July 10, 2026 · 11 min read
Every GPU cluster needs a nervous system: switches, routers, SerDes silicon, and optical transceivers that let thousands of GPUs communicate at microsecond latency. As hyperscalers deploy $200B+ in annual AI CapEx, the networking companies supplying this infrastructure — Arista, Marvell, Cisco — are seeing demand the industry has never seen before. Here are the four AI networking stocks worth understanding in 2026, with full bull/bear cases and analyst targets.
Every conversation about AI infrastructure focuses on GPU supply (NVIDIA H100, B100) or data-centre power. The networking layer — the switches, routers, and silicon that move data between GPUs — gets far less attention. That is a mistake.
Hyperscalers will spend an estimated $50–60B on AI data-centre networking in 2026 — switches, optical transceivers, cabling, and interconnect silicon. This is 25–30% of total AI infrastructure CapEx and growing proportionally as GPU cluster sizes scale from 10K to 100K GPU configurations.
400G networking, the standard of 2023–2025, is insufficient for 100K-GPU AI clusters. 800G Ethernet and InfiniBand are the mandatory standards for 2026+ AI training infrastructure — and every hyperscaler is in the middle of a forklift upgrade right now, driving a multi-year super-cycle for AI networking vendors.
Every major hyperscaler (Microsoft, Meta, Google, AWS) is in the middle of upgrading AI training fabric from 400G to 800G Ethernet. This cycle — estimated to generate $30–50B in incremental networking spend from 2025–2028 — is already visible in Arista's record backlog. The 800G upgrade happens at the switch, optics, and NIC layer simultaneously, creating demand across the stack.
AWS Trainium, Microsoft Maia, Google TPUs, and Meta MTIA are all AI accelerators designed with custom networking interconnects. This is a $15–25B annual TAM for companies like Marvell (SerDes, interconnect IP), Broadcom (custom ASIC), and Cadence/Synopsys (EDA tools used in chip design). The hyperscalers chose custom silicon to reduce NVIDIA dependency — and networking companies are the primary beneficiary.
NVIDIA's InfiniBand (via Mellanox) has dominated GPU-to-GPU communication in AI clusters for the past five years. Ethernet proponents (Arista, Broadcom, Intel) argue that 800G Ethernet now matches InfiniBand on latency and congestion control, at lower cost. The Ultra Ethernet Consortium (which includes Microsoft, Meta, and AMD) is the most visible signal of hyperscalers pushing Ethernet. If Ethernet wins more than 50% of new AI cluster builds in 2026–2027, Arista's TAM doubles; if InfiniBand holds, Arista's AI revenue is capped.
AI data centres in 2026 are hitting power and cooling walls before they hit networking walls. This is accelerating demand for power-efficient optical interconnects (replacing copper at 400G+) and programmable switching silicon that can be placed closer to GPU clusters to reduce cable-plant power draws. Arista's white-box routing architecture and Marvell's optical DSP chips are well positioned for this trend.
Click any stock to expand the full bull case, bear case, and analyst targets.
| Ticker | Price | Rating | AI Revenue Mix | FY2026 Rev Growth | Fwd P/E | Consensus Target |
|---|---|---|---|---|---|---|
| ANET | ~$118 | Strong Buy | ~70% | +24% | 44x | $145 |
| MRVL | ~$89 | Strong Buy | 74% | +36% | 38x | $115 |
| CSCO | ~$68 | Hold | ~10% | +6% | 19x | $72 |
| JNPR | ~$40 | Hold | ~8% | +3% | 28x | $42 |
All data as of July 2026. AI Revenue Mix = estimated % of total revenue from AI-specific products/contracts. Prices approximate.
The closest competitor to Marvell in custom AI silicon. Broadcom designs AI ASICs for Google (TPU) and Meta (MTIA). The company guided $60–90B in AI semiconductor revenue by FY2027 — but this includes custom AI accelerators, not just networking. Already a mega-cap; upside is more muted than Marvell but with less execution risk.
Makes optical transceivers — the pluggable modules that convert electrical signals to light inside data-centre switches at 400G/800G. Every Arista switch order includes an associated optical transceiver order. Coherent, II-VI, and Lumentum are the three major transceiver vendors. Higher-risk, higher-volatility exposure to the same AI networking CapEx cycle.
Another optical transceiver and laser vendor. Lumentum's silicon photonics technology (co-packaged optics) is viewed as the next-generation architecture that will replace pluggable transceivers in future 1.6T/3.2T AI switch designs. Earlier stage than Coherent but potentially higher upside if co-packaged optics reach volume by 2028.
Storage networking play. As AI training datasets grow into the hundreds of petabytes, the storage fabric connecting GPU clusters to data becomes a bottleneck. NetApp's ONTAP storage OS and its AI-specific StorageGRID product are seeing increased demand from enterprises deploying private AI models. More of an enterprise AI infrastructure play than hyperscale pure-play.
Get AI prediction signals, unlimited stock comparisons, portfolio analytics, and personalized watchlists — free for 14 days, no credit card required.
14-day free trial · No credit card required · Cancel anytime