Networking Watch: The Pure-Play AI Networking Winner, The Custom Silicon and Switching Powerhouse & Emerging Technologies

The CODEW Networking Watch | August 25, 2026

The networking layer of the AI infrastructure buildout is emerging as the hottest trade in technology. As AI clusters scale from thousands to hundreds of thousands of GPUs, the network fabric — not compute — is becoming the primary determinant of cluster performance and the next critical bottleneck. This Watch examines how switching silicon, optical interconnects, and new architectures are reshaping the competitive landscape, with Arista Networks and Broadcom leading the charge while Cisco, NVIDIA, and a wave of optical startups scramble to catch up.

The Lead

The Network Layer Is Now the Hottest Trade in AI

What happened: The networking layer of the AI infrastructure buildout has posted numbers that demand serious analytical attention. Arista Networks and Broadcom have separated from the pack in a way the market has not fully priced. JPMorgan's equity research team identifies AI as the primary catalyst for above-trend earnings growth, with the investment cycle extending beyond early infrastructure buildout into broad enterprise adoption. Capital rotation has shifted from pure-play GPU names to the infrastructure that ties compute together: switching, custom silicon, optics, and fabric management software. Arista and Broadcom are the two largest beneficiaries of that rotation.

Why it matters: The common framing for AI infrastructure spending focuses on GPU compute, but that framing is incomplete. Every GPU cluster, every hyperscaler data center campus, every AI training run requires switching silicon, custom accelerators, and high-speed fabric to tie compute nodes together. As AI clusters grow from thousands to hundreds of thousands of GPUs, the network is becoming a primary determinant of cluster efficiency. Dell'Oro Group's Sameh Boujelbene put it bluntly: "A company can spend billions on GPUs, but if the fabric cannot deliver predictable bandwidth and low latency, they did not buy an AI supercomputer; they bought an expensive collection of stranded chips".

Who is affected: Networking vendors (Arista, Broadcom, Cisco, NVIDIA), optical component suppliers, hyperscalers building AI infrastructure, and enterprise AI buyers facing network-related costs.

What to watch next: Whether the networking trade continues to outperform as hyperscalers pour billions into switching and optical infrastructure, and whether emerging technologies like co-packaged optics (CPO) and hollow-core fiber reshape the competitive landscape.

The Scale-Up, Scale-Out, Scale-Across Architecture

AI infrastructure grows along three complementary dimensions, each placing distinct demands on the network:

Scale-Up: Packs more compute per server or rack, often via high-bandwidth intra-node links like NVIDIA's NVLink. NVIDIA classifies this as connecting GPUs within a single rack.

Scale-Out: Adds more racks and combines their resources to run larger models or concurrent jobs. This is where Ethernet switching dominates, with vendors like Cisco, Broadcom, and NVIDIA converging on 51.2-102.4 Tbps-class switching silicon. NVIDIA's Spectrum-X Ethernet platform promises up to 1.6x higher AI networking performance.

Scale-Across: Connects multiple data centers over optical networks, enabling clusters to work together across regions. Broadcom estimates the Scale-Across TAM at $15 billion to $20 billion by 2030. Hyperscalers are increasingly building and controlling these connections themselves, bypassing telecom carriers.

The Scale-Across dimension is particularly significant because the primary constraint for hyperscalers is no longer computing capacity but access to electrical power. As AI infrastructure expands beyond individual campuses in search of available power, geography is becoming a first-order constraint on AI scaling. Relativity Networks calls this shift the "AI Geography Era".

Vendor Landscape: The Leaders and Challengers

Arista Networks: The Pure-Play AI Networking Winner

Arista Networks delivered its first $3 billion quarter in Q2 2026, with revenue of $3.036 billion — exceeding guidance of $2.8 billion and Wall Street expectations of $2.83 billion. The company maintained its target of at least $3.5 billion in AI fabrics revenue for the year and reported that its Etherlink AI platforms have now passed 100 cumulative customers. Arista has announced the 7060XE7 Series range of 1.6 Terabit networking platforms designed as the foundation for rack-scale AI infrastructure.

Arista's CEO recently stated: "Arista just proved the AI networking buildout is still accelerating, not slowing, with faster growth, margin resilience where it counts, and a raised outlook to back it up".

Broadcom: The Custom Silicon and Switching Powerhouse

Broadcom has become the "leading AI supplier in custom ASIC (XPU) and networking". BMO's bullish thesis is anchored in Broadcom's increasingly important role in the AI infrastructure buildout. The company is seeking to raise $70-$100 billion in debt to expand custom AI chip capacity, supporting AI firms like Anthropic. Broadcom has partnered with Anthropic and others to pour $35 billion into expanding Anthropic's computing infrastructure through Broadcom's custom chips and networking solutions.

Broadcom provides switching chips covering scale-up, scale-out, and scale-across scenarios, and has introduced AI network interface cards to address surging GPU-to-GPU communication demand. Hasan Siraj, Broadcom's VP of Product, emphasized: "When you're training these models or doing distributed inference, you can't fit the model on a single XPU. You need to connect tens, hundreds, thousands, or even 500 million units. That's the network".

Notably, Broadcom's networking NICs use Ethernet rather than NVIDIA's InfiniBand. Siraj stated: "I think there is now near-consensus in the industry that for most use cases, Ethernet should be the de facto standard".

NVIDIA: From GPU Supplier to Full-Stack Networking Player

NVIDIA has introduced Spectrum-X Ethernet, a transformative networking platform engineered specifically for giga-scale AI workloads. Traditional Ethernet falters under the synchronized, high-bandwidth demands of AI training. Spectrum-X addresses these bottlenecks with hardware-accelerated adaptive routing, congestion control, and load balancing.

The company's Spectrum-X Ethernet Photonics (co-packaged optics) switch entered full production on August 14. It is the world's first production 200G/lane co-packaged optics switch system, delivering 5x network power efficiency improvement, 5x AI application uptime improvement, and 10x mean time between failures compared to traditional pluggable optics.

At Hot Chips 2026, NVIDIA announced full production of Groq 3 LPX racks, which can achieve 4x performance over current solutions in specific conditions. Each LPX rack integrates 256 LPUs with 128GB of on-chip SRAM and 640 TB/s bandwidth. The racks can be deployed at massive scale and treated as a single compute unit, working alongside Vera Rubin systems. NVIDIA also unveiled a Multi-Rail network architecture through Spectrum-X Ethernet that can connect up to 512,000 GPUs in a single system — far exceeding the 2,048 GPU limit of previous two-layer switch topologies.

Cisco: The Incumbent Responds

Cisco is responding to the AI networking challenge with its Silicon One G300 chip, designed for scale-out within the data center. The G300 class exposes a large number of high-speed channels and emphasizes congestion avoidance and burst tolerance. Aggregate bandwidth reaches 102.4 Tbps via 512 lanes at 200 Gbps per lane. The new devices are expected to ship broadly before the end of the year. Cisco recently posted record results on hyperscaler AI infrastructure orders.

Emerging Technologies: The Next Bottlenecks and Breakthroughs

Co-Packaged Optics (CPO): The Interconnect Revolution

SK hynix and global researchers have published a CPO roadmap in Nature Electronics, positioning optical interconnects as the key to overcoming AI bandwidth bottlenecks. The paper sets aggressive targets: single-node bandwidth exceeding 100 Tb/s, energy consumption below 1 pJ/bit, and chip-to-chip latency under 10 nanoseconds. SK hynix is moving beyond HBM innovation to help define the architecture of next-generation AI infrastructure at the system level.

TrendForce data shows CPO penetration in AI data center optical modules at approximately 0.5% in 2026, projected to surge to 35% by 2030. NVIDIA's production CPO switch is a major milestone, with analysts identifying 2026 as the industry's inflection point for CPO commercialization.

Optical Startups Attract Massive Capital

Lumilens emerged from stealth on August 6 after just two years, with its first product certified and shipping to a hyperscaler under a multi-billion-dollar agreement. The company raised over $700 million in new funding at a $5.55 billion valuation, with cumulative funding exceeding $900 million. Lumilens addresses both Scale-out (800G, 1.6T, and above pluggable transceivers) and Scale-up (NPO and CPO silicon) networks. The company's founder and CEO, Ankur Singla, stated: "The bottleneck in AI has shifted from how many GPUs you can buy to how many GPUs you can connect".

Relativity Networks raised a $22 million SAFE investment — more than double its target — from investors including Rhapsody Venture Partners, Bell Ventures Inc., and Faster Than Glass. The company secured a $40 million follow-on order from a hyperscaler after successful testing of its ChronoCore hollow-core fiber technology. Relativity Networks produced its highest-density hollow-core fiber cable to date: 24 fibers in a single 10-millimeter cable. The technology addresses the "Propagation Tax" — the unavoidable latency penalty of distance — enabling AI infrastructure to scale across geographically distributed campuses.

Market Signals

Several patterns emerged from this week's networking developments:

The network is becoming the primary determinant of AI cluster performance. Dell'Oro Group estimates nearly 40% CAGR in AI-driven front-end network demand. Data Center Networking is the largest engine in networking spending, with 43% of buyers planning significant increases in infrastructure spending.

Ethernet is winning the AI networking debate. Industry consensus is shifting toward Ethernet as the de facto standard for most AI use cases. NVIDIA's Spectrum-X Ethernet platform, Broadcom's Ethernet-based NICs, and Arista's Etherlink AI platforms all reflect this trend.

Optical interconnects are entering commercial production. NVIDIA's CPO switch production, SK hynix's Nature Electronics roadmap, and Lumilens' multi-billion-dollar hyperscaler agreements signal that optical technology is moving from lab to factory.

Geography is becoming a first-order constraint. The shift to distributed AI infrastructure — driven by power availability — is creating demand for new technologies like hollow-core fiber that can bridge data centers with low latency.

📊 THE CODEW STAT

$15-20 billion — Projected Scale-Across (inter-data center AI connectivity) TAM by 2030

102.4 Tbps — Aggregate bandwidth of next-generation switching silicon via 512 lanes at 200 Gbps

35% — Projected CPO penetration in AI data center optical modules by 2030 (up from 0.5% in 2026)

Three Networking Signals

Signal 1: The Networking Trade Is Outperforming GPU Pure-Plays

Arista's $3 billion quarter and Broadcom's AI networking momentum suggest capital is rotating from GPU pure-plays to the infrastructure that ties compute together. Watch whether this trend accelerates as hyperscalers pour billions into switching and optical infrastructure. The networking layer represents a "pick and shovel" play on the AI buildout with potentially more durable economics.

Signal 2: CPO Commercialization Will Reshape the Supply Chain

NVIDIA's production CPO switch and SK hynix's Nature Electronics roadmap signal that co-packaged optics is moving from lab to factory. Watch for supply chain shifts as optical component suppliers scale to meet hyperscaler demand. The 0.5% to 35% penetration trajectory implies significant opportunities for optical component manufacturers.

Signal 3: Hyperscalers Are Bypassing Telecom Carriers for Scale-Across

Broadcom reports hyperscalers are increasingly building and controlling data center interconnects themselves. Watch whether this trend accelerates and what it means for telecom carriers' AI revenue aspirations. If hyperscalers build their own long-haul optical networks, telecom carriers could lose a significant growth opportunity.

THE CODEW TAKE

The most important signal from this week's networking developments is that the AI infrastructure buildout is shifting decisively from compute to connectivity. The GPU is no longer the constraint — the network that connects GPUs is. As AI clusters scale from thousands to hundreds of thousands of GPUs, the fabric's ability to deliver predictable bandwidth and low latency is becoming the primary determinant of cluster efficiency.

Three trends are driving this shift. First, switching silicon is entering a new performance tier. Cisco's 102.4 Tbps G300, Broadcom's Tomahawk 6, and NVIDIA's Spectrum-X are all pushing the boundaries of what Ethernet can deliver. Second, optical interconnects are moving from lab to factory. NVIDIA's CPO switch production and SK hynix's Nature Electronics roadmap signal that 2026 is the inflection point for commercial optical interconnects. Third, geography is becoming a design constraint. As AI infrastructure expands in search of power, technologies like hollow-core fiber that can bridge data centers with low latency are becoming strategically critical.

For enterprise buyers, the implication is clear: the era of buying GPUs and plugging them into standard Ethernet is ending. AI clusters now require purpose-built networking infrastructure optimized for synchronized, high-bandwidth traffic patterns. The vendors that can deliver this infrastructure — Arista, Broadcom, Cisco, NVIDIA — are positioned to capture significant value as the AI buildout continues. The networking trade is not just the hottest trade in AI; it may be the most durable.

Source Attribution

DataCenterKnowledge — AI Data Center Networking: Scale Up, Out, and Across (August 20, 2026)
The Cheap Investor — The Network Layer Is Now the Hottest Trade in AI (August 23, 2026)
Blockchain.News — NVIDIA's Spectrum-X Ethernet Redefines AI Networking (August 25, 2026)
Fiscal.ai — NVIDIA and Marvell Expand NVLink Fusion Partnership (August 24, 2026)
Investing.com — Arista Networks at Rosenblatt AI summit (August 18, 2026)
Yahoo Finance — Arista's (ANET) Record Quarter Reignites The AI Networking Debate (August 10, 2026)
GuruFocus — Broadcom Rebounds as Wall Street Sees 25% AI Upside (August 24, 2026)
Financial Express — Broadcom's $80 billion AI bet (August 23, 2026)
SK hynix Newsroom — SK hynix's CPO roadmap in Nature Electronics (August 20, 2026)
中国电子元件行业协会 — Lumilens emerges from stealth with $900M+ funding (August 24, 2026)
The Fast Mode — Relativity Networks Raises $22M for AI Infrastructure (August 23, 2026)
TechBang — NVIDIA Hot Chips 2026: Gorq 3 LPX and Spectrum-X (August 24, 2026)
讯石光通讯网 — Broadcom: Hyperscalers bypassing telecom carriers for Scale-across (August 21, 2026)
TrendForce — CPO penetration forecast (August 24, 2026)

The CODEW