AI data centers are not just GPUs. They are GPUs wired together at 800G and 1.6T, with switches, optics, and custom silicon doing the heavy lifting between racks.
Three public names sit at the center of that wiring layer. Arista Networks, Cisco, and Marvell Technology each take a different slice of the same pie.
With NVIDIA reporting Q1 FY27 results on Wednesday May 20, networking peers will trade off NVDA's data center commentary all week.
The Networking Stack Inside AI Data Centers
An AI training cluster has two networks. The front-end network connects servers to storage and the outside world. The back-end network connects GPUs to each other.
The back-end is where speed matters most. A single training job can stall if one link is slow. That is why hyperscalers are pushing 800G ethernet today and 1.6T into 2027.
Ethernet versus InfiniBand
InfiniBand has long been the default for tightly coupled GPU clusters. It is fast and low-latency, but it is also proprietary and tied to one vendor's roadmap.
Ethernet is the open alternative. The Ultra Ethernet Consortium is rallying hyperscalers around standards that match InfiniBand performance at scale. That shift is the single biggest tailwind for Arista and the merchant silicon ecosystem behind it.
Where the spend lands
Microsoft, Meta, Alphabet, and Amazon have guided combined capex above $320 billion for 2026, per investor commentary. A meaningful slice flows into switches, optics, and custom accelerators rather than GPUs alone.
ANET Profile: Cloud Titan Spend Tailwind
Arista is the merchant ethernet pure-play. Full-year 2025 revenue hit $9 billion, up 29% year-over-year, and management raised its 2026 AI networking revenue target to $3.25 billion from $2.75 billion.
Cloud and AI customers made up 48% of 2025 revenue, with enterprise and financial services at 32%. That mix is concentrated but diversifying, which matters when one hyperscaler trims capex.
The 7800R and 7060X moat
Arista's 7800R and 7060X switches anchor AI back-end fabrics at Meta and Microsoft. The company's EOS software layer is the sticky part. Network engineers train on it, and ripping it out is painful.
What can break the story
Hyperscaler concentration is the risk. If Meta or Microsoft pauses an AI build-out, Arista feels it fast. The 2026 guidance assumes hyperscaler spending keeps accelerating through year-end.
CSCO Profile: Stabilizing Growth and Splunk Synergies
Cisco is the incumbent that lost ethernet share to Arista, then tried to buy its way back through software. The $28 billion Splunk acquisition is finally showing up in numbers.
Q1 FY26 revenue reached $14.9 billion, up 8%, with Networking up 15% to $7.8 billion. AI infrastructure orders hit $1.3 billion in Q1 alone, and management now expects roughly $3 billion in FY26 AI hyperscaler revenue.
Silicon One and the 8223 router
Cisco's Silicon One chip family is the in-house answer to merchant ethernet. The new 8223 router uses Silicon One P200 at 51.2 Tbps. Cisco expects to ship its one-millionth Silicon One chip in Q2 FY26.
Why Splunk matters
Splunk gives Cisco a data and security layer that pure networking peers cannot match. Splunk's ARR grew double digits in Q1, and joint Cisco-Splunk sales motions closed one of Splunk's largest deals to date. Software and services now sit above 50% of revenue, which softens the hardware cycle.
MRVL: Optical, DSP, and Custom Silicon Crossover
Marvell is the most leveraged to two trends at once. The first is optical interconnect, where its DSPs power 800G and 1.6T modules. The second is custom silicon for hyperscalers building their own AI chips.
FY2026 revenue was $8.2 billion. Q4 data center revenue hit $1.65 billion, 74% of the total and up 21% year-over-year. Q1 FY27 guidance points to $2.4 billion plus or minus 5%.
AWS Trainium and Microsoft Maia
Marvell co-designs AWS Trainium silicon, with visibility extending through Trainium 3 in 2026. It also supplies IP and back-end services for Microsoft Maia accelerators. Both programs ramp on 3nm nodes through this fiscal year.
The optical leg
Marvell introduced a 2nm 800G optical DSP and the industry's first 1.6T ZR and ZR+ coherent modules. Optical content per AI cluster keeps rising as link counts climb, even if pricing per port compresses.
Conclusion
These three names are not interchangeable. Arista is the cleanest bet on ethernet share gains, Cisco is the slower-growth software-plus-silicon story, and Marvell is the indirect play on custom AI silicon and optical interconnect.
NVDA earnings on Wednesday May 20 will reset assumptions across the group. Strong data center commentary lifts all three setups. Soft commentary pressures Arista and Marvell first, with Cisco's software mix offering some buffer.
Investors who want exposure to the AI networking theme can buy fractional shares of ANET, CSCO, and MRVL on Gotrade and build a basket without taking a single-name view.
FAQ
Is ethernet really replacing InfiniBand in AI clusters?
Ethernet is closing the gap, especially at 800G. Hyperscalers prefer open standards, and the Ultra Ethernet Consortium is driving the shift. InfiniBand still leads in some tightly coupled clusters.
Which stock is most exposed to a single hyperscaler?
Arista has the highest cloud and AI customer concentration at 48% of 2025 revenue. Cisco is more diversified across enterprise. Marvell depends heavily on AWS Trainium volumes.
How does Splunk fit Cisco's AI story?
Splunk adds data analytics and security software that pairs with Cisco networking. Software and services are above 50% of revenue, which smooths hardware cycles and lifts margins.
Why does NVDA earnings move ANET, CSCO, and MRVL?
NVDA data center commentary signals hyperscaler GPU demand. Strong GPU orders translate into more switches, optics, and custom silicon spend at the three networking peers.
