AI networking startups race to replace Nvidia's NVLink

Cornelis Networks and Delos Data are entering the AI scale-up networking market with open, protocol-agnostic alternatives to Nvidia's NVLink ecosystem, targeting larger GPU clusters and reduced communications bottlenecks. Cornelis introduced its Active Compute Fabric architecture and is preparing 800 Gbps CN6000 switches and NICs, while Delos outlined 30+ Tbps I/O chiplets, 10+ Tbps near-packaged optics, and 400+ Gbps NICs. Cornelis raised about $205 million and Delos has raised more than $100 million, signaling investor appetite for challengers to Nvidia's AI interconnect dominance.

Analysis

The near-term monetization accrues to AVGO, not the venture-backed challengers: open scale-up protocols are currently being implemented over Ethernet, increasing the content value of merchant switch silicon before dedicated alternatives reach volume production. This also improves AMD's rack-level total-cost-of-ownership proposition, but only if its software stack can translate fabric openness into materially higher effective accelerator utilization; interoperability alone does not overcome CUDA ecosystem inertia. NVDA's valuation should be largely insulated over the next 1-3 quarters because hyperscalers buy validated systems and uptime, not theoretical bandwidth, and its integrated switch/GPU roadmap remains the reference implementation.

The more consequential 6-18 month risk to NVDA is margin rather than GPU unit demand. A credible open fabric with in-network collective offload could reduce the proprietary-network premium embedded in rack-scale systems, particularly for hyperscalers designing custom accelerators; that shifts bargaining power toward switch ASIC, optics, and packaging suppliers. The cited utilization savings should not be capitalized: cluster idle time also reflects model parallelism, storage, software, and scheduling constraints, so network-only improvement is likely far below vendor modeling until independent deployments demonstrate end-to-end training throughput gains.

Optics is the underappreciated second-order beneficiary as cluster topology moves beyond rack boundaries. Copper remains favored inside racks on power and serviceability grounds, but row-scale architectures raise the optical-content opportunity for COHR and LITE; adoption timing depends on failure-rate qualification and whether customers accept near-packaged rather than fully co-packaged designs. CSCO has strategic optionality through Ethernet switching and software, but faces greater pricing risk than AVGO if open standards commoditize system control planes.

Consensus is likely overestimating the immediacy of a competitive threat to NVDA and underestimating qualification friction. Protocol standards without broadly available endpoint silicon, reference systems, and production software create customer choice but not necessarily deployable substitution; the key catalyst is a named hyperscaler design win with independently measured collective-latency and fault-recovery results.

Key Decisions for Investors

• Maintain/establish long AVGO versus short CSCO over the next 3-6 months: merchant Ethernet scale-up deployments should favor AVGO's switching content, while CSCO's differentiated software value is less proven in hyperscale AI fabrics. Reassess if a major cloud customer discloses a non-AVGO switch design win or AVGO signals AI-networking inventory digestion.
• Use AMD as a tactical long only into evidence of production MI-series cluster deployments, not protocol announcements. A positive catalyst is a hyperscaler disclosing rack-scale performance or expanded accelerator orders; falsify on another material delay to accelerator availability, software readiness, or weak AI revenue guidance.
• Keep NVDA core exposure but hedge 6-18 month network-premium risk with a modest NVDA/AMD relative-value position rather than an outright NVDA short. The spread should tighten only after open-fabric systems demonstrate comparable uptime and training throughput at scale; NVDA's next platform guidance and gross-margin outlook are the near-term invalidation points.
• Create an alert basket in COHR and LITE for row-scale optical adoption disclosures. Do not initiate solely on architecture claims; require evidence of qualified near-packaged optics, customer sampling, and a measurable increase in datacenter optical revenue guidance before sizing exposure.