Cornelis Raises $205M to Challenge Nvidia Dominance with Active Compute Fabric

TL;DR

• Cornelis has raised $205M in new funding to scale production of its Active Compute Fabric networking platform aimed at reducing GPU idle time in large AI clusters.
• The company's fabric technology focuses on faster, more efficient data movement between GPUs to boost utilization and offer an open alternative to Nvidia's InfiniBand and NVLink ecosystem.
• The raise positions Cornelis as a key challenger in AI infrastructure as hyperscalers and enterprises seek to diversify suppliers and lower the cost of training and inference.

The Spinout Taking Aim at Nvidia's Moat

Cornelis is not a typical AI startup. Born as a spinout of Intel's Omni-Path high-performance computing interconnect business, the company has spent years building networking hardware for some of the world's most demanding supercomputers. Now it is redirecting that HPC DNA squarely at the AI boom.

The $205M funding round marks a major escalation in that strategy. In a market where Nvidia dominates not just GPUs but the networking that ties them together, investors are betting that the next bottleneck — and the next big opportunity — is the fabric itself. As AI models grow to trillions of parameters and clusters scale to tens of thousands of GPUs, how fast data moves between chips matters as much as how fast chips compute.

Why Networking Is Now the AI Battleground

For AI labs, the math is brutal. Nvidia H100 and Blackwell GPUs can cost tens of thousands of dollars each, yet in many large training runs they sit idle 30% or more of the time waiting for data from other nodes. That idle time is wasted money and wasted time.

Nvidia recognized this early, which is why its $6.9B acquisition of Mellanox, now the foundation of its InfiniBand and Spectrum-X networking portfolio, has become such a powerful moat. Customers who buy Nvidia GPUs are strongly incentivized to buy Nvidia networking to connect them.

Cornelis, along with rivals like Broadcom, Marvell, and Ultra Ethernet Consortium startups, argues this closed coupling is bad for the industry. Its pitch is simple: deliver equal or better performance with open, standards-based networking that works across GPUs, custom accelerators, and CPUs, while dramatically cutting the cost per bit moved.

Inside Active Compute Fabric

At the center of that pitch is Active Compute Fabric, Cornelis' new networking architecture for AI data centers.

Unlike traditional passive fabrics that simply shuttle packets from point A to point B, Active Compute Fabric adds intelligence directly into the network path. The company says its CN5000 series switches and SuperNICs use advanced congestion control, adaptive routing, and in-network collective operations to keep data flowing to GPUs without interruption.

In practice, that means faster collective communications like all-reduce operations that are critical for distributed training, lower tail latency at massive scale, and the ability to maintain high throughput even when thousands of GPUs are communicating simultaneously.

Cornelis claims the result is a measurable drop in GPU idle time, higher cluster utilization, and faster time-to-train for large language models and multimodal systems. For operators running 10,000-GPU clusters, even a single-digit percentage improvement in utilization can translate into millions of dollars in savings.

What the $205M Will Fund

The company plans to use the fresh capital to ramp manufacturing of its next-generation chips and systems, expand its software stack, and grow go-to-market teams targeting hyperscalers, neoclouds, and sovereign AI projects.

That focus is deliberate. Hyperscalers like Microsoft, Meta, Amazon, and Google are all actively seeking alternatives to single-vendor lock-in to improve negotiating leverage and supply chain resilience. At the same time, a wave of government-backed AI factories in Europe, the Middle East, and Asia is looking for open, high-performance infrastructure.

Funding will also go toward interoperability work with the Ultra Ethernet Consortium, where Cornelis has been a prominent contributor pushing an open Ethernet-based standard for AI and HPC networking as a counterweight to proprietary InfiniBand.

What This Means for Competition in AI Infrastructure

Cornelis' raise will not dethrone Nvidia overnight. Nvidia still controls the full stack, from CUDA software to NVLink chip-to-chip interconnects to InfiniBand scale-out, and its execution pace remains relentless.

But the funding signals a shift in investor confidence. After years where capital flowed almost exclusively to GPU challengers like AMD, Cerebras, and Groq, attention is moving to the connective tissue around them. Networking, power, cooling, and memory are now seen as investable frontiers where openness and efficiency can win.

If Cornelis can prove Active Compute Fabric delivers sustained performance gains in real-world AI deployments — not just benchmarks — it could become a crucial second source for builders desperate to scale faster and cheaper than Nvidia's ecosystem currently allows.

In the race to chip away at Nvidia's dominance, the fight may be won not inside the GPU, but in the fabric between them.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.