Cornelis Unveils 'Active' Network with $205M to Tackle AI's Billion-Dollar Waste

WAYNE, PA – September 14, 2026

In the frantic gold rush to build ever-larger artificial intelligence models, there’s a secret that the balance sheets of major tech companies wish would stay hidden: immense, staggering waste. We're not just talking about energy consumption, but the quiet idling of the most expensive hardware in the data center. The powerful GPUs that are the engines of the AI revolution are spending nearly half their time waiting for data, stuck in a digital traffic jam. It’s a problem that, by one estimate, costs a single large-scale system over $1.6 billion a year in wasted capacity.

This is the story hiding in the data, and it's the challenge that networking company Cornelis aims to solve with a trio of major announcements: a new architecture called Active Compute Fabric, a hefty $205 million funding round, and a strategic collaboration with chip giant Qualcomm Technologies. The company is betting that the solution to AI’s inefficiency isn’t just faster chips, but a fundamentally smarter network.

“AI infrastructure is reaching a point where faster endpoints alone are not enough. The fabric has to become an active part of the compute system,” said Lisa Spelman, CEO of Cornelis, in a statement that gets to the heart of the issue. For years, the network has been treated as passive plumbing, a set of pipes to move data from A to B. Cornelis is arguing it's time for the pipes to start thinking.

The High Cost of Waiting

The core issue is a bottleneck. As AI models grow to astronomical sizes, they are trained across vast clusters of thousands of GPUs. These GPUs need to constantly communicate and synchronize with each other, a process that generates a chaotic storm of data traffic. Traditional networks, even high-speed ones, struggle to manage this storm, leading to congestion and delay. The result is that expensive GPUs—costing upwards of $4 per hour to run—sit idle, waiting for their next instruction.

Cornelis puts a number on this inefficiency, modeling that in a 100,000-GPU system, roughly half of all GPU hours are spent waiting. This translates not only to a staggering $1.68 billion in squandered investment annually but also 500 GWh of wasted power—enough to supply nearly 48,000 U.S. homes for a year. In the world of economics, this is a classic case of a system whose components have outpaced the infrastructure connecting them.

This is where the company’s new ‘Active Compute Fabric’ comes in. Instead of just transporting data, this new architecture aims to move the network from a passive conduit to an active participant in the computing process. By embedding programmable compute directly into the network switches, the fabric can operate on data as it moves, offload common tasks that would otherwise clog the GPUs, and adapt to the specific needs of a given workload. The goal is to have the network do work that would otherwise stall the GPU, turning wasted time back into productive computation.

A Bet on Open Roads

Perhaps the most significant part of Cornelis's strategy isn't just the technology, but the philosophy behind it: a commitment to open standards. For years, the high-performance networking space for AI has been dominated by NVIDIA's proprietary InfiniBand and NVLink technologies. While highly performant, this ecosystem creates vendor lock-in, forcing customers to buy into a single, vertically integrated stack.

Cornelis is taking a different path, building its architecture on a foundation of emerging open standards. For scale-up networking (connecting accelerators within a server rack), it will use UALink and ESUN. For scale-out (connecting racks across the data center), it will use specifications from the Ultra Ethernet Consortium. This isn't just a technical detail; it's a strategic gambit. These standards are backed by a powerful coalition of industry giants—including Intel, Google, Microsoft, AMD, and Meta—who are all seeking an open, interoperable alternative to NVIDIA's walled garden.

By aligning with this movement, Cornelis is positioning itself as the champion for customer choice. “Customers want complete rack-scale solutions without being locked into a single vendor or architecture,” Spelman noted. “We see a significant opportunity to deliver that choice with an open approach built for AI infrastructure demands.” This move could transform the market, where the projected $55 billion opportunity in AI networking by 2030 is currently a prize fought over by a handful of incumbents.

A Formidable Alliance and a War Chest

Challenging an incumbent as dominant as NVIDIA requires more than just a good idea; it requires capital and powerful allies. Cornelis appears to have secured both. The $205 million in funding is a massive vote of confidence from investors like IAG Capital Partners. "Open-standard scale-up and scale-out networking for AI represents more than $55 billion of opportunity by 2030, and we believe the network will decide how much of the total AI build-out delivers real return," said Joel Whitley, a Partner at the firm.

The funding will fuel the production of its next-generation CN6000 hardware—which is already sampling with customers—and accelerate its move into the scale-up networking market. But even more telling is the public collaboration with Qualcomm Technologies.

Qualcomm, a powerhouse in mobile and edge AI, is making a serious push into the data center. Its partnership with Cornelis signals a shared vision for a more open and efficient AI infrastructure. Tony Pialis, a top executive at Qualcomm Technologies, joined Spelman for her keynote at the AI Infra Summit, a clear signal of a deep strategic alignment.

“As AI systems continue to scale, moving data efficiently across the rack becomes just as important as the compute itself,” Pialis stated. “Improving utilization and AI economics will require a more integrated approach across compute, memory, and networking, and Cornelis’ vision for an open, programmable fabric aligns with that industry direction.”

This alliance provides Cornelis with critical validation and a powerful partner in building an ecosystem around open standards. For customers, it signals the emergence of a viable, multi-vendor alternative for building the next generation of AI supercomputers, one where the network is no longer an afterthought, but a first-order design decision.