← research bank

Arista: Customer concentration is diluting, not concentrating, as AI scales

gen-arista · conviction — · status open · horizon — · as of 2026-08-07

Arista's customer HHI of 2,529 implies a structurally diversified book—roughly equivalent to five equal 20% customers—even as it ships into the most concentrated AI capex cycle in history. Oracle is ramping cloud infrastructure at 17% growth, CoreWeave at 200%, and Apple is building private capacity at scale.
Robust to undisclosed shares. 3 derived inputs under this thesis; redrawing every supply weight the industry does not publish moves none of them by more than 25%. Computed from evidence at most 17 days old (oldest input: alphabet).

Exhibits

Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.
68177286SOXX 225ANET 13612mo, indexed to 100 at start · dashed = SOXX benchmark

Series available as data/gen-arista.csv

Exhibit 2What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.
Arista's customer base is diversifying structurally as AI infrastructure spending broadens beyond the hyperscale fiveArista Networks — customer concentration (H… †100.0%Oracle Corporation †80.0%CoreWeave, Inc.75.0%Ai Capex Cycle85.0%COMPOSED (and)63.7%

† 2 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

64% if the 2 gates are independent, 75% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: CoreWeave, Inc. at 0.75 — CoreWeave growth is filed but scale vs. Arista's revenue base uncertain

The 19-point gap between capex pull and demand pull signals orders from customers whose own revenue is not yet scaling proportionaArista Network… — capex pull at least 35.21… †100.0%Arista Networ… — demand pull at least 15.97… †100.0%Arista Networks75.0%COMPOSED (and)75.0%

† 2 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

One gating premise, so the conclusion is exactly as strong as it. The figure is an ordering device, not a calibrated probability — see how the numbers are made.

Weakest link: Arista Networks at 0.75 — Gap interpretation assumes orders lead revenue; timing or working-capital effects possible

Arista operates at the fabric layer where every GPU architecture and every new infrastructure entrant becomes a customer, decoupliScale-out network (between nodes)85.0%Broadcom Inc. †90.0%Arista Networks — gross margin 64.1% †100.0%COMPOSED (and)85.0%

† 2 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

One gating premise, so the conclusion is exactly as strong as it. The figure is an ordering device, not a calibrated probability — see how the numbers are made.

Weakest link: Scale-out network (between nodes) at 0.85 — Ethernet dominance in AI back-end assumed but InfiniBand or proprietary fabrics compete in some topologies

The variant

Consensus

Arista is seen as a hyperscaler play with binary AI backend risk. The market prices 62× EV/FCF on the assumption that Microsoft and Meta dominate revenue and that success requires defending share with two or three logos. Customer concentration is viewed as rising because AI spend is concentrated.

Variant

Arista's customer HHI of 2,529 implies a structurally diversified book—roughly equivalent to five equal 20% customers—even as it ships into the most concentrated AI capex cycle in history. Oracle is ramping cloud infrastructure at 17% growth, CoreWeave at 200%, and Apple is building private capacity at scale. The hyperscaler narrative misses that Arista sells middleware—Ethernet fabric—not compute, so every new GPU buyer becomes a new switching customer. AI capex concentration at the chip layer is *creating* customer diversity at the interconnect layer.

Differentiator

Supply-chain models conflate chip-layer and fabric-layer concentration. Arista sits one layer below compute: every entrant into AI infrastructure—cloud, sovereign, enterprise—buys the same back-end Ethernet regardless of GPU vendor. The 35% capex pull and 16% demand pull gap shows orders arriving faster than hyperscaler revenue growth can explain, evidence of breadth the earnings multiple ignores.

Falsifiers

Open questions

Reasoning chain

Arista's customer base is diversifying structurally as AI infrastructure spending broadens beyond the hyperscale five VALID
premises

Low HHI during peak hyperscaler AI build suggests non-traditional customers already material; composition supports diversification claim

The 19-point gap between capex pull and demand pull signals orders from customers whose own revenue is not yet scaling proportionally STALE
premises

Capex outpacing revenue growth by 19 points indicates infrastructure pre-build by customers not yet monetizing AI at scale

Arista operates at the fabric layer where every GPU architecture and every new infrastructure entrant becomes a customer, decoupling it from chip-layer winner-take-all dynamics VALID
premises

Middleware position and merchant silicon create structural customer diversity as infrastructure proliferates regardless of compute architecture outcomes

Write-up

Pre-filled skeleton: gen-arista.md