AI Networking
AI networking covers the high-speed interconnect layer inside AI data centers - the InfiniBand and Ethernet fabrics, switch silicon, optical and copper interconnects, and physical connectors that link GPUs, servers and racks within a facility. The competitive set spans switch and networking-silicon designers, high-speed connectivity semiconductor specialists, contract manufacturers, and the broader accelerated-computing ecosystem that these fabrics serve.
Active institutional funds added a net $13.9 billion in Q2 2026, versus $4.4 billion in Q1 2026 — an inflow 218% larger. 65% of companies in this sub-theme saw increased active positions in Q2 2026.
Overview
AI networking is the set of technologies that move data between GPUs, servers and racks inside an AI cluster, distinct from wide-area or last-mile connectivity. It spans two competing fabric approaches - InfiniBand and high-speed Ethernet - and the silicon, optics and physical connectors that implement them. Nvidia and Broadcom design GPU networking, switch, network-adapter and cable solutions and compete directly with each other and with Cisco Systems in this space. Cisco, Arista Networks and Extreme Networks build data-center and campus switching platforms; Marvell, Credo Technology, Astera Labs, MACOM and Semtech supply high-speed connectivity, interconnect and analog semiconductors used in these fabrics; Nvidia licenses its NVLink Fusion interconnect technology to chip designers including Marvell and Qualcomm to connect custom silicon into its AI infrastructure ecosystem. Amphenol supplies connector and interconnect systems, while Jabil, Sanmina and Foxconn serve as contract manufacturers producing networking hardware for vendors such as Arista Networks. AMD and Intel compete with Nvidia and Broadcom in accelerated computing and networking product categories, and Nokia is expanding into Indium Phosphide optical component production, a technology used in optical interconnects. TSMC serves as the sole or exclusive foundry partner for several interconnect chip designers, including Credo Technology and Astera Labs.
Investment Thesis
Investment thesis details are part of the Pro plan.
Bottleneck Tracker
Bottleneck details are part of the Pro plan.
Key Companies
Related Events
Marvell is buying Lumentum in a $115B all-stock deal to become a one-stop shop for AI photonics. The deal fills gaps in lasers, transceivers, optical circuit switches and indium phosphide fabrication. It follows Marvell's roughly $3.5B Celestial AI purchase in February 2026 and Nvidia's $2B investment in Lumentum eight months ago. Copper reach falls as per-lane speeds pass 400 Gbps, which pushes AI rack-scale builders toward optics. Marvell's market cap is about $180B, so the stock-funded deal is large relative to the company, and it does not close until the first half of 2028.
Google is turning TPUs into a merchant AI-chip business by distributing them through Oracle's cloud, widening its reach beyond its own data centers. Analysts model TPU sales of $9-10 billion in the second half of 2026 and $62-99 billion in 2027, which would make Google a credible Nvidia challenger. External TPU sales carry a roughly 20% operating margin versus about 35% for overall cloud, so the mix shift could dilute Google Cloud margins even as revenue grows near 90%. Anthropic and OpenAI as customers deepen Google's role as supplier to its model rivals, and Broadcom's $63 billion lease vehicle shows capital markets financing TPU buildouts. Nvidia faces more competitive pressure on price-performance, while Oracle gains a differentiated AI offering.
A pluggable optic at 12.8 Tbps would sharply raise bandwidth density for AI cluster cabling, favoring optical module, DSP and laser suppliers who can meet 400G-per-lane and 100 W thermal demands. Nvidia's participation, despite its move to co-packaged optics on its switches, shows pluggables remain strategic for NICs and flexible deployments. The open standard widens the addressable market beyond Cisco and Nvidia but raises thermal, reliability and blast-radius challenges for module and cage makers. Microsoft's use of Cisco's P200 for scale-across networking strengthens Cisco's position against Nvidia Spectrum-XGS and Broadcom Jericho 4. Commercial impact is gradual, tied to 204.8 Tbps switches and Vera Rubin Ultra-class racks later in the decade.
The Vultr order is a concrete proof point that AI-driven demand for networking gear is outstripping supply, prompting HPE to double its component purchase commitments to convert backlog into revenue. HPE raised its FY2027 networking growth guidance to high-teens/low-20% and expects operating margins to expand into the mid-to-high 20% range, signaling networking could become a structurally higher-margin business post-Juniper. The deal also validates AMD's Helios rack-scale platform as a viable alternative to Nvidia-centric AI clusters among neocloud buyers like Vultr. Execution risk remains high: HPE must convert a backlog growing 3.5x faster than revenue into actual shipments and cash flow. Investors reacted positively, with shares up about 4%, reflecting optimism about HPE's AI infrastructure positioning relative to rivals.
Nvidia's strategy of licensing NVLink and building an IP toll on the custom AI chip ecosystem is a structural market event that changes the competitive calculus for every company designing bespoke AI accelerators. Hyperscalers and cloud providers building custom silicon — Google's TPUs, Amazon's Trainium, Microsoft's Maia — face the risk of their supply chains being partially encircled by Nvidia's IP licensing requirements if NVLink becomes a de facto interconnect standard. MediaTek, as the first major partner, gains access to Nvidia's ecosystem while helping Nvidia extend its toll-collection reach into mobile and edge AI segments. Companies like AMD and Intel that compete at the chip level now also face Nvidia as an IP licensor embedded in their customers' infrastructure decisions. This durable IP moat, if it solidifies, could compress margins across the custom silicon industry for years.