The problem Lisa Spelman keeps describing is a simple one: expensive accelerators spend too much time waiting. In a modern AI data center, a rack of processors can sit idle while data crawls toward it through a network that was never designed for the job. On September 14, Spelman’s company, Cornelis Networks, raised $205 million to fix that, and announced an architecture it calls the Active Compute Fabric.
The round was led by IAG Capital Partners and is the largest venture investment in the Pennsylvania region this year. Cornelis, based in Wayne, Pennsylvania, also said it has entered a strategic collaboration with Qualcomm. The two pieces of news arrived together because they describe the same bet: that the next performance gains in AI computing will come not from faster chips alone but from a network that does some of the work itself.
The idea behind the Active Compute Fabric is to embed programmable computing directly into the fabric of the network. Instead of merely shuttling data between accelerators, the network processes a portion of the data as it moves, using lossless transport so nothing is dropped along the way. The goal is to reduce the time costly accelerators spend idle waiting for data, which is among the largest hidden costs in a data center.
The underlying economics explain why this matters now. AI models have grown so large that they no longer fit on a single accelerator, or even a single rack. Training one requires thousands of chips coordinating across a network, and every moment a chip waits for data is a moment its enormous cost is not being recovered. The network, once an afterthought, has become the place where a cluster’s efficiency is won or lost.
Spelman, the company’s chief executive, came from Intel, where she spent years in the data center business before taking the top job at Cornelis. Her argument is that compute, memory, and storage have all grown smarter in recent years while the network has not. The Active Compute Fabric is an attempt to close that gap, treating the network as a participant in computation rather than a dumb pipe between boxes.
The technical foundation leans on open standards. Cornelis said the architecture runs on Ethernet, UALink, and Ultra Ethernet, the open protocols that the industry has rallied around as an alternative to proprietary interconnects. UALink, formed by a consortium of chip and cloud companies, targets the short, fast links inside a rack, while Ultra Ethernet addresses the wider fabric between racks. Cornelis is betting it can span both.
That positioning puts Cornelis in a crowded and well-funded race. Broadcom and Arista Networks are among the established players chasing the same opportunity, as the interconnect becomes the limiting factor in ever-larger AI clusters. The logic is simple: as clusters grow, the network increasingly determines how much of the expensive compute actually gets used, and whoever controls the network captures a share of the value.
The distinction between scale-up and scale-out is the technical crux. Scale-up links a small number of accelerators into a single fast system; scale-out connects thousands of systems into a data center. Most networking companies specialize in one or the other. Cornelis is promising an architecture that covers both, which would let customers buy one fabric instead of two.
The Qualcomm collaboration adds a second dimension. Qualcomm has been pushing its own AI silicon toward the data center, and a partnership with a networking company signals an interest in the systems that surround those chips. For Cornelis, the alliance provides credibility and, potentially, a path into deployments that a startup would struggle to reach on its own.
The $205 million is a statement of confidence in a moment of uncertainty. AI-related stocks have been volatile this week as investors debate the sustainability of capital spending. A company raising money to build networking for AI data centers is, in effect, betting that the buildout continues, even if the pace of the public debate has shifted.
Analysts said the timing is notable for another reason. The infrastructure that supports AI has moved in waves, from chips to memory to networking, and the networking wave is only now forming. Cornelis is trying to position itself ahead of that wave, with a product designed for the clusters that are still being planned rather than the ones already running.
The challenge is execution. Building a fabric that can process data in motion without adding latency or dropping packets is an engineering problem that has humbled larger companies. The open standards Cornelis is embracing are themselves still maturing, and the company will have to prove that its approach works at the scale its customers demand.
Spelman’s pitch is that the industry has no choice. The accelerators are fast enough, she argues, and the real bottleneck has shifted to the network. Whether that argument persuades the hyperscalers who write the checks is the question the next year will answer. For now, Cornelis has the capital to make its case.
