AI fabric architect designing rail / fat-tree / dragonfly topologies at 100k+ GPU scale.

Audience Profile

• Age / Experience: Senior IC, 10-20 years
• Current role: Network Architect / AI Fabric Lead / Principal Engineer (Hyperscaler / AI lab)
• Top pain points:
• Spectrum-X vs Tomahawk 6 vs Silicon One real benchmark comparison data
• Optical vs AEC tradeoff per intra-rack segment
• Multi-tenant fabric isolation guarantees
• Top decision blockers:
• Switch silicon supplier consolidation reducing leverage
• Vendor-proprietary congestion control
• (Only two declared for this segment)

What This Segment Needs

• Information: Independent Spectrum-X vs Tomahawk 6 vs Silicon One throughput/latency-under-incast data; collective-comm bandwidth contention models at 100k+ GPU.
• Tools: Topology planners that simulate rail / fat-tree / dragonfly blast radius and per-segment optical-vs-AEC BOM/power.
• Services: Vendor-neutral congestion-control bake-offs and multi-tenant isolation conformance testing.

Top 5 Companies for You (Fit Score)

| Rank | Company | Score | Why | |------|---------|-------|-----| | 1 | Arista Networks | 84/100 | Sequential accel Q1 $2.005B→Q3 $2.31B (~27% YoY); reaffirmed ~$750M AI back-end Ethernet for 2025; UEC steering member (2025-06-11); Tomahawk 6 (102.4 Tbps) HW/SW co-design; AI Networking SA hire 2026-04-21. | | 2 | Credo Technology | 84/100 | Q1 FY2026 record $222.8M (+274% YoY); production 200G/lane SerDes + 1.6T AEC frontier work. Caveat: single hyperscaler ~86% of FY2025 revenue, second >10% only Q1 FY2026. | | 3 | Cisco | 83/100 | Q1 FY2026 webscale AI orders >$1.3B in one quarter, guidance raised; Silicon One P200/8223 shipped Oct 2025 (51.2 Tbps, 64x800G); RoCEv2/ECN/PFC + congestion-mgmt in NX-OS — directly relevant fabric stack. | | 4 | Semtech | 78/100 | Data center +100% YoY Q2 FY2026; CopperEdge ACC in volume production; Staff DSP roles on 200G/lane PAM4. But ~$50M/qtr run-rate is a small base, one-product concentration. | | 5 | Celestica | 76/100 | FY2025 ~$12.2B (+26.4% YoY), FY2026 outlook ~$16B; builds 1.6T Ethernet / Tomahawk 6 / co-packaged optics. EMS execution shop: thin margin, Broadcom roadmap + sole-source whitebox dependence. |

Deal-Breakers (Your Hard Preferences)

No hard preferences declared for this segment.

How to Evaluate Any Company in this Niche (Checklist)

• [ ] Check growth signals: sustained 3+ consecutive quarters of back-end/scale-out Ethernet revenue raises, not single-quarter order spikes.
• [ ] Check comp data: triangulate Principal/Staff Network Architect bands via levels.fyi + recent H1B/PERM filings; vendor pages give none.
• [ ] Check learning signals: live reqs naming 224G PAM4 SerDes, co-packaged optics, RoCEv2/PFC congestion control — not generic 'cloud networking'.
• [ ] Check stability signals: single-customer revenue concentration disclosure (>40% = fragility) and merchant-silicon dependence on one Tomahawk/Spectrum-X roadmap.
• [ ] Check culture signals: ask interviewers whether congestion control is open (UEC/RoCEv2) or vendor-proprietary, and who owns multi-tenant isolation SLAs.
• [ ] Check standards posture: confirm UEC/ESUN/OCP membership level (steering vs participant) and ship date of last frontier ASIC.

Reverse-Hype Watch

Material warnings: Semtech's data-center 'grew >100% YoY' is off a small ~$50M/quarter base concentrated in one CopperEdge product at a few hyperscale scale-up customers, with optical/co-packaged substitution risk. Generalize the pattern: Credo and Celestica carry the same single-customer / single-silicon-roadmap concentration even without explicit warnings.

Under-reported for this segment: public coverage chases revenue growth and port speed, but almost nothing is published on multi-tenant fabric isolation guarantees or whether a vendor's congestion control is proprietary lock-in. Job posts and earnings calls won't tell you incast/collective-comm behavior at 100k+ GPU — that only surfaces in reference-customer back-channels and standards-body (UEC/ESUN) participation depth.