Sep 22, 2026, CSCO +4/10, ANET +5/10, CIEN +4/10, NVDA +7/10, AVGO +3/10
Takeaway: AI's networking impact stems from the escalating costs of GPU racks, which makes idle bandwidth expensive. Hyperscalers are pushing for the move from 800 gbps to 1.6 tbps bandwidth and pulling the AI-interconnect to grow 3x faster than networking hardware overall. Arista (+5/10) captures that shift best and with the strongest validation; Ciena (+4/10) and Cisco (+4/10) sit on the optical-component backlog and legacy incumbent sides respectively; and Nvidia (+7/10) and Broadcom (+3/10) bookend the underlying fight over whether bandwidth runs on proprietary or open standards. Worth watching: how much of this spend is incremental versus funded by CIOs pulling budget from everything else.
Research series: This is the summary of our research about how AI is driving Networking stocks. Relevant research:
Market growth
Gartner estimated worldwide IT spending in 2026 at $6.37T, +14% YoY.
2026 growth is ~$800B.
$316B of this is from Data Centre Systems (AI-optimized servers & hyperscale expansions) and ~$200B from Software (GenAI model integration & AI-ready enterprise applications).
The remaining $280B of the $800B growth is distributed between IT services (system implementation & managed cloud), devices (memory price inflation & AI-capable endpoint refreshes), cloud infrastructure (demand for cloud-hosted AI model workloads) and telecom (network bandwidth).
Data Center Systems is the single fastest-growing category. Gartner projects spend will reach $822B in 2026, 62.5% YoY, one of the most drastic single-year surges in enterprise technology history.
Gartner’s analyst, John-David Lovelock, said that building the compute capacity is
“the largest infrastructure project ever attempted by humanity.”
The $822B figure spans servers, storage, power, cooling and networking.
Data center networking hardware market is estimated at ~$32.7 billion in 2026, projected to reach $71.4 billion by 2033, 11.8% CAGR.
Inside that, the AI-specific sub-segment, high-speed interconnects and fabrics for 800G and 1.6T switching, is growing faster, at over 30% CAGR.
This isn’t just more boxes, it’s the price of each box.
Standard enterprise servers costing $10-15K are being displaced by AI server nodes costing $200-300K each, driven largely by the GPU modules. These servers are 8-GPU configurations, and a single high-end card is $22-$33K.
Memory now accounts for >80% of GPU’s bill of materials, a real constraint given the three companies producing HBM (MU, SKHY and Samsung) have already pre-sold their 2026 output.
Fully built AI racks now cost roughly $3.9M vs $500K for a traditional rack, a 7x increase, with the jump from 10-25kW to 80-132kW power density (and 200kW+ for next-generation systems), the second major cost driver behind memory.
How this drives networking growth
Data center networking hardware addressable market is $32.7B (2026) → $71.4B (2033), and within it, the >30% CAGR AI-interconnect sub-segment.
The more expensive a GPU rack becomes, the more costly it is for it to sit idle waiting on data. That’s the reason 800 gbps is giving way to 1.6 tbps: hyperscalers are willing to pay to keep the network from bottlenecking the compute. This is the mechanism behind Arista’s and Cisco’s order growth.
The power and cooling requirements explain why racks are getting denser (more GPUs, tighter together): the “east-west” traffic within a rack and between racks has exploded. This in turn means that more compute packed closer together needs more bandwidth.
The networking hardware market is $32.7B in 2026, growing to $71.4B by 2033, modest next to the headline AI numbers, but the specific AI-interconnect slice inside it is growing three times faster than the category average, because the exploding cost of the compute is what’s making faster interconnects worth paying for.
Networking trends
The Ethernet-vs-InfiniBand.
Hyperscalers are migrating AI clusters from Nvidia’s proprietary InfiniBand to open Ethernet, specifically RoCEv2 (RDMA over Converged Ethernet).
The reasons are: lower cost, an open multi-vendor ecosystem, and superior port density at 800G/1.6T, set against growing hyperscaler pushback against Nvidia’s proprietary lock-in (InfiniBand and Spectrum-X, bundled with its DGX GPU racks).
This is an ongoing vendor fight, not a settled outcome.
2. The 800G-to-1.6T speed transition
800G ports represent ~38% of active capacity in 2026, with 1.6T transceivers now entering production for next-generation Blackwell and custom-ASIC clusters.
The technical mechanism: moving from 8×100G to 8×200G SerDes signaling doubles per-lane throughput without doubling physical port or cable density, enabled by next-generation switch silicon like Broadcom’s Tomahawk 6 and Nvidia’s Quantum-X800/Spectrum-6, both running at 102.4 Tbps.
Higher-capacity chips let operators build multi-tier fabrics with fewer total switches, lower power per bit, and fewer optical modules.
3. Enterprise and WAN convergence
Forrester’s Q3 2026 research found networking vendors unifying data center, campus, and WAN fabrics into single-pane operating systems (Arista’s EOS) to reduce operational friction as enterprise workloads move into hybrid AI models.
4. Power and thermal constraints as a market driver
As switching fabrics and optics consume a growing share of total rack power, vendors innovating in power-efficient silicon, co-packaged optics (CPO), or liquid-cooled networking gear are positioned to capture share specifically on that basis — not just on raw switching speed.
How stocks fair against these trends
Arista Networks (ANET, +5/10) is best positioned vs trend 1.
Barclays’ TMT desk named Arista the primary pure-play beneficiary of the RoCEv2 Ethernet migration, setting a street-high $289 target based on multi-year supplier purchase commitments extending into 2027, the same as order-backlog strength (purchase commitments nearly tripling to $9.7B).
Forrester’s Q3 2026 Wave credited Arista’s advantage specifically to software consistency: modular EOS spanning data center, campus, and SD-WAN, simplifying Zero Trust policy enforcement and AI-driven telemetry across domains - trend 3.
Arista is named as “the hyperscaler standard” for cloud-scale Ethernet fabrics at Microsoft and Meta specifically, with the risk being high customer concentration among a small number of top-tier hyperscalers.
Cisco (CSCO, +4/10) sits on the adapting side of Trend 1, not the leading side.
Its $9.3B hyperscaler AI infrastructure orders in 2026 came via Silicon One, Cisco’s switching architecture built to compete in the same open-Ethernet shift Arista is capturing.
Sell-side consensus frames this order backlog as validating Silicon One’s ability to bridge legacy enterprise networks into hyperscale AI backbones, a defensive move against the risk of losing hyperscale share.
Cisco’s core strength remains its enterprise campus base, where refresh cycles are slower than hyperscale data center buildouts.
Ciena (CIEN, +4/10) operates on Trend 2, one layer beneath the switches — the optical transceivers and interconnects that move data at 800G and 1.6T.
The component-level bottleneck already documented for Ciena (pump laser and indium phosphide supply) is the direct physical constraint behind management’s own “supply, not demand” framing of its swollen backlog.
Nvidia (+7/10) and Broadcom (+3/10) complete the two ends of the switching-silicon fight described above.
Nvidia's vertically-integrated push (InfiniBand and Spectrum-X) bundled into its GPU racks sits alongside customer friction. Hyperscalers have begun absorbing 15%+ price increases on AI servers carrying Nvidia chips, driven by HBM memory cost pass-through, without any sign of delaying deployment. A test of pricing power Nvidia's lock-in commands even as the Ethernet migration builds against it.
Broadcom, meanwhile, confirms its role as the silicon backbone under the non-Nvidia switching market, though its AI-semiconductor growth (143% YoY last quarter, guided past 200%) is diluted in the headline numbers by slower legacy RF and infrastructure-software segments.
Step 4: What to invest in
Arista is the highest-conviction name: the clearest, most externally validated beneficiary of the single largest structural trend, the Ethernet migration.
Ciena offers the most direct, if narrower, exposure to the speed transition, a bet on the component layer rather than the switch layer, with strong, if occasionally punished by the market, execution record.
Cisco is the name to hold with the least urgency. AI order momentum exists, but it’s a small fraction of a larger, slower-moving business, making it more a “does the incumbent successfully defend share” story than a “who captures the fastest-growing market segment”.
This report synthesizes public market data gathered via web research and reports about Cisco, Arista Networks, Ciena, Nvidia and Broadcom. It is for information only and is not investment advice.