Networking Watch: The AI Data Center Is Rewriting the Network
The AI Data Center Is Rewriting the Network
Is Networking Becoming as Strategically Important as the GPUs Themselves?
Cisco doesn't make a single GPU. Yet the company just reported $9.3 billion in AI infrastructure orders from hyperscale customers in fiscal 2026 — including roughly $4 billion in its fourth quarter alone — up from just over $2 billion the year before. That's not a rounding error in a company known for routers and switches; it's a signal that the AI buildout's second most expensive line item, after the accelerators themselves, is the network required to connect them. Cisco's networking revenue for the quarter came in around $9.8 billion, up 28% year over year, entirely on the strength of AI-driven demand.
Dell'Oro Group's research puts a number on where this is heading: back-end AI data center switching — the fabric connecting GPUs to each other and to storage — is on pace to become a roughly $1 trillion cumulative market opportunity, with scale-up networking alone (the tightest, rack-level tier connecting GPUs within a single compute domain) expected to account for more than half of that AI back-end switch market by 2030. Every major systems vendor, from Nvidia to Cisco to Broadcom to AMD, is now selling networking as a first-class architectural decision rather than a commodity afterthought, and the capital flowing into optics, switching silicon, and interconnect standards this year confirms the market agrees.
Networking Market This Week
Cisco's AI order growth is the standout number this cycle, but the broader financial picture across networking vendors tells a consistent story of demand outrunning supply. Arista Networks' most recent quarter crossed $3 billion in revenue for the first time, up 37.7% year over year, even as CEO Jayshree Ullal described the supply chain as "tight." The more consequential figure sat below the headline: Arista's multi-year forward purchase commitments to component suppliers nearly tripled year over year, from $3.6 billion to $9.7 billion, as the company locks in memory, integrated circuits, and optical components years in advance rather than risk getting caught short — the clearest evidence yet that vendors are treating AI networking component scarcity as a multi-year structural condition, not a temporary bottleneck to wait out.
Underneath the public earnings, the underlying market-sizing data keeps getting revised upward. Dell'Oro's upgraded forecast attributes the AI networking market's growth specifically to "scale-up, scale-out, and scale-across" — a three-tier taxonomy that's become the industry's standard shorthand for describing GPU-to-GPU, server-to-server, and data-center-to-data-center connectivity, respectively.
The AI Data Center Challenge
Conventional enterprise Ethernet was built on assumptions that collapse under AI training workloads. Standard networks tolerate oversubscription, expect occasional packet loss handled by TCP retransmission, and distribute traffic using flow-based hashing that works well for web servers and databases. For large-scale distributed AI training — where thousands of GPUs must complete synchronized "all-reduce" operations, exchanging gradient data across an entire cluster in lockstep — that same behavior is catastrophic. A single dropped packet or congested link can stall an entire training step across every GPU in the job.
That structural mismatch is why the industry now designs distinct fabrics for each layer: scale-up (GPU-to-GPU within a rack), scale-out (server-to-server and rack-to-rack within a data center), and scale-across (connecting entire data centers into a single logical AI supercomputer). Power is the compounding constraint layered on top: networking silicon and optics now account for a growing share of total AI data center power draw, at a moment when power availability, not chip supply, is emerging as the binding constraint on how fast new AI capacity can come online.
GPU-to-GPU Networking (Scale-Up)
Scale-up networking has become its own distinct competitive battleground, separate from the fabric connecting racks to each other — and increasingly, the industry's biggest rivals are choosing to standardize it together rather than compete on it. At last month's Hot Interconnects Symposium, engineers from AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI detailed progress on the Optical Compute Interconnect (OCI) standard, an open, silicon-centric specification the six companies jointly launched in March specifically to move data optically inside a single GPU rack.
AMD's answer on the silicon side is Pensando, built into its Helios rack-scale system from the outset; the Pensando Vulcano 800 AI NIC combines three 800Gbps NICs per GPU for up to 2.4Tbps of aggregate bandwidth per accelerator. Nvidia's answer is NVLink and NVSwitch, increasingly extended into the optical domain — the company's Rubin Ultra platform is expected to integrate co-packaged optics directly into scale-up interconnects starting in 2027. Nvidia's SHARP protocol, which pushes gradient-reduction computation directly into switch silicon, shaves an estimated 30-40% off all-reduce latency by eliminating a full network round-trip.
Ethernet vs. InfiniBand
The long-running architectural debate between InfiniBand and Ethernet for AI clusters has tipped decisively toward Ethernet, though InfiniBand retains a durable niche. Dell'Oro data shows Ethernet passing InfiniBand in AI back-end switching market share during early 2026, reaching roughly two-thirds of the market — up from InfiniBand's roughly 80% share as recently as late 2023, a near-total reversal in under three years.
The Ultra Ethernet Consortium (UEC) — an open, Linux Foundation-governed standards effort launched in 2023 by AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, and Microsoft, now including more than 100 member companies — has moved from specification to shipping silicon. The consortium released its 1.0 specification in mid-2025, and Broadcom's Thor Ultra, the industry's first 800G AI Ethernet NIC, ships fully UEC-compliant, giving the standard its first flagship shipping silicon.
Nvidia's Networking Strategy
No company has invested more aggressively in owning the full networking stack than Nvidia. Spectrum-X, its Ethernet-based AI networking platform, pairs Spectrum switch ASICs with BlueField SuperNICs and LinkX optics to deliver what the company claims is up to 1.6x better AI workload performance than off-the-shelf Ethernet. In August, Nvidia extended the platform with Spectrum-XGS Ethernet, a "scale-across" technology explicitly designed to link geographically distributed data centers into a single, unified "giga-scale AI super-factory," with CoreWeave among the first hyperscale operators announced to deploy it.
The Cisco relationship is itself part of Nvidia's networking strategy: Cisco's N9100 data center switch runs on Nvidia's Spectrum-4 ASIC as part of an Nvidia Cloud Partner-compliant reference architecture. Layered onto Nvidia's roughly $4 billion combined investment in optical suppliers Coherent and Lumentum, announced in March, the pattern is consistent: Nvidia is vertically integrated from the accelerator die through the optics that connect it to everything else.
Optical Networking & Silicon Photonics
As AI clusters scale toward gigawatt power envelopes, copper interconnects run out of reach at the distances and bandwidths required, forcing a transition to optical links even within a single rack. Co-packaged optics (CPO), integrating optical engines directly into the switch package rather than using pluggable transceivers, is the leading architectural answer. Broadcom's third-generation CPO switch, Tomahawk 6 "Davisson," shipped with 102.4 terabits per second of optically enabled switching capacity.
The capital and M&A activity in optics has been extraordinary. Beyond Nvidia's Coherent/Lumentum investment, Marvell paid up to $5.5 billion for silicon-photonics startup Celestial AI. Private optics startups have collectively raised roughly $2.5 billion, with Ayar Labs closing a $500 million Series E at a $3.75 billion valuation. Nokia's work on Linear Pluggable Optics addresses a specific, underappreciated problem: 1.6T optical modules consume nearly double the power of 800G modules, potentially adding a full kilowatt to a single 64-port switch.
Cloud & Hyperscaler Infrastructure
Hyperscaler capital expenditure is increasingly a networking capex story as much as a GPU capex story. AWS's approval to build a cable landing station in West Cork, Ireland — supporting the new Fastnet transatlantic cable connecting Ireland to Maryland — is a direct, physical instance of the "scale-across" networking challenge: hyperscalers now need to move AI training and inference traffic between continents with the same urgency they once reserved for connecting racks within a single building.
Neocloud operators are becoming a networking growth market in their own right: Green Mountain, the Norwegian data center operator, secured a neocloud customer covering 14 megawatts at its LON-East campus in London. Cisco's broader AI infrastructure push, meanwhile, extends into Kubernetes-native territory: the company's Isovalent platform has been validated for inference workloads on AI PODs, enabling high-performance Kubernetes networking as part of its "Secure AI Factory" reference architecture.
Enterprise Networking Impact
While hyperscalers and frontier labs dominate the AI networking headlines, the effects are cascading into mainstream enterprise IT on a lag. Arista's Etherlink AI fabric customer base has grown from four or five hyperscaler deployments in 2024 to more than 100 cumulative customers today, spanning cloud titans, frontier AI labs, and neoclouds — direct evidence that AI networking demand is broadening beyond the handful of trillion-dollar cloud platforms into a wider enterprise tier. This is creating a market bifurcation: hyperscalers building custom, vertically optimized fabrics at the frontier of what's technically possible, while a broader enterprise tier adopts increasingly standardized, UEC-compliant Ethernet gear as the pragmatic default.
Security & Resilience
AI infrastructure's growing centrality has made the network layer an increasingly explicit target. The clearest example remains the FCC's drafted rule, covered in a prior edition of this series, that would restrict imports of new Chinese-made optical transceivers over concerns they could be used to steal data, install malware, or disrupt data center operations — treating the optical layer of AI infrastructure with the same national-security scrutiny previously reserved for telecom carrier equipment.
Cisco's "Secure AI Factory" architecture, which folds Hypershield security directly into its Nexus switch line via AMD Pensando DPUs, reflects the industry's broader response: enforcing zone-based segmentation and stateful firewalling at wire speed within the fabric itself. As AI clusters scale toward "scale-across" architectures, the security perimeter necessarily expands with them.
M&A, Funding & Competitive Landscape
Capital allocation in 2026 has made networking's new strategic status explicit in dollar terms. Nvidia's roughly $4 billion combined investment in Coherent and Lumentum secured dual-vendor silicon photonics supply; Marvell's up to $5.5 billion purchase of Celestial AI added competing scale-up optical capability. On the systems side, Cisco and Nvidia's expanded partnership shows two of the industry's largest players choosing deep technical partnership over head-to-head competition in switching silicon.
On the venture side, Ayar Labs' $500 million Series E at a $3.75 billion valuation remains the standout independent optics bet, while Celestica's decision to contribute its Broadcom Tomahawk 6-based DS6000/DS6001 switch designs to the Open Compute Project signals continued industry commitment to keeping at least part of the AI networking supply chain open and multi-vendor.
Three dynamics define networking's new strategic weight in AI infrastructure, and each has a distinct implication for how executives should think about the next 12-18 months of buildout.
First, the AI networking market's own size estimates keep getting revised upward faster than almost any other infrastructure category this year. Dell'Oro raising its forecast toward a roughly $1 trillion cumulative back-end switch opportunity mirrors Cisco's AI order book nearly quadrupling year over year. Executives building multi-year AI infrastructure plans should treat networking capacity and component availability as a genuine, independently binding constraint.
Second, the industry's biggest rivals keep choosing collaboration over competition specifically at the interconnect standards layer. AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI co-authoring the OCI standard suggests the physics problem at the heart of AI interconnect is now bigger than any single company's competitive roadmap. That collaborative posture may not last — as the technology matures, the incentive to vertically integrate and lock in proprietary advantage will only grow.
Third, optical networking has completed its transition from a component category to a strategic chokepoint with real geopolitical weight attached. A $4 billion Nvidia investment, a $5.5 billion Marvell acquisition, an FCC-drafted import restriction on Chinese transceivers, and a new transatlantic cable landing station approval all landing within the same year signals that photonics and physical interconnect infrastructure are now viewed with the same strategic seriousness as advanced chip manufacturing.
What to Watch Next
- Cisco's fiscal 2027 AI order guidance — the clearest near-term signal of whether the fiscal 2026 order boom represents a sustainable growth curve or a one-time catch-up cycle.
- Dell'Oro's next forecast revision — another upward revision would reinforce that AI networking demand keeps outrunning even aggressive projections.
- Whether Nvidia's Rubin Ultra platform delivers on integrated co-packaged optics for scale-up interconnects starting in 2027 — the first real test of embedding photonics directly into the GPU package.
- The FCC's optical transceiver import rule — whether it's finalized before year-end 2026, with direct implications for Coherent, Lumentum, and Chinese suppliers Innolight and Eoptolink.
- Additional cable landing station and subsea infrastructure approvals following AWS's West Cork/Fastnet project.
- Whether Arista, Cisco, and HPE's next earnings cycles show continued AI-networking customer broadening or renewed concentration in a handful of hyperscaler mega-deployments.
The CODEW Stat
Cisco's AI infrastructure orders from hyperscale customers quadrupled in a single fiscal year, from just over $2 billion to $9.3 billion — remarkable growth for a company that manufactures no GPUs at all, and the clearest evidence yet that the network, not just the chip, has become a primary line item in AI infrastructure spending.
Reviewed by Erwin Castro
on
Wednesday, September 02, 2026
Rating: