Arista Networks Passed $3 Billion — And Three Structural Questions Nobody Is Asking

Arista Networks hit its first $3 billion quarter in Q2 2026. Revenue came in at $3.036 billion, up 37.7% year-over-year and well above guidance of roughly $2.8 billion. Non-GAAP EPS was $1.02, a 40% increase. Management raised full-year revenue guidance for the third time to approximately $12.6 billion, implying 40% growth. Management also unveiled the 7060XE7, a next-generation 1.6-terabit switch, and announced that AI fabric revenue is on track for at least $3.5 billion in FY2026.

The consensus response has been reflexive. Analysts upgraded. The sell-side ran headlines declaring the buy thesis intact. The stock is up 44% year-to-date.

Most of the industry is reading Arista's results backwards. The question isn't whether Arista executed. The question is whether Arista's structural position supports a valuation that has already baked in flawless execution through 2028.

The stock trades at 59 times trailing earnings, 22.6 times trailing sales, and 48 times EV/EBITDA. That is not a multiple for a company that buys its silicon, assembles it, and sells it through two customers that account for 42% of revenue.

The Nvidia Problem Nobody Wants to Acknowledge

In Q1 2026, Nvidia generated $2.1 billion in data center Ethernet switch revenue — up 192.7% year-over-year — and surpassed both Cisco and Arista to become the largest vendor in data center Ethernet switching with a 21.5% market share. Two years ago, Nvidia held less than 4% of that market.

This is not a sidebar about Nvidia diversifying. This is the competitive earthquake that Arista's earnings narrative sidesteps.

Nvidia's Spectrum-X platform bundles switches, DPUs, optics, cables, and software into a pre-optimized "AI factory" package. Hyperscalers no longer buy networking as an isolated silo; they buy it as a validated GPU cluster component. When Meta, Microsoft, Oracle, and xAI deploy Ethernet-based AI clusters, a significant fraction of that switch content flows through Nvidia's integrated stack, not through a standalone Arista switch.

Arista's management told investors on the Q2 call that the industry "lives in an NVIDIA world." They also said the fourth customer from the original Ethernet AI training cohort has "officially moved from InfiniBand to Ethernet at production scale."

When the industry lives in Nvidia's world and Nvidia now sells more Ethernet switches than you do, the Ethernet win story becomes significantly more complicated.

IDC's Brandon Butler identified the real battleground: Nvidia needs to translate its hyperscale AI momentum into enterprise networking. That is where Arista still holds decades of installed base and relationships. But the hyperscale AI revenue that drives Arista's current growth trajectory — the scale-out, scale-across, and front-end AI fabric layers — is increasingly a bundled Nvidia problem. IDC data shows Nvidia captured 21.5% of the total Ethernet switch market, which includes enterprise. If that share compounds, Arista's "default across all hyperscalers" claim — which independent analysis has already flagged as unsupported — becomes harder to sustain.

The Broadcom Tax

Arista does not design its own switching silicon. It relies on Broadcom Jericho-class processors — Jericho3-AI powers the 7800R4 platform — and Broadcom Tomahawk, Trident, and StrataDNX silicon across its broader portfolio. This is not a strategic choice; it is a structural dependency.

Gross margins tell the story that management commentary obscures. GAAP gross margin fell to 62.9% in Q2 2026 from 65.2% a year earlier. Non-GAAP gross margin dropped from 65.6% to 63.4%. Management guided for a full-year range of 62% to 64%, explicitly "inclusive of expected supply-chain cost increases." The CEO characterized memory costs as "horrendous." On the Q1 call, management stated "every chip is challenged." Purchase commitments jumped from $6.8 billion to $8.9 billion sequentially.

Meanwhile, Broadcom reported a 49% operating margin in Q1 2026, up from 32% two years ago. Broadcom has the structural advantage: it designs the silicon that everyone needs, sells it at high volume, and retains full pricing power. Arista sits in the middle, absorbing supplier cost inflation to maintain its customer commitments.

The margin profile in this supply chain belongs to the silicon designer, not the system integrator. Arista is the assembly layer.

Operating margin stayed at 45.4% GAAP through expense discipline — operating costs held flat at $520 million despite 9% sequential revenue growth. That is legitimate execution. But the gap between gross margin compression and operating margin stability is closing. You cannot hold operating expenses perfectly flat while revenue triples. The math works until it doesn't.

How Much of This Is Just Cisco Collapsing?

Arista's share gains in high-speed switching are real. But a significant portion of those gains comes from Cisco's structural decline, not from Arista's competitive superiority.

Cisco reported a 25% operating margin in the most recent quarter — flat since Q2 2024 — weighed down by restructuring costs and a painful hardware-to-software transition. Cisco Silicon One G300, announced in February 2026, delivers 102.4 Tbps throughput and claims 28% improvement in job completion time. But Cisco's execution has been substandard for years, and the market has moved on. Arista is capturing share from a collapsing incumbent.

Amazon and investors credited Graviton's "success" to Amazon's custom chip strategy. Graviton only looked good because of years of Intel incompetence. The same pattern is at work here.

When incumbent incompetence drives challenger growth, the question isn't how great the challenger is. The question is what happens when the incumbent stops collapsing, or when a better entrant appears.

Cisco Silicon One is not a dead end. It's merchant silicon from Broadcom's partner channel, running on Cisco's management plane. If Cisco ever executes its software transition successfully — or if hyperscalers adopt Silicon One hardware from multiple vendors — the incumbent-collapse narrative reverses.

The Customer Concentration Binary

Two customers — Microsoft at 26% and Meta at 16% — account for 42% of Arista's revenue. Management said these partnerships "could never be stronger." They also acknowledged that at least one, possibly two, additional customers could exceed the 10% revenue threshold.

Management has told investors that AI fabric adoption has broadened from 4-5 customers in 2024 to more than 100 cumulative customers. That is a genuine expansion, and it matters. But the revenue math doesn't lie: the top two still control nearly half the business.

The structural vulnerability is straightforward. If either Microsoft or Meta pauses a data center buildout, shifts capital priorities toward custom silicon or white-box alternatives, or redirects spending toward Nvidia's integrated stack, the impact on Arista's revenue is immediate and material. At 59x earnings, the market has priced in zero tolerance for customer-level disruption.

International revenue rose to 23% of total in Q2, up from 15.5% a quarter earlier. Enterprise networking is at "record" levels. These are positive diversification signals. But they don't replace the two customers that still drive the majority of AI fabric revenue.

The Product Roadmap Is Real — With Caveats

Arista's engineering work is genuine. The 7800R4 series delivers a deep-buffer, VOQ (Voice Queue) architecture with up to 32GB of dynamically assigned packet memory per line card and 460 Tbps aggregate throughput in a 16-slot chassis. The XPO (extended pluggable optics) connector, endorsed by 100+ vendors, is positioned to reduce switch racks by up to 75% and provide 4x OSFP density. The 7060XE7 targets 1.6T with linear pluggable optics that cut interconnect power by approximately 60%.

Independent analysis of the 7800R4 architecture confirms that the VOQ scheduling and deep packet memory are genuine technical differentiators for RoCE-based (RDMA over Converged Ethernet) AI fabrics. The per-slot throughput of 28.8 Tbps checks out against the published specs.

But the architecture also has constraints. The 7808R4 requires specific fabric and cooling module configurations that independent teardowns flagged as mislabeled in public diagrams — a sign that even the engineering community struggles to parse Arista's chassis composition accurately. More importantly, the modular chassis advantage is conditional: ODMs like Celestica now hold nearly 50% combined share with Nvidia in the AI back-end Ethernet segment. Cisco Silicon One and disaggregated white-box strategies limit Arista's exclusive control over scale-across economics.

The product is competent. The moat is conditional.

The Cross-Currents

Where this leaves the investment thesis:

• AI networking demand is real and broadening. Scale-across networking alone is projected to represent 30% of Arista's $3.5 billion AI target — a category that was "virtually nonexistent" a year ago. The TAM for scale-across is estimated at $15 billion by 2030. That is a genuine expansion, not a marketing invention.

• Nvidia's Ethernet switch revenue now exceeds Arista's in absolute terms. Nvidia's 21.5% market share, up from under 4% two years ago, represents a structural competitive shift. Arista competes in the same Ethernet layer that Nvidia is now dominating through integration. Directionally, this is a negative for Arista's long-term share in hyperscale AI.

• Gross margin compression is structural, not cyclical. Broadcom's 49% operating margin versus Arista's compressed 62-64% gross margin reflects the fundamental economics of this supply chain. Arista absorbs supplier costs to secure allocation. This is sustainable as long as demand exceeds supply, but it is not a margin profile that justifies 59x earnings.

• Customer concentration remains a binary risk. Microsoft and Meta at 42% combined is a structural vulnerability that diversification signals haven't yet offset. The market's zero-error tolerance at current multiples is the real question.

• Cisco's collapse is real but not permanent. Silicon One exists. The incumbent-failure tailwind is genuine but temporary.

Directionally, the cross-currents point to a company whose growth is genuine but whose structural economics — merchant silicon dependency, Nvidia integration threat, and hyperscaler concentration — are being priced as if they were structural advantages.

The earnings beat doesn't resolve these questions. It just makes them more urgent.