Competitors, stated neutrally

1Rules of engagement

Competitive conversations are where FAEs damage themselves fastest. Three rules keep you safe and, incidentally, more persuasive.

Only vendor-published figures. If it is not on the vendor’s own page, you do not say it. Every number below was re-fetched on 2026-09-07 from the vendor that publishes it.^[1][2]

Use their vocabulary. Arista calls its design a “cell-based fabric” with “Virtual Output Queues (VOQ)” and “Distributed Credit Scheduling”.^[1] Say that, not a paraphrase that sounds like a flaw. Translating a competitor’s architecture into your own vocabulary is how you end up wrong in front of an engineer who has run it.

Concede your own gaps first. The three sentences that do the most work in a competitive meeting are all admissions: this figure I cannot verify; this capability we do not have; this switch you can buy from Dell today instead of ours.^[3][5]

2Arista, in Arista's words

The 7800R4 is a chassis line: 7816LR4 (16 slots), 7812R4, 7808R4 and 7804R4.^[1] Maximum 800G ports run 576 / 432 / 288 / 144 across those four, and maximum 400G ports 1,152 / 864 / 576 / 288.^[1] Peak capacity runs from “115 Tbps (230 Tbps FDX)” on the 7804R4 to “460 Tbps (920 Tbps FDX)” on the 7816LR4.^[1] Power efficiency is published as “6W per 100Gbps”.^[1]

The architecture claim is the substantive one, and it is not a marketing line: a cell-based fabric that sprays uniform cells across all fabric links, virtual output queues to remove head-of-line blocking, and distributed credit scheduling so every egress port schedules independently.^[1] Arista describes the result as operating “like a single extremely large switching chip that is 100% internally lossless and fair”.^[1]

Two honest notes. First, that is a genuinely different answer to the same problem NVIDIA solves with per-packet adaptive routing plus SuperNIC reordering — a chassis-internal scheduling approach versus a distributed fabric approach.^[7] Neither page publishes a head-to-head. Second, the fetched Arista page does not name a chipset; the widely repeated Jericho3-AI attribution comes from a datasheet PDF surfaced in search, not from the page, so do not assert it.^[1]

3Cisco, including the sentence you have to be able to hear

Cisco’s AI POD design pairs a Nexus 9364E-SG2 backend switch — “64x 800GbE ports”, “51.2 Terabits per second (Tbps)” and a “256 MB” on-die buffer — with a Nexus 9364D-GX2A frontend switch at “64x 400GbE QSFP-DD ports” and 51.2 Tbps.^[2] Both fabrics are “a two-tier, non-blocking spine-leaf (Clos) topology”, with the backend built as an “MP-BGP EVPN VXLAN fabric” providing lossless Ethernet and RoCE.^[2]

Then the part that matters more than any port count. Cisco’s GPU servers connect through “eight NVIDIA BlueField-3 backend NICs” — the document specifies the B3140H SuperNIC with a single 400GbE port — into “four upstream leaf switches” in a “4-way rail-optimized topology”.^[2] And Cisco states plainly that the design “is compliant with NVIDIA Enterprise Reference Architecture and endorsed by NVIDIA for Infrastructure Configuration and Spectrum-X”, referencing the 2-8-9-400 Enterprise RA node — two CPUs, eight GPUs, nine NICs, 400 GbE per GPU.^[2][12]

Internalise the consequence: a competitor ships an NVIDIA-endorsed implementation of NVIDIA’s own reference architecture, using NVIDIA NICs, on their switches. The NIC is not the contested layer. The switch is, and so is the software stack around it.^[2][6]

4Broadcom: the row you cannot fill

Broadcom’s Tomahawk 6 product page and its associated press release were fetched for the research notes and re-fetched on 2026-09-07 while writing this lesson. Both returned content-free shells: no switching capacity, no SerDes count, no co-packaged-optics variant name, no power claim.^[3][4] Every Tomahawk 6 figure in circulation — the 102.4 Tbps single chip, the 512 lanes of 200G SerDes, the “TH6-Davisson” CPO variant, the 70% optical-power reduction — reaches us through search summaries, not through a primary page.^[3][4]

So the Broadcom row in your competitive sheet has one honest value: could not verify on a primary page. Not “Broadcom does not publish”, not “the claims are inflated” — just that you could not source them and you will re-check.

This is not an evasion, it is the strongest available position. If a customer reads you a Tomahawk 6 number and you counter with a different unverified number, you have taught them that both vendors’ engineers repeat marketing. If you say you could not find it on Broadcom’s page and offer to look again with them, you have demonstrated exactly the standard you apply to your own claims.

The commercial reality sits alongside it: Dell resells Broadcom-Tomahawk-based PowerSwitch Z9864F-ON at “64x 800 GbE OSFP112 + 2x SFP+” and 51.2 Tbps, and Z9664F-ON at 64x 400 GbE and 25.6 Tbps, on the same AI switch catalog page as the NVIDIA SN5610, SN5600D and SN5400.^[5] Your customer does not have to leave Dell to buy the alternative. That is the single most useful fact in this lesson, because it tells you what your job is: not to win a vendor argument, but to be the person who explains which design the customer’s reference architecture and support path actually require.^[10]

5The differentiator you can actually defend

Every vendor in this lesson publishes a large throughput number. Cisco publishes 51.2 Tbps per switch, the same figure as SN5600 and SN5610; Arista publishes chassis capacities up to 460 Tbps; NVIDIA publishes SN6600 at 102.4 Tb/s and SN6800-LD at 409.6 Tb/s.^{[2][1][8][11]} A throughput duel is a duel you cannot win and do not need.

What no competitor page publishes is a version-matched stack. The Spectrum-X validated solution stack tracks rows from “April 2024 (v 1.0.1)” to “September 2026 (v2.3.1)”, and the current row names the switch NOS, ConnectX-8 firmware, BlueField-3 firmware, DOCA-Host build, NetQ, NCCL, HPC-X, Network Operator and DTS versions for three validated GPU platforms.^[6] That is the artifact to put on the table, because it answers the question a customer is really asking in a competitive meeting — “what happens in eighteen months when something breaks” — with a document instead of a promise.