The NVLink Alternatives: Genuine Disruption or VC-Funded Pipe Dreams?

Nvidia has built a formidable moat around its AI empire not just with GPUs, but with NVLink Fusion, the interconnect technology that allows hundreds of GPUs to function as a single accelerator. As *The Register* first reported, two challengers are now attempting to break that lock-in by offering open alternatives, though they are attacking the problem from opposite ends of the hardware stack.

**The Architecture Play: Cornelis Networks** Reporting from *The Register* notes that Cornelis Networks, an Intel spin-off from 2020, is positioning itself as the standard-bearer for open scale-up networking. At the AI Infra Summit, the company introduced the Active Compute Fabric (ACF). This open architecture aims to integrate programmable compute directly into the network fabric, moving beyond the company's Omni-Path roots used in supercomputers like Lynx and Trinity.

Cornelis is targeting the massive inefficiency of GPU idle time. The company claims that in a 100,000 GPU system, roughly half of all GPU hours are spent waiting for data—a waste they value at $1.68 billion annually and 500 GWh of power. To solve this, Cornelis is launching CN6000-series switches and NICs capable of 800 Gbps. Their ACF architecture seeks to offload functions like KV cache offloading, MoE expert dispatch, and in-fabric checkpointing to the network, mirroring the collective acceleration seen in Nvidia's NVSwitch ASICs.

**The Physical Play: Delos Data** While Cornelis focuses on the architecture, Delos Data—founded by former Intel and Barefoot Networks executives—is targeting the physical layer. According to *The Register*, Delos is marketing a "Nonstop AI" portfolio of network reference designs.

Their most aggressive claim is an I/O die capable of over 30 Tbps of aggregate bandwidth (approximately 8 TB/s in either direction). This would more than double the 3.6 TB/s interconnect bandwidth found in current accelerators from AMD and Nvidia. However, the viability of this hardware depends entirely on integration; these I/O dies must be integrated directly into the accelerator package, a process requiring intense co-design.

**The Ecosystem Battle** The broader industry is already attempting to pivot. *The Register* reports that AMD has embraced the Ultra Accelerator Link (UALink) protocol, utilizing custom I/O dies on the MI455X connected via Broadcom’s 102.4 Tbps Tomahawk 6-based switches.

Unlike Nvidia, which requires customers to use NVSwitches for scale-up networking, Delos claims its chiplets are protocol agnostic. Cornelis is similarly pushing for a "least common denominator" through the Ultra Ethernet protocol and Ethernet for Scale-Up Networking. Whether these open standards can actually lure customers away from Nvidia's integrated ecosystem remains the trillion-dollar question.