The problem the money is chasing
A graphics processing unit (GPU) training or running a large model is not always busy. A meaningful share of its time goes to waiting: waiting for the next batch of data to travel across the network from another machine in the cluster. Cornelis Networks, a company that spun out of Intel in 2020, has raised $205 million in a round led by IAG Capital Partners to sell a fix for exactly that stall, and it is betting the fix belongs inside the network itself rather than on the chip.
What Active Compute Fabric actually does
Cornelis calls its answer the Active Compute Fabric. The idea is to stop treating the network as a passive pipe that only moves bytes and instead give it enough of its own processing power to act on data as it passes through. Concretely, that means putting real compute, RISC-V processor cores capable of artificial intelligence (AI) acceleration, into both the network interface card (NIC, the hardware that connects a server to the network) and the switch itself. RISC-V is an open instruction set architecture that anyone can build chips around without paying a licensing fee, which is part of why a smaller networking vendor can credibly add general-purpose compute to its hardware rather than depending on a chip design it would have to license from someone else, potentially a competitor.
Cornelis chief executive Lisa Spelman described the approach directly: "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." The effect, if it works as claimed, is that a chip can process and send information at the same time instead of sitting idle while it waits for a fresh batch to arrive over the wire.
Two product generations, two different bets
Cornelis already has hardware in the market. The CN5000, its first generation of standalone products, started shipping in 2025 and is built entirely on Omni-Path, an interconnect architecture that requires Cornelis's own SuperNICs and switches end to end, a closed loop by design.
The CN6000, its second generation, changes that. It runs at 800 gigabits per second and uses multimode silicon that speaks both Ethernet and Omni-Path. That lets a customer pair CN6000 gear with third-party switches, including Broadcom's Tomahawk line, while running in Ethernet mode, and Cornelis says it preserves Omni-Path's performance benefits underneath through an internal encapsulation layer even when the outward-facing protocol is standard Ethernet. In practice, that is Cornelis hedging: sell a proprietary advantage to customers who want it, but do not force them to rip out their existing Ethernet switches to get it.
Why this is a story about Nvidia
Nvidia does not just sell GPUs; it sells the networking gear that connects them too, and that combination is a large part of why data center operators find it easier to buy the whole Nvidia stack rather than mix vendors. Nvidia hardware can technically run on other companies' networking fabrics, but it is optimized for Nvidia's own, which makes switching away from Nvidia networking harder than it looks on a spec sheet.
Cornelis is one of a small wave of infrastructure startups trying to pry that bundle apart by selling networking that works across GPU and accelerator vendors rather than locking a buyer into one. An open fabric is only a compelling pitch, though, if it performs competitively, and Cornelis's argument rests on the idea that moving compute into the network, not just faster wires, is what closes the performance gap with a vertically integrated rival.
What it means in practice
For data center operators, the near-term decision is which generation to evaluate: the CN5000 for its most mature, single-vendor Omni-Path performance, or the CN6000 for its ability to slot into an existing Ethernet-based cluster without a full fabric replacement. For enterprise buyers watching AI infrastructure costs, the underlying signal is that network idle time, not just raw GPU count, is becoming a line item companies are willing to fund a $205 million round to fix. Whether Active Compute Fabric delivers on that in production workloads, rather than in a funding announcement, is the thing to watch as Cornelis's newer hardware reaches wider availability later in 2026.
