Most of the AI spending conversation still starts with chips. Arista Networks’ latest quarter suggests the wiring between those chips deserves equal billing, and that AI networking is turning into a line item buyers scrutinize rather than an afterthought. Analyst Zeus Kerravala of ZK Research made that case in a piece first reported by SiliconANGLE, and the company’s numbers back much of it up.
Key Takeaways
- Arista Networks reported a 58.5% revenue increase to $3.66 billion, with strong AI networking growth projected to double to about $7 billion.
- AI networking has become a performance decision, as efficiency directly impacts costs and output in AI projects.
- Arista stands to gain from three factors: increasing accelerator diversity, scale-across capabilities, and potential entry into rack-scale Ethernet solutions.
- Investors expressed concerns about deferred revenue and rising R&D costs, but Arista remains optimistic about its long-term margins and demand visibility.
- Going forward, technology buyers should integrate network considerations into AI planning, as AI networking plays a crucial role in overall system performance.
Table of contents
The quarter in numbers
Arista booked third-quarter revenue of $3.66 billion. That’s up 58.5% from a year earlier and 20.5% from the prior quarter, the fastest growth since the company went public, and well above its own $3.3 billion forecast. Adjusted earnings per share reached $1.20, compared with 77 cents a year ago.
Gross margin came in at 64% and operating margin at 49.1%. The company holds $14.4 billion in cash and carries no debt.
Management raised its 2026 outlook for the fourth time this year, to about $13.3 billion, or 48% growth. It expects fourth-quarter revenue between $3.9 billion and $4 billion. The first look at 2027 calls for 38% to 40% growth, to roughly $18.5 billion.
Inside that forecast, AI networking revenue is expected to double from $3.5 billion to about $7 billion. Campus networking is projected to grow 60% to around $2 billion. Shares jumped about 12% in after-hours trading before settling to a gain of roughly 5% the next day.
Why AI networking is now a performance decision
The reasoning here is about money. In a large cluster, a dropped packet or a failed link leaves costly accelerators waiting for data. Inference makes that worse, because the work is split into stages such as prefill, decode, storage and cache, each running on different resources that have to talk to each other constantly.
CEO Jayshree Ullal described this as a shift from a stack built around compute to one built around the network. She tied network design directly to token efficiency, which is the cost of producing each unit of AI output.
That matters because boards and finance chiefs increasingly judge AI projects on cost per token and accelerator utilization. Both depend on the network. Kerravala’s example is simple: a 5% gain in GPU utilization across a build measured in gigawatts is worth far more than the switches that produced it. Once buyers run that math, AI networking stops looking like a commodity purchase.
Three reasons Arista stands to gain in AI networking
Every networking vendor benefits from this awareness, but Kerravala argues Arista benefits more. He points to three drivers.
Accelerator diversity
Three years ago, Arista talked about four or five AI customers running 50,000 to 100,000 Nvidia GPUs, often over InfiniBand. Today it counts more than 100 Etherlink customers. It expects more than 10 gigawatts of deployments across 10 or more accelerator types by the end of 2027.
Nvidia is a serious rival, and its Spectrum-X Ethernet has landed hyperscalers including Meta and Oracle. But its networking is tuned for its own full stack. As buyers bring in silicon from AMD, Microsoft’s Maia, Meta’s MTIA, Qualcomm and Cerebras, many will want a network supplier that doesn’t also sell a competing accelerator.
In Kerravala’s view, each new accelerator widens the slice of AI networking open to an independent Ethernet vendor, and Arista leads that category. Co-design work with Cerebras and Maia puts the company inside system design from the start.
Scale-across
Power limits are pushing AI capacity across multiple buildings and regions. Linking those sites takes deep buffers, large routing tables and very high reliability. Ullal said scale-across made up about 30% of AI demand this year and should grow in 2027, anchored by the 7800 platform. Arista’s routing heritage and EOS software count for the most in this corner of AI networking.
Scale-up
At the OCP summit, Arista showed rack-scale Ethernet designs based on the ESUN specification, with trials planned for 2027 and production in 2028. If ESUN displaces proprietary options such as UALink or NVLink, as Ullal expects, Arista would sell inside the rack for the first time. That would open a segment of AI networking the company has never served.
Security adds a fourth angle. CTO Ken Duda said Arista is working with Anthropic, Google and OpenAI to bring AI-driven vulnerability discovery into its software pipeline, including through Project Glasswing. President Todd Nightingale credited that effect for campus growing well ahead of the market.
What worried investors
The stock had slid more than 20% in the month before earnings, according to Chartmill, so analysts came in with questions. The answers were mostly reassuring, with some caveats.
- Deferred revenue: The balance was flat at about $6.8 billion, which some read as softening demand. CFO Chantelle Breithaupt said heavy turnover in and out of the balance masked steady delivery and that the year-end figure will top last year’s.
- Spending and margins: R&D rose to $408.5 million from $278.1 million, which Duda tied to hyperscaler prototypes and liquid cooling. Gross margin slipped from 65.2% a year ago, but the long-term range stays at 62% to 64%.
- Supply and visibility: Purchase commitments climbed to $11.4 billion and inventory to $3.2 billion, with multiyear agreements covering silicon, circuit boards and memory. Arista has two quarters of firm order visibility and says it has about two years of work behind that.
Read Next
A few related pieces worth your time:
- Top Managed IT Service Providers in Spokane – 2026 Reviews
- Machine-Assisted Search Helps You Spot Great Toronto Listings Faster
- Bitcoin Price Prediction: Can BTC Hold $76,000 After the Binance Hack and $1 Billion in ETF Outflows?
What to watch next
For years the Arista debate was whether Ethernet could beat InfiniBand. Kerravala thinks InfiniBand will keep a role but that Ethernet’s wider reach means it grows faster. The sharper question for AI networking now is how much of an AI system the network ends up defining.
William Blair’s Sebastien Naji echoed that view in a note, keeping an Outperform rating and calling Arista a core AI infrastructure holding. Kerravala sees the main risk as execution rather than demand or deferred revenue: staying precise while the company gets larger and orders speed up.
For technology buyers, the practical takeaway is to price the network into AI plans early. Ask how it affects utilization, token cost and the freedom to mix accelerators. If Arista’s forecast holds, AI networking will be one of the more closely watched parts of the AI buildout through 2027.
Source: SiliconANGLE
