Networking Watch: NVIDIA Enters AI Infrastructure Battle, Pushes Networking into Standalone Growth Line
Watch Tech Series · Networking Watch | September 29, 2026
The CODEW Networking Watch tracks the technologies, companies, and infrastructure developments reshaping AI, cloud, data centers, Ethernet, optical networking, and enterprise connectivity.
Networking has stopped being the plumbing underneath the AI story and become a headline of its own. The global data center networking market is on track to roughly triple, from about $39.5 billion in 2025 to more than $93 billion by 2032, and every major networking vendor's most recent earnings call has turned into an AI infrastructure update. Cisco booked $9.3 billion in AI infrastructure orders from hyperscalers in fiscal 2026 — a roughly 4.5x increase year-over-year — and is guiding to $7.5 billion in AI hyperscaler revenue for fiscal 2027, up from about $4 billion recognized in fiscal 2026.
The single most important development shaping the market right now is the contest between Ethernet and NVIDIA's proprietary InfiniBand/NVLink stack for control of AI cluster networking — a fight that pulls in every major vendor covered below, and one where the outcome will determine how open or closed the AI infrastructure stack becomes.
Networking Developments to Watch
1. Arista Networks — AI networking target raised to $3.5 billion
What happened: Arista Networks reported Q1 2026 revenue of $2.71 billion (up 35.1% year-over-year) and raised its full-year AI networking revenue target to $3.5 billion — roughly double the prior year. Two customers, Microsoft and Meta, each account for more than 10% of revenue.
Why it matters: Arista has taken the No. 1 data center Ethernet switch position from Cisco and holds roughly a 20.7% share of the data center switching segment, with its Etherlink portfolio spanning 10 gigabit to 800 gigabit speeds.
Technical significance: Arista is preparing customers for the next transition, from 800G to 1.6-terabit platforms, and is leaning on liquid cooling and linear-drive pluggable optics (LPO) — which strip the digital signal processor out of the optic — to manage power constraints that are now as binding as bandwidth.
Competitive implications: Arista's single software stack, EOS, spanning everything from campus switches to 1.6T AI platforms, is its core differentiator against both Cisco and white-box competitors.
Companies affected: Cisco, Broadcom (Arista's primary silicon supplier), NVIDIA.
2. NVIDIA — Spectrum-X and Quantum-X push networking into a standalone growth line
What happened: Analyst consensus pegs NVIDIA's fiscal 2026 networking revenue at roughly $29.3 billion, up about 125% from fiscal 2025's $13.0 billion, driven by the Spectrum-X Ethernet platform (switches, BlueField DPUs, and LinkX cabling) and Quantum-X InfiniBand/silicon-photonics switching.
Why it matters: NVIDIA is no longer just a GPU company selling networking as an accessory — networking has become a multi-billion-dollar product line in its own right, and one increasingly sold as an integrated bundle with GPU clusters.
Technical significance: Spectrum-X targets scale-out Ethernet networking between racks; NVLink remains NVIDIA's proprietary scale-up interconnect inside a rack — a distinction that shapes where Ethernet can realistically compete with NVIDIA in the near term.
Competitive implications: NVIDIA is simultaneously a customer, partner, and competitor to Cisco and Arista — evidenced by Cisco's own N9100 switch, which is built on NVIDIA Spectrum-X silicon, even as Cisco and Arista also sell Ethernet gear that competes directly with Spectrum-X.
Companies affected: Cisco, Arista, Broadcom, Marvell.
3. Cisco — Silicon One G300 and a $9.3 billion AI infrastructure order book
What happened: Cisco Systems introduced a new generation of AI-focused data center networking platforms built around its 102.4 Tbps Silicon One G300 switch silicon, alongside the Silicon One P200-based 8223 router. AI infrastructure orders from hyperscalers reached $9.3 billion in fiscal 2026, a roughly 4.5x increase year-over-year, with fiscal 2027 revenue guided to $7.5 billion.
Why it matters: Cisco's networking product orders grew 40% year-over-year in the most recent quarter — the eighth consecutive quarter of double-digit growth — evidence that its comeback in AI-era networking is broad-based, not a single hyperscaler win.
Technical significance: Cisco is positioning its vertically integrated stack — silicon, optics, networking, security, and observability — as a differentiator against point-product competitors, and reported that optics content per AI deployment is expanding as back-end clusters require higher optical port density per rack.
Competitive implications: Cisco is fighting to reclaim ground lost to Arista in data center Ethernet while also partnering with NVIDIA (the N9100) — a hedge that lets Cisco compete and cooperate with the same rival simultaneously.
Companies affected: Arista, NVIDIA, Broadcom, HPE/Juniper.
4. HPE completes its $14 billion Juniper Networks acquisition
What happened: HPE closed its acquisition of Juniper Networks on July 2, 2025, after an 18-month regulatory process, paying $40.00 per share (roughly $13.4 billion in cash consideration). The DOJ settlement required HPE to divest its Instant On campus/branch business and provide limited licensing access to Juniper's Mist AIOps platform.
Why it matters: The deal doubled the size of HPE's networking business, which HPE expects to eventually contribute more than half of the company's total operating income — a structural bet that networking, not servers, is now HPE's core growth engine.
Technical significance: The combination merges HPE's Aruba campus/wireless networking with Juniper's AI-native, Mist-powered enterprise and data center portfolio into a single vendor offering silicon, hardware, operating system, security, and software.
Competitive implications: Creates a third full-stack networking challenger to Cisco and Arista, with the scale to compete for large enterprise and service-provider contracts that neither HPE nor Juniper could credibly pursue alone.
Companies affected: Cisco, Arista, Nokia, Dell.
5. Marvell — building an optical interconnect portfolio through acquisition and R&D
What happened: Marvell Technology acquired Celestial AI, a specialist in photonic fabric technology for scale-up optical interconnects, and completed its acquisition of XConn Technologies, adding PCIe and CXL switching silicon. At ECOC 2026 in Málaga, Marvell demonstrated 2nm 400G-per-lane PAM4 optics, an 800G ZR/ZR+ pluggable with MACsec, 1.6T ZR coherent optics, and a 102.4 Tbps co-packaged optics platform.
Why it matters: Marvell is assembling, through acquisition, a full portfolio spanning custom AI silicon, advanced packaging, and optical interconnect — positioning itself as an alternative to Broadcom for hyperscalers building custom AI infrastructure.
Technical significance: Marvell's argument, presented at AI Infra Summit 2026, is that multi-trillion-parameter models and growing KV caches push memory and compute demand beyond what fits in a single 72-XPU rack — meaning scale-up networking has to leave the rack, and copper interconnect can't make that jump; only optics can.
Competitive implications: Marvell's optics roadmap now directly overlaps with Broadcom's and with Coherent's PhotonLink platform; the company still needs to publish measured efficiency data and firm production timing to convert technology demonstrations into design wins.
Companies affected: Broadcom, Coherent, NVIDIA, hyperscalers building custom silicon.
6. Broadcom — Tomahawk 6 and a raised AI networking target
What happened: Broadcom raised its fiscal 2026 AI networking revenue target to $3.25 billion, up from a prior estimate of $2.75 billion, and continues shipping Tomahawk 6 switch silicon alongside its custom XPU design program for hyperscalers.
Why it matters: Broadcom sits underneath much of the industry as a merchant silicon supplier — Arista's switches run on Broadcom chips, and Broadcom's custom XPU business gives hyperscalers a path to accelerators that don't depend on NVIDIA at all.
Technical significance: Broadcom captures value primarily at the silicon and accelerator level with higher margins than system vendors, while depending on partners like Arista to turn that silicon into deployable systems — a genuine interdependence rather than a simple supplier relationship.
Competitive implications: If NVIDIA's Spectrum-X gains further share in AI networking, it threatens both Arista's system sales and Broadcom's silicon attach rate simultaneously — the two companies' fortunes are more tightly linked to each other than either is to NVIDIA.
Companies affected: Arista, NVIDIA, Marvell, hyperscalers with custom silicon programs.
7. Ultra Ethernet Consortium — an open standard takes aim at InfiniBand
What happened: The Ultra Ethernet Consortium released its first full specification in June 2025, defining an Ethernet-based networking system designed specifically for AI and high-performance computing at scale — with hyperscalers and enterprises reportedly migrating away from proprietary InfiniBand toward open Ethernet since its release.
Why it matters: Until recently, Ethernet wasn't considered fast or reliable enough for the back-end networks connecting GPU clusters; InfiniBand was the default. UEC is the industry's coordinated attempt to close that gap through an open standard rather than a single vendor's proprietary technology.
Technical significance: Over time, the consortium's backers expect migration to extend beyond scale-out networking (between racks) into scale-up networking (within a rack), where NVIDIA's NVLink currently dominates — via a complementary open standard called UALink.
Competitive implications: A credible open alternative to both InfiniBand and NVLink would reduce customer lock-in to NVIDIA's full-stack networking approach — directly benefiting Arista, Cisco, Broadcom, and Marvell, all of which are UEC members.
Companies affected: NVIDIA (most exposed), Arista, Cisco, Broadcom, Marvell, hyperscalers.
8. Co-packaged optics moves from lab to early hyperscale deployment
What happened: Co-packaged optics (CPO) — integrating optical I/O directly onto or adjacent to switch silicon, eliminating the electrical bottleneck between ASIC and transceiver — is moving from lab validation toward early commercial hyperscale deployment, with Broadcom, Marvell, NVIDIA, Intel, and Ayar Labs all advancing competing CPO roadmaps.
Why it matters: As switch bandwidth climbs toward 51.2T and 102.4T, pluggable optics start hitting power and density limits that CPO is specifically designed to solve.
Technical significance: NVIDIA has already announced Spectrum-X and Quantum-X switch silicon incorporating co-packaged optics; the industry-wide roadmap is moving from 800G through 1.6T toward 3.2T optical data rates over the next several years.
Competitive implications: CPO adoption timing will shape which optics and silicon suppliers capture the next phase of AI data center spend — the technology remains early enough that no single approach has yet become the default.
Companies affected: Broadcom, Marvell, NVIDIA, Intel, Coherent, Ayar Labs.
Table 1 · Developments to Watch, Summary
| Company/Tech | Category | Key Figure |
|---|---|---|
| Arista Networks | AI/data center Ethernet | $3.5B FY26 AI networking target |
| NVIDIA Spectrum-X/Quantum-X | AI cluster networking | ~$29.3B FY26 networking revenue (est.) |
| Cisco Silicon One | Switch silicon, routers | $9.3B FY26 AI infra orders |
| HPE–Juniper | Enterprise/campus/DC networking | $14B acquisition, closed July 2025 |
| Marvell (Celestial AI, XConn) | Optical interconnect, CXL/PCIe | 102.4 Tbps CPO platform demoed |
| Broadcom Tomahawk 6 | Merchant switch silicon | $3.25B FY26 AI networking target |
| Ultra Ethernet Consortium | Open standard | First full spec, June 2025 |
| Co-packaged optics | Optical interconnect | Lab to early hyperscale deployment |
AI Is Changing Networking
Bandwidth requirements are compounding faster than any prior networking cycle: the industry moved from 400G to 800G in roughly two years and is now pushing toward 1.6T, with 3.2T already on vendor roadmaps. Latency matters differently for AI than for traditional enterprise traffic — GPU clusters are latency-sensitive in a way that makes even small delays in chip-to-chip communication costly at scale, which is why NVIDIA's proprietary NVLink still leads for the tightest, rack-internal connections even as Ethernet gains ground everywhere else.
Network architecture is bifurcating into scale-up (within a rack, dominated by NVLink today) and scale-out (between racks, where Ethernet is winning share fastest). Data-center fabrics are being redesigned around this split, with vendors increasingly selling complete rack-to-cluster networking systems rather than individual switches. Interconnects now have to handle east-west traffic patterns — GPU-to-GPU communication — that dominate AI back-end clusters, a fundamentally different traffic profile than the north-south client-server patterns most enterprise networks were built for.
Ethernet's credibility for AI back-end networking is a genuinely new development — as recently as two years ago it was considered a second-tier option behind InfiniBand, and the Ultra Ethernet Consortium's 2025 specification is the industry's formal answer to that gap. Optical infrastructure is under the most direct pressure of any layer: as switch bandwidth climbs past 51.2T and 102.4T, pluggable optics are running into power and density limits, which is what's pulling co-packaged optics out of the lab and into early hyperscale deployment.
The CODEW Lens: AI didn't just increase demand for networking — it changed what "good" networking means, shifting the design target from minimizing latency for a single request to maximizing sustained throughput across tens of thousands of simultaneously communicating chips.
Competitive Landscape
Cisco is executing a broad-based AI networking comeback, backed by vertically integrated silicon, optics, and security, while also partnering with NVIDIA on the Spectrum-X-based N9100 switch. Arista Networks holds the No. 1 data center Ethernet switch position with roughly 20.7% segment share, differentiated by its single EOS software stack. Broadcom sits underneath much of the market as the dominant merchant switch-silicon supplier (Tomahawk 6) while also running a fast-growing custom AI accelerator (XPU) design business for hyperscalers.
NVIDIA has turned networking into a standalone, multi-billion-dollar business through Spectrum-X and Quantum-X, and remains the only vendor with a fully proprietary, end-to-end AI networking stack from chip to switch to cable. Juniper Networks, now part of HPE following the $14 billion 2025 acquisition, brings AI-native, Mist-powered networking into HPE's broader hybrid cloud and AI portfolio. HPE expects the combined networking business to eventually generate more than half of total company operating income.
Nokia is investing in AI-specific networking research, including a new AI Networking Innovation Lab, as it works to extend its telecom infrastructure position into enterprise and data center AI networking. Huawei posted $895 million in Ethernet switch revenue in Q1 2026 (up 27.2% year-over-year, roughly 5.8% global share), remaining a significant player outside markets affected by U.S. export restrictions. Dell continues to compete primarily as a systems integrator bundling third-party switch silicon into its infrastructure offerings rather than developing proprietary switch silicon of its own. Marvell is the fastest-moving challenger to Broadcom in custom silicon and optical interconnect, building its position through acquisition (Celestial AI, XConn Technologies) rather than organic development alone.
The CODEW Lens: The clearest structural feature of this landscape is that competitors are also suppliers to each other — Arista buys from Broadcom, Cisco partners with NVIDIA on one product while competing against it on others. Few of these companies can afford to treat any other as a pure rival.
Table 2 · Competitive Positioning
| Company | Core Position | Primary Rivalry |
|---|---|---|
| Cisco | Vertically integrated networking, security, silicon | Arista, HPE/Juniper |
| Arista Networks | #1 data center Ethernet switching | Cisco, NVIDIA Spectrum-X |
| Broadcom | Merchant switch silicon, custom XPUs | Marvell, NVIDIA |
| NVIDIA | Proprietary end-to-end AI networking | Ultra Ethernet Consortium members |
| HPE / Juniper | Full-stack enterprise + AI-native networking | Cisco, Arista |
| Marvell | Custom silicon, optical interconnect | Broadcom, Coherent |
| Nokia / Huawei / Dell | Telecom infra, regional strength, systems integration | Regionally and segment-specific |
The Infrastructure Shift
The through-line across every development above is a shift from general-purpose networking toward increasingly specialized, high-speed infrastructure purpose-built for AI. That shift shows up in silicon (switch chips designed specifically for GPU cluster traffic patterns rather than adapted from enterprise designs), in optics (co-packaged optics emerging specifically because pluggable transceivers can't keep pace with AI-driven bandwidth and density requirements), and in commercial structure (vendors increasingly selling complete networking systems tied to specific AI cluster architectures rather than general-purpose switches sold independently).
Power has become as binding a constraint as bandwidth — Arista's move toward liquid cooling and linear-drive pluggable optics, and the entire industry's push toward co-packaged optics, are both direct responses to the reality that AI data centers are increasingly power-limited rather than space-limited. This is pulling networking companies deeper into conversations about power delivery and thermal design that were traditionally the data center operator's problem alone.
The CODEW Lens: Networking is being pulled into the same specialization trend already reshaping semiconductors — general-purpose products give way to workload-specific silicon, optics, and systems, with power efficiency now as much a design constraint as raw speed.
What to Watch Next
Technology transitions: The 800G-to-1.6T switch and optics transition is the most immediate near-term shift to track, with 3.2T already appearing on vendor roadmaps for the years beyond. Watch whether Ultra Ethernet Consortium-compliant products see real hyperscaler deployment volume in 2027, and whether UALink gains similar traction in scale-up networking against NVLink.
Product launches: Marvell's 2nm optics and 102.4 Tbps co-packaged optics platform still need measured efficiency data and confirmed production timing before they translate into design wins — a gap worth watching close in coming quarters. NVIDIA's Quantum-X silicon-photonics switches represent the highest-profile CPO commercialization effort to track.
Infrastructure investments: Cisco's guided jump from roughly $4 billion to $7.5 billion in AI hyperscaler revenue between fiscal 2026 and 2027 is a useful benchmark for whether hyperscaler networking capex continues compounding at its current pace.
Partnerships and competitive battles: The Cisco–NVIDIA N9100 relationship is worth tracking as a template — expect more "coopetition" arrangements where networking vendors both compete with and build on top of NVIDIA's silicon. HPE's integration of Juniper, and whether the combined company can convert its doubled networking scale into share gains against Cisco and Arista, is a multi-year story still in its early innings.
M&A opportunities: Marvell's acquisition-led build-out (Celestial AI, XConn Technologies) suggests optical interconnect and custom silicon IP remain active M&A categories — watch for similar moves from Broadcom or Cisco to fill gaps in their own optics roadmaps.
Standards developments: Continued Ultra Ethernet Consortium specification work and early co-packaged optics interoperability standards are the two standards efforts most likely to shape which vendors' products can compete on a level playing field over the next two to three years.
The CODEW Lens: What is changing in networking, and why does it matter to the broader infrastructure stack? Networking has moved from a commodity layer beneath compute to a co-equal design constraint alongside it — GPU cluster performance is now gated as much by interconnect and power as by chip architecture itself.
The CODEW Stat
$39.5B → $93B: data center networking market, 2025-2032 Across Cisco, Arista, NVIDIA, and Broadcom alone, AI-related networking revenue targets for 2026 already sum to well over $40 billion — evidence that the networking layer beneath AI has become one of the most closely watched, and fastest-growing, corners of the entire technology infrastructure stack.
Reviewed by Erwin Castro
on
Tuesday, September 29, 2026
Rating:
No comments: