Much of the discussion surrounding AI infrastructure lately has focused on compute and power. Faster GPUs, denser accelerator clusters, bespoke, specialized silicon, and the power required to run it all have dominated market chatter as of late. But anyone who understands the complexities of modern AI data centers knows that compute is not the sole factor that determines performance and efficiency. A significant bottleneck, that often erodes ROI and leaves multimillion dollar racks underutilized, is the network fabric and Cornelis’ new Active Compute Fabric technology is designed to address this limitation in novel ways.

With the AI Infra Summit looming, Cornelis Networks, a high-performance, low-latency interconnect fabrics company that spun out of Intel just made a handful of announcements that could alter how the industry thinks about AI rack-scale system design moving forward. With the introduction of its Active Compute Fabric, a move into scale-up networking, a $205 million funding round, and a public collaboration with Qualcomm, Cornelis is clearly committed to its assertion that the network itself must evolve from a passive transport layer into an active participant in compute.

Why The Network Has Become A Limiting Factor

AI clusters have grown from dozens of GPUs to many thousands. Models have ballooned from billions of parameters to hundreds of billions, and even trillions. Yet the fundamental behavior of most data center networks hasn’t evolved as quickly—they mostly move data from point A to point B, as reliably as possible.

The problem is that modern AI workloads don’t always behave like traditional High-Performance Compute and other distributed applications. They are dominated by synchronization, collective operations and dependencies, and massive east-west traffic patterns. When the network can’t keep up with the huge bandwidth requirements of such workloads, GPUs and other accelerators often sit idle, sometimes for surprising amounts of time.

Underutilized accelerators are not just an inconvenience—they have a direct economic impact. Every percentage point of lost utilization translates into longer training times, higher inference costs, and lower overall system efficiency, among other negative effects. In an era where AI infrastructure spending is measured in billions, and power is at a premium, this inefficiency compounds quickly and Cornelis’ announcements address this exact pain point.

Active Compute Fabric: Making The Network A First-Class Compute Resource

The core idea behind Active Compute Fabric may seem simple on the surface. Instead of treating the network as a dumb pipe, make it programmable, adaptive, and capable of performing work as data moves through it.

There are three pillars Cornelis describes regarding its Active Compute Fabric architecture. The first is lossless transport to eliminate or minimize congestion and ensure predictable performance, the second is in-fabric acceleration to offload collective operations and reduce synchronization overhead, and third is programmable compute that allows the fabric to be optimized for particular AI algorithms, applications and software.

Ideally, this means the network would be able to reshape traffic patterns dynamically, execute operations that would otherwise load the accelerators or CPU handling orchestration, and adapt to workload changes. Cornelis’ technology marks a shift from the network acting as a straightforward interconnect to the network offloading part of the compute workload.

Cornelis is grounding the architecture in open standards as well, including UALink and Ethernet for scale-up, Ultra Ethernet for scale-out, and explicitly supporting a wide range of accelerators. This is an important consideration as companies further diversify the hardware and technologies used in their data centers. AI customers increasingly want rack-scale solutions and are growing more averse to vendor lock-in, and Cornelis is positioning itself as a neutral, open alternative.

Why Qualcomm’s Participation Is Significant

Tony Pialis, who leads Qualcomm’s data center business, will join Cornelis’ CEO Lisa Spelman during her keynote. Qualcomm will be joining Cornelis on stage at the AI Infra Summit to talk about how this technology may help shape Qualcomm’s data center ambitions in the future, as it details how passive fabrics are no longer optimal for large-scale AI systems.

Qualcomm has been vocal about the economics of inference, particularly the need to squeeze every bit of efficiency out of accelerator clusters. When accelerators stall and sit idle waiting for data, inference costs rise. And when synchronization is off, throughput drops. Both companies see the network as a key design consideration that can help address these issues and something that must be architected alongside compute, not bolted on afterward.

A $205 Million Vote of Confidence

Cornelis also announced approximately $205 million in new funding, aimed at scaling production, expanding customer engagements, and accelerating go to market efforts. For a company operating in a capital-intensive segment like high-performance networking, this level of investment is meaningful and suggests that the market believes the networking bottleneck is real and that Cornelis has demonstrated enough traction through its current deployments in AI and HPC data centers to justify further expansion.

The company’s product cadence is paramount to maintain momentum. The Cornelis CN5000 fixed-function ASIC for standard packet forwarding is already shipping, while the CN6000 series of DPUs / Smart Network Interface Cards is sampling with customers now and expected to reach broader availability in Q4 of this year. Cornelis also plans to preview its next-generation scale-up and scale-out roadmap at the AI Infra Summit, spearheaded by the upcoming CN7000 series which will be the foundation of the Active Compute Fabric. The CN7000 will add many RISC-V-based cores and SRAM on every NIC and network switch controller, to distribute compute engines and memory across the network fabric.

Why This Matters for AI Data Center Economics

The industry has spent years chasing b igger accelerators, faster interconnects , and more exotic memory hierarchies. But the next wave of efficiency gains will come from maximizing the utilization of the hardware already deployed.

If a data center has 20,000 GPUs and they’re only utilized 70% of the time, it loses the equivalent of 6,000 GPUs worth of performance. That’s potentially tens of millions of dollars in stranded value. Cornelis’ Active Compute Fabric is designed to reclaim that value by reducing synchronization stalls, offloading collective operations, improving traffic predictability, allowing the network to adapt to workload changes, and supporting heterogeneous accelerator environments. To put it more simply, it aims to make every GPU, NPU, or custom accelerator work closer to its theoretical peak, for longer periods of time.

Cornelis Active Compute Fabric: The Bottom Line

Cornelis’ announcements could mark a shift in how the industry discusses AI networking. The company is betting that the future of high-performance infrastructure will be defined not just by faster endpoints, but by smarter fabrics that participate in computation, reduce overhead, and unlock higher utilization across entire racks.

With fresh funding, open-standard alignment, and a notable partnership with Qualcomm, Cornelis is positioning itself as a key player in the emerging era of active, programmable networking, over and above the Smart Network Interface Cards that are currently available in market.

As AI models continue to scale and the cost of underutilized accelerators grows, architectures like Cornelis’ Active Compute Fabric could be critical technology that helps extract maximum utilization, efficiency, and ROI, which is what every operator of an AI data center wants.