Rival networking vendors are rushing to introduce alternative interconnects and switches as Nvidia expands its influence through NVLink Fusion technology. Delos Data and Cornelis Networks officially entered the scale-up networking race at the AI Infra Summit. Scale-up fabrics like NVLink allow massive groups of GPUs to behave as one enormous AI accelerator. Competitors such as AMD are adopting emerging protocols like Ultra Accelerator Link to catch up to these capabilities.
Protocols today largely rely on standard Ethernet switches. AMD uses Broadcom Tomahawk 6 switches connected to custom I/O dies on the MI455X. Purpose-built UALink switches and physical interconnects remain rare, but companies like Cornelis and Delos plan to change that dynamic. Both firms approach the challenge through standardization, software optimization, and physical hardware.
Cornelis Networks introduced the Active Compute Fabric at the AI Infra event. This open architecture for scale-up and scale-out networking integrates programmable compute directly into the fabric. Spun out from Intel in 2020, Cornelis originally designed its Omni-Path technology for high-performance supercomputers. The company is now launching 800 Gbps-capable CN6000-series switches and NICs to bring its technology to a broader audience.
Active Compute Fabric establishes a baseline for in-network compute capabilities across diverse hardware. Nvidia uses NVSwitch ASICs with support for SHARP to offload in-network collectives, which cuts communication overhead. Cornelis aims to accelerate functions including KV cache offloading, Mixture-of-Experts expert dispatch, message passing interface offload, and in-fabric checkpointing. The company stated that communications overhead and synchronization can leave expensive accelerators underutilized.
Cornelis modeling suggests that in a 100,000 GPU system, half of all GPU hours are spent waiting for data. This amounts to roughly $1.68 billion in wasted capacity and 500 GWh of power annually. While these manufacturer estimates require caution, reclaiming idle GPU time presents substantial potential savings.
Delos Data tackles the physical layer with network reference designs marketed under its Nonstop AI portfolio. Founded by former Barefoot Intel and Broadcom executives, the startup provides system-level blueprints for speeding up data movement beyond a single rack. The data interface will be offered in three distinct form factors.
The first form factor is an I/O die capable of more than 30 Tbps of aggregate bandwidth or about 8 TB/s in either direction. This bandwidth doubles the capacity of the latest accelerators from Nvidia and AMD, which cap out at 3.6 TB/s. Successful deployment depends on integration directly into the accelerator package through tight co-design.
Delos maintains protocol agnosticism rather than trapping users in a walled garden like Nvidia does with NVLink Fusion. Its chiplets support ESUN, UALink, or other protocols, allowing hyperscalers to license the design directly. This protocol agnosticism also applies to the company near-packaged optics technology, which integrates a 10-plus Tbps data interface with optics engines.
Rack-scale systems largely rely on copper interconnects due to power limits, but larger clusters scaling past 576 GPUs require optics. Near-packaged optics use copper within the rack and user-serviceable optical modules between racks to limit failure impact. Delos is also developing a 400-plus Gbps NIC for scale-out training clusters and storage networks.
Investors are eager for competition in the scale-up market. Cornelis announced approximately $205 million in funding to support its next-generation products. Delos Data announced that it has raised more than $100 million with backing from Matrix, Playground, and Socratic Partners.
