Cornelis takes aim at AI bottlenecks with programmable network fabric.
Networks built for AI and HPC workloads have mostly focused on moving data faster between endpoints. That leaves compute idle while packets travel, a bottleneck vendors have tried to address mainly by adding bandwidth rather than changing what the fabric itself does.
Cornelis Networks is taking a different approach. The company raised $205 million to fund its expansion into scale-up networking and introduced Active Compute Fabric, an architecture that puts programmable compute inside the network rather than using it only to connect endpoints.
Cornelis emerged from Intel in 2020, when the team behind Intel’s Omni-Path architecture spun out as an independent company. Omni-Path is a proprietary network connectivity protocol that provides a competitive alternative to InfiniBand and Ethernet. In 2025, the company shipped its first generation of standalone products, the CN5000 platform. Cornelis Networks has since continued to develop its own SuperNICs, switches and software built for AI and HPC workloads.
“I want Cornelis to be synonymous with game-changing technology that improves AI efficiency and eases some of the burdens that we’re facing in the world today about keeping it funded and responsibly so,” Lisa Spelman, CEO of Cornelis Networks, told Network World.
The ‘path’ forward for enabling AI
Cornelis built its first generation of independent products around Omni-Path exclusively. That architecture requires both Cornelis SuperNICs and switches end to end. The company’s second generation, the 800 gigabit CN6000, adds Ethernet support without giving up the proprietary architecture that differentiates its products.
Spelman describes the CN6000 SuperNIC as multimode silicon capable of running either protocol. “You can use it in Ethernet mode, like RoCEv2 Ethernet mode, or you can use it as OmniPath end to end with the OmniPath switch,” she explained. “It’s one piece of silicon, both modes contained in it.”
In Ethernet mode, the CN6000 can pair with switches from other vendors, including Broadcom’s Tomahawk line. Traffic sent over Ethernet still passes through an internal encapsulation layer built around Omni-Path. Spelman said that internal architecture preserves congestion management, credit-based flow control, and packet spraying, aimed at low latency and high message rates even when the wire protocol is Ethernet rather than Omni-Path.
“We have the Ethernet protocol inside the chip, it’s an encapsulation layer, and it races around the chip like an Omni-Path packet, and then it comes back out and is retranslated back out to Ethernet,” Spelman said. “So, you still get some of the benefits of the Omni-Path native architecture, even though on the wire it’s Ethernet.”
Putting compute inside the network
Active Compute Fabric extends Cornelis Networks’ combination of proprietary performance and open standards into scale-up networking. It adds a new element—compute embedded directly into networking silicon—in an approach that differs from the functions found in a typical smart NIC or DPU.
“If you think of a standard NIC, there might be a DPU or a smart NIC that does offload of network overhead,” Spelman said. “We’re taking it a step further and putting real compute, think RISC-V cores, into both the NIC and the switch that are capable of AI acceleration.”
Cornelis does not design its own RISC-V cores. The company licenses the IP from a third-party RISC-V supplier and integrates it alongside its own ASIC designs. That approach is consistent with a fabless model, in which Cornelis builds only the components that differentiate its products.
The compute embedded in the fabric targets functions such as KV cache acceleration and mixture of experts (MoE) routing. Spelman said that work can run inside the network rather than occupying GPU cycles, with inference as the clearest opportunity for those capabilities.
“Inference acceleration is absolutely a sweet spot for us,” Spelman said.
She said Cornelis continues to serve training workloads as well, with demand extending beyond hyperscalers and frontier labs to enterprise, government, and academic customers building out their own AI infrastructure.
“The biggest thing we’re trying to solve is we’re giving you your GPUs back,” she said. “You’ve bought all these GPUs, and they’re being used at about a 50% rate, some people at 42, some at 48, some at 54. At that utilization, with a better, higher performing active network that has compute in the network, you can drive that GPU utilization up five points, 10 points, and really push your capabilities.”
The scale-up push extends Cornelis beyond its scale-out heritage and into rack-scale interconnects. Spelman said that area is aimed squarely at accelerators outside Nvidia’s NVLink ecosystem. Active Compute Fabric for scale-up relies on open standards including UALink and ESUN rather than a proprietary link.
“Our heritage has been in scale-out, but we see a huge opportunity in scale-up for all of the compute that doesn’t want to be on the NVLink ecosystem, isn’t looking to get so tied to Nvidia, wants options but needs that incredibly high performance and industry standards,” Spelman said.
