> ## Content Index > Fetch the complete content index at: https://www.edgewisely.com/llms.txt > Use this file to discover other available public pages before exploring further. # Cornelis Puts Compute Inside the AI Network - URL: https://www.edgewisely.com/cornelis-active-compute-fabric-205m-ai-networking/ - Published: 2026-09-16T01:46:05.000Z - Updated: 2026-09-16T07:38:54.000Z - Description: How a $205 million round and a Qualcomm alliance are being spent on an argument that the 15% of an AI system nobody optimizes determines the value of the other 85%. - Author: John Karpentar - Tags: Chips, Deep Tech How a $205 million round and a Qualcomm alliance are being spent on an argument that the 15% of an AI system nobody optimizes determines the value of the other 85%. **Half the accelerators in a large AI cluster are idle at any given moment, waiting for data. Cornelis Networks just raised $205 million on the claim that the fix is not faster endpoints but a network that stops behaving like a pipe.** The most expensive silicon in the world spends a great deal of its life waiting. In large AI deployments, accelerator utilization commonly sits near half of installed capacity — meaning that of every two GPUs a company finances, provisions, powers and cools, roughly one is doing nothing at any given instant. Not because it is slow. Because the data hasn't arrived. On Monday, Cornelis Networks announced [$205 million led by IAG Capital Partners](https://siliconangle.com/2026/09/14/cornelis-networks-raises-205m-and-scales-up-and-scales-out-with-its-new-active-compute-fabric/?ref=edgewisely.com), a strategic collaboration with Qualcomm, and a product architecture called [Active Compute Fabric](https://www.cornelis.com/technology/active-compute-fabric?ref=edgewisely.com) built around a single unfashionable proposition: the network should compute, not just carry. ## What "active" actually means here Conventional data center networking moves packets from A to B as fast as possible and does nothing else. Cornelis calls these passive pipes, and the company's argument is that every collective operation, every synchronization step and every cache transfer therefore costs accelerator time at both ends of the wire. Active Compute Fabric integrates programmable compute into the network itself, so that work happens while data is in transit. Chief Marketing Officer Brandon Draeger gave SiliconANGLE four concrete examples: assembling KV cache data for disaggregated inference, coordinating expert dispatch for mixture-of-experts models, accelerating collective operations such as AllReduce, and compressing gradients as they move. "The payload does not arrive the way it left," Draeger said — which is the cleanest one-line description of the architecture available. In a collective operation, partial results from thousands of endpoints get combined inside the fabric so that a single reduced result lands at the destination instead of thousands of separate contributions. The reduction happens once, in the path the data was already taking, rather than consuming accelerator cycles at both ends. Cornelis puts the effect at up to a 50% reduction in network traffic, based on pre-production simulations. That number deserves the caveat attached to it: simulated, pre-production, and by the vendor's own admission dependent on model, cluster size and customer stack. Draeger was appropriately careful, saying improvements will vary by workload and deployment. The structural claim underneath is more interesting than the benchmark. Cornelis and Qualcomm both put the network at roughly 15% of the cost of an AI system — and argue it determines how much value you extract from the other 85%. If that ratio is right, networking has been systematically under-designed relative to its leverage, chosen after the accelerator decision rather than alongside it. ## The open-standards play The architecture is built on Ethernet and UALink for scale-up and Ultra Ethernet for scale-out. That choice is the entire competitive strategy. Nvidia's advantage in networking is not that InfiniBand and NVLink are unbeatable on the merits. It is that they arrive pre-integrated with everything else a customer is buying. Nvidia chips can technically run on other fabrics, but the full stack is optimized end to end, and deviating from it means accepting integration risk for a benefit that is hard to quantify in advance. Cornelis, which [spun out of Intel in 2020](https://techcrunch.com/2026/09/14/ai-infrastructure-company-cornelis-raises-205m-to-chip-away-at-nvidias-dominance/?ref=edgewisely.com), is selling the opposite: bring whichever accelerator you want. The bet is that as deployments grow, the cost of being locked into a single vendor's fabric starts to exceed the cost of integrating an open one. That is a bet on customer scale and customer sophistication, and it is the same bet several credible challengers have now placed — [Positron's wager against HBM](https://edgewisely.com/positron-ai-875-million-series-c-lpddr5x-inference/?ref=edgewisely.com) is the memory-side version of the same argument. CEO Lisa Spelman framed it as demand rather than ideology, saying customers want an open alternative that gives them more choice in how they build. That framing is convenient but not wrong. Buyers financing multi-billion-dollar clusters have every incentive to create a second source, and networking is the layer where a second source is most achievable, because the standards bodies already exist. The money goes toward scaling production of the [CN5000 and CN6000](https://www.cornelis.com/products/cn6000?ref=edgewisely.com) switches and deploying Active Compute Fabric. The 400Gbps CN5000 is shipping. The 800Gbps CN6000 is sampling with customers, with broader availability expected in the fourth quarter. ## The Qualcomm signal The Qualcomm collaboration is the part worth reading carefully, because Qualcomm does not need a networking partner for the business it is known for. Draeger described the two companies as being in advanced stages of joint technology evaluation focused on rack-scale inference across scale-up and scale-out environments, with the goal of a system where network and accelerators are designed together from the start. Read that backwards and it says something about Qualcomm's data center ambitions: you do not co-design a rack-scale fabric unless you intend to sell something that fills racks. For Cornelis, the value is validation rather than volume. A startup selling against Nvidia, Cisco and Arista needs a reference architecture that a conservative buyer can point at. For Qualcomm, it is optionality — a credible open fabric to pair with inference silicon, without having to build one. ## Who should care **Anyone operating clusters at scale** should treat the utilization number, not the bandwidth number, as the metric that matters. If accelerators sit near 50% utilization, the marginal dollar spent on better data movement outperforms the marginal dollar spent on more accelerators — and the financing structures now carrying AI infrastructure, which we examined in [AI's Capex Bill Moved to the Credit Desk](https://edgewisely.com/ai-capex-revenue-gap-2026-hyperscalers/?ref=edgewisely.com), make that arithmetic sharper every quarter. **Buyers evaluating Cornelis** should ask for the utilization delta on their own workloads, not the traffic-reduction figure. Traffic reduction is an input. Accelerator utilization is the output anyone actually pays for, and the gap between them is where vendor claims go to die. **Nvidia** loses nothing this quarter. What it loses is inevitability. Each credible open-standards alternative reduces the assumption that the full stack must be bought together — and Nvidia's position at the center of the AI supply chain is already drawing scrutiny, as we covered in [the DOJ's look at the Groq transaction](https://edgewisely.com/doj-nvidia-groq-antitrust-probe-reverse-acquihire/?ref=edgewisely.com). **Everyone else in networking** now has a category to respond to. Cisco and Arista have scale and incumbency; neither has shipped an answer to in-fabric compute for AI collectives. *In every hardware cycle, the bottleneck migrates. Whoever notices the migration one product generation early gets to define the category.* Cornelis is not attacking Nvidia where Nvidia is strong. It is attacking the seam between the accelerator and the network, a seam that only became load-bearing when clusters got large enough for collective operations to dominate the time budget. That is a narrow opening. Narrow openings are how open standards have always gotten in. ## Frequently Asked Questions ### What is Cornelis Active Compute Fabric? Active Compute Fabric is a networking architecture announced in September 2026 that embeds programmable compute directly into the network fabric. Rather than only forwarding packets, it performs operations on data in transit — assembling KV cache, coordinating mixture-of-experts dispatch, accelerating collectives such as AllReduce, and compressing gradients before they reach their destination. ### How much did Cornelis Networks raise? Cornelis Networks announced a $205 million funding round on September 14, 2026, led by IAG Capital Partners. The company said it will use the capital to scale production of its CN5000 and CN6000 network switches and accelerate deployment of its Active Compute Fabric architecture. ### How does Cornelis compete with Nvidia InfiniBand and NVLink? Cornelis builds on open standards — Ethernet and UALink for scale-up, Ultra Ethernet for scale-out — so customers can pair the fabric with accelerators from any vendor. Nvidia's networking is optimized end to end with its own stack, which makes it easier to adopt but harder to leave. ### Why are AI accelerators often idle? In large AI deployments, accelerator utilization commonly sits near half of installed capacity because GPUs wait on data movement and synchronization between nodes. Collective operations, cache transfers and synchronization steps consume accelerator time at both ends of every network hop, leaving expensive silicon idle. --- *Editor's note — sources: SiliconANGLE (September 14, 2026), TechCrunch, and Cornelis Networks' own Active Compute Fabric and CN6000 product documentation. Executive quotes are as reported by SiliconANGLE. Additional reporting referenced: HPCwire on the UALink scale-up interconnect, and Forbes on rack utilization. Traffic-reduction and utilization figures are Cornelis's own pre-production simulations and have not been independently verified.*