AI Networking Market Analysis — May 2026

Market Size & Growth

• **Disclosed AI-networking revenue/orders floor: ≈ $3.4B+ (FY2025)** across the named L3 set — Arista ~$750M AI back-end (scale-out Ethernet) line [signal fdd4bbeb / beb3a8d2], Cisco ~$2B webscale AI-infrastructure *orders* [signal ceb9081c], Credo $436.8M FY2025 revenue [signal 9b0f5caa / 04e9ab97], Semtech data-center ~$200M annualized (≈$50.3M Q1 FY2026) [signal 6dc937f5]. This is a directly-disclosed *floor*, not a clean TAM — it mixes orders, systems revenue and component revenue (see Reverse-Hype).
• **3-year CAGR proxy: ≈ 40–50%.** Component lines are compounding 27–30% YoY (Arista, three FY2025 quarters $2.005B→$2.205B→$2.31B [signal 2db4f47d]), ~126% (Credo FY2025 [signal 9b0f5caa]), and >100% (Semtech data-center YoY Q2 FY2026 [signal 4ec7d5c5]). The cleanest multi-year anchor: Celestica's AI-led revenue $9.65B (FY2024) → ~$16B (FY2026 outlook), ≈29% 2-yr CAGR [signal 8dbb4243 / 436ae9ad].
• **Key drivers:**
• Single-tier 102.4 Tbps switch silicon (Broadcom Tomahawk 6) collapsing hop count for 100,000+ GPU clusters [signal fe5310e7].
• Hyperscaler clusters scaling toward 50,000–100,000 GPUs on Ethernet scale-out fabric [signal beb3a8d2].
• Sovereign-AI buildouts adding national-scale fabric demand (Cisco–HUMAIN, Saudi PIF) [signal 24a20019].

Supply Chain

• **Upstream:** merchant switch ASICs (Broadcom Tomahawk 6, 102.4 Tbps [signal e0c88073]), Cisco Silicon One P200 [signal 33fc99d1], 200G/lane SerDes & retimer IP [signal 968f78b1], optical DSP/PMDs and active copper cable ICs [signal b43b1685], plus ODM/EMS assembly (Celestica whitebox [signal d2b13d0f]).
• **Downstream:** hyperscalers (sole-source whitebox builds [signal db93312d]), sovereign-AI operators (HUMAIN [signal 24a20019]), and GPU-cluster/neocloud operators running 50k–100k-GPU training fabrics [signal beb3a8d2].

`[Broadcom/SerDes IP] → [switch+router systems / AEC / optics: Arista, Cisco, Celestica, Credo, Semtech] → [hyperscaler + sovereign GPU clusters]`

Trend Lines (ranked by importance)

**1. Ethernet invades the scale-up domain (the NVLink/UALink wall is now contested).**

• Evidence: "the Ethernet for Scale-Up Networking (ESUN) workstream launched ... extending Ethernet into the scale-up domain to compete with NVLink/UALink interconnects" [signal 2b699c6d].
• Why now: this is not "AI got better" — it is two discrete 2025 standards/silicon events: UEC 1.0 spec shipped 2025-06-11 defining an Ethernet transport scaling to ~1M endpoints [signal 88a64169], and the OCP ESUN workstream formed 2025-10-13 with AMD, Broadcom, Cisco, Meta and Microsoft [signal 2b699c6d]. Scale-up was a closed NVLink moat until these landed.
• Implication (6–12 mo): copper scale-up suppliers (Semtech CopperEdge, Credo 1.6T AEC) inherit a *second* TAM inside the rack [signal 968f78b1]; expect first ESUN-aligned design discussions to surface in hyperscaler 2026 fabric RFQs.

**2. Copper's revenge for sub-7m runs.**

• Evidence: "full-year FY2025 revenue of $436.8 million, up roughly 126% ... driven mainly by hyperscaler demand for its AEC connectivity products" [signal 04e9ab97].
• Why now: 200G-per-lane SerDes entering production ramp in 2025 (Credo 1.6T AEC [signal 968f78b1]) plus GPU-rack power/thermal ceilings make optical power untenable for short in-rack runs — flipping active-copper economics. Semtech CopperEdge moved to *volume* production with data-center net sales >100% YoY [signal 4ec7d5c5].
• Implication: AEC/ACC attach per GPU rack rises; the unglamorous cable layer is compounding faster than the switch headlines, and a second Credo hyperscaler crossed >10% of revenue [signal 7398907d].

**3. Scale-across: routing (not just switching) becomes an AI line item.**

• Evidence: "Cisco launched the 8223 fixed router ... 51.2 Tbps ... 64 x 800G ports for scale-across AI data center interconnect ... linking distributed GPU clusters across multiple AI data centers" [signal 33fc99d1].
• Why now: single-site gigawatt power limits in 2025 force training across geographically distributed datacenters; deep-buffer 800G DCI routers (Silicon One P200) shipped 2025-10-08 specifically for this [signal 33fc99d1] — a topology that did not exist as a product category in early 2024.
• Implication: deep-buffer routing TAM opens alongside leaf/spine switching; expect scale-across to appear as a distinct budget line in hyperscaler 2026 capex.

Key Inflection Points (Watch List)

1. **First production UEC-conformant fabric at scale** — UEC 1.0 spec shipped 2025-06 [signal 88a64169]; the signal to watch is silicon, not spec.
2. **Tomahawk 6 GA execution** — platforms were "planned" as of 2025-06 [signal fe5310e7]; watch GA shipping Etherlink/whitebox systems through 2026.
3. **Cisco webscale AI order run-rate** — >$1.3B in a *single* quarter (Q1 FY2026) [signal 6d46e8ab]; watch whether it sustains >$1B/qtr or decelerates.
4. **Credo customer de-concentration** — 86% of FY2025 revenue from one hyperscaler [signal 2a71d353]; second >10% customer is the normalization signal [signal 7398907d].
5. **Celestica FY2026 ~$16B reaffirm/raise** — the cleanest whitebox-fabric demand tell [signal 436ae9ad].
6. **Juniper–HPE close (regulatory 2026)** — reshuffles the L3 incumbent map.

Reverse-Hype Warnings

**Overhyped: a clean "AI networking TAM," and "Ethernet has beaten NVLink."** Every confident market-size chart for this space is double-counting. Cisco's headline ~$2B is *orders*, not revenue [signal ceb9081c]; Arista's $750M is recognized revenue [signal fdd4bbeb]; Celestica's contribution is whitebox *systems* revenue that already embeds Broadcom silicon [signal db93312d]; Credo and Semtech sell the components inside those same systems [signal 04e9ab97, signal 6dc937f5]. Summing them inflates the pool. Likewise, ESUN is a *launched workstream* (2025-10) [signal 2b699c6d] and UEC 1.0 is a *spec* (2025-06) [signal 88a64169] — neither is shipping production scale-up Ethernet against NVLink's installed base. Specs are not silicon.

**Underrated: concentration fragility and the copper layer.** The entire disclosed pool is hostage to ~3–5 hyperscaler capex programs — Credo at 86% single-customer [signal 2a71d353], Celestica on disclosed sole-source whitebox builds [signal db93312d]. One deferred hyperscaler training cluster reprices the whole group. Separately, the most boring layer — active copper cable — is compounding fastest (Credo +274% Q1 FY2026 [signal 60d62e6c]; Semtech data-center >100% [signal 4ec7d5c5]) while attention sits on switch silicon. And the incumbent transformation is overstated: Cisco's blended growth is still ~8% with AI a sub-4% revenue mix [signal fbb66b0e] — AI is the lever, not yet the base.