Startups Challenge Nvidia’s Dominant AI Interconnect Standard

Open Standards Threaten Proprietary Lock-In

Two emerging companies, Cornelis Networks and Delos Data, have formally entered the competitive arena of high-speed AI networking. They aim to provide open alternatives to Nvidia’s proprietary NVLink technology. This move occurred during the recent AI Infra Summit, where both firms unveiled their new scale-up fabric solutions. Their goal is to help data centers connect thousands of GPUs efficiently without relying on a single vendor’s ecosystem.

Nvidia has long controlled this space through its tightly integrated hardware and software stack. The company recently expanded its reach with NVLink Fusion, a technology designed to allow third-party chips to join its high-bandwidth network. However, this expansion only deepens Nvidia’s influence over the industry. Rival vendors see an urgent need to offer distinct pathways for scaling artificial intelligence workloads. They argue that open standards will prevent lock-in and improve performance for massive model training.

The core of the competition lies in how different processors communicate with each other. Scale-up fabrics act as the nervous system for AI clusters, enabling rapid data exchange between chips. Without fast interconnects, even the most powerful GPUs sit idle waiting for information. Delos Data and Cornelis Networks propose designs that prioritize openness and flexibility. They believe their architectures can outperform closed systems by reducing latency and increasing throughput.

Can New Entrants Displace Market Leaders?

Cornelis Networks, which spun off from Intel, brings significant engineering pedigree to the table. The company focuses on creating efficient switch fabrics that can handle the massive data demands of modern AI models. Delos Data approaches the problem from a different angle, emphasizing modular design and ease of integration. Both companies presented their technologies to a crowd of infrastructure leaders eager for alternatives. They hope to prove that their solutions are not just theoretical concepts but viable commercial products ready for deployment.

Challenging Nvidia is a daunting task. The chip giant has built a moat around its ecosystem that includes software libraries, drivers, and hardware compatibility. New entrants must convince customers that switching costs are worth the potential gains. Analysts note that while the demand for AI compute power continues to surge, the supply chain remains fragile. Any disruption in component availability could favor flexible, open architectures.

The timing of these announcements is critical. As large language models grow in size, the need for connecting more chips becomes paramount. Traditional PCIe links are reaching their bandwidth limits. High-speed interconnects are no longer optional; they are essential for cost-effective AI training. By entering the market now, Delos and Cornelis position themselves to capture early adopters who fear future dependency on Nvidia.

Frequently Asked Questions

Why do startups want to replace NVLink? Companies seek open standards to avoid vendor lock-in and reduce costs. They also aim to improve communication speeds between different types of processing units.

Who are the main competitors mentioned? Delos Data and Cornelis Networks are the primary challengers highlighted in this report. They compete directly against Nvidia’s NVLink Fusion technology.

When did these companies announce their plans? The announcements took place during the AI Infra Summit this week. Both firms officially entered the scale-up networking race at this event.