloads.

  2. L5 · Networking

    Interconnect, switching, and high-speed data-movement silicon linking compute clusters.

  3. L5 · Optical

    Optical transceivers and fiber connectivity for datacenter-scale networking.

  4. L6 · AI Server

    GPU servers, racks, and storage systems integrating accelerators into deployable units.

Icons mark layer roles — not company logos. Position only; not a price chart.

Why this layer matters

A frontier training run spans tens of thousands of accelerators that must behave as one machine. That produces a distinct hierarchy: links inside the package, scale-up inside the rack, scale-out between racks, plus the switches and NICs that carry it. This is a first-class industry, not a footnote under semiconductors.

What is the bottleneck?

Cluster scale is increasingly limited by what the fabric can sustain rather than by how many accelerators can be bought.

A structural observation about supplier concentration and lead times — not a view on any security.

Who sits in this layer?

  • Company

    MRVLPrimary business sits here

    Marvell Technology

    Data-center interconnect and networking silicon.

  • Company

    ANETPrimary business sits here

    Arista Networks

    Datacenter switching systems used in large AI cluster fabrics.

  • Company

    AVGOPrimary business sits here

    Broadcom

    Merchant switch silicon and datacenter networking components.

  • Company

    ALABPrimary business sits here

    Astera Labs

    Connectivity semiconductors for accelerator and memory interconnect.

  • Company

    CRDOPrimary business sits here

    Credo Technology

    High-speed connectivity products for datacenter interconnect.

  • Company

    CSCOAdjacent / mixed

    Cisco Systems

    Enterprise and service-provider networking broadly; AI datacenter fabric is one segment.

  • Company

    APHPrimary business sits here

    Amphenol

    High-speed interconnect, cable assemblies, and connectors for datacenter systems.

Weekly digest: AI-map layer changes and new public 13F filings, each with its source. No score, no grade, no trade signal. Unsubscribe anytime.

Public 13F filings citing these securities

Research context, not a recommendation. These are facts on file with the SEC; the classification that puts a security in this layer is ours, and the filer made no such claim.

  • ARK Investment Managementreported a position inAVGO
    Form 13F-HR for the period ending 2026-09-30, filed 2026-11-13 · SEC filing
  • Bridgewater Associatesreported a position inAVGO
    Form 13F-HR for the period ending 2026-09-30, filed 2026-11-16 · SEC filing
  • Coatue Managementreported a position inAVGO
    Form 13F-HR for the period ending 2026-09-30, filed 2026-11-16 · SEC filing
  • Tiger Global Managementreported a position inAVGO
    Form 13F-HR for the period ending 2026-09-30, filed 2026-11-16 · SEC filing

Rows below are public SEC Form 13F-HR filings, reported by the filing entity for the period shown and filed later — they are never current holdings. Form 13F covers long US equity positions only: it excludes short positions, cash, non-US holdings and most derivatives, and a reported position may have been sold since. Assigning a ticker to a layer is XingAI's classification, not the filer's. Nothing here is a recommendation, and an institution reporting a position is not a reason to take one.

Coverage boundary

Coverage boundary: US-listed tickers or ADRs only (same rule as TSM / ASML). Important suppliers without a US listing are omitted even when they matter to the stack. That is a coverage rule, not a quality ranking.

No hand-listed notable omissions for this layer yet. The US-list / ADR rule still applies.

Checked 2026-10-09. Not exhaustive.

Same track

Frequently asked questions

What is the bottleneck?

Cluster scale is increasingly limited by what the fabric can sustain rather than by how many accelerators can be bought.

Why does this layer matter?

A frontier training run spans tens of thousands of accelerators that must behave as one machine. That produces a distinct hierarchy: links inside the package, scale-up inside the rack, scale-out between racks, plus the switches and NICs that carry it. This is a first-class industry, not a footnote under semiconductors.

Who sits in this layer?

Listed alphabetically, not ranked: MRVL: Data-center interconnect and networking silicon. ANET: Datacenter switching systems used in large AI cluster fabrics. AVGO: Merchant switch silicon and datacenter networking components. ALAB: Connectivity semiconductors for accelerator and memory interconnect. CRDO: High-speed connectivity products for datacenter interconnect. CSCO: Enterprise and service-provider networking broadly; AI datacenter fabric is one segment. APH: High-speed interconnect, cable assemblies, and connectors for datacenter systems.

Descriptive positioning only. The positioning label says whether a company's business is exposed to a layer, never whether to buy it. Layers with no exposure are not recommendations to add any.