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

TL;DR

• Cornelis Networks has raised $205M in new funding to scale production and deployment of its Active Compute Fabric interconnect for large-scale AI clusters.
• Its Active Compute Fabric moves collective communication and data orchestration into the network itself to cut GPU idle time waiting for data and boost utilization.
• The raise positions Cornelis as one of the most credible independent challengers to Nvidia's InfiniBand and NVLink dominance as hyperscalers seek open, efficient alternatives.

The Money Behind the Move

Cornelis Networks has closed $205M in new financing, a major vote of confidence for an independent interconnect player in an AI infrastructure market largely controlled by Nvidia.

The round, announced this week, will be used to accelerate manufacturing, expand go-to-market with hyperscalers and neoclouds, and ramp its next-generation Active Compute Fabric platform for AI training and inference clusters. For Cornelis, a company born out of Intel's Omni-Path heritage, the funding marks a shift from high-performance computing specialist to full-stack AI networking contender.

Led by former Intel Xeon chief Lisa Spelman as CEO, Cornelis has been steadily building credibility with supercomputing centers and enterprise HPC customers. This new capital gives it the balance sheet to compete for massive AI data center deals that require scale, global supply, and long-term support.

What Is Active Compute Fabric

At the heart of the announcement is Active Compute Fabric, Cornelis's new network architecture designed specifically for AI workloads.

Unlike traditional passive fabrics that simply move packets between GPUs, Active Compute Fabric adds intelligence and compute capability directly into the network fabric. The idea is to offload and accelerate collective operations - like all-reduce, broadcast, and all-to-all that dominate distributed training - so GPUs spend less time stalled waiting for data to arrive.

The company says the approach reduces tail latency, minimizes congestion collapse at tens of thousands of GPUs, and maintains high effective bandwidth even as clusters scale. It is built around an open, Ethernet-compatible foundation with Cornelis's own silicon, smart switches, NICs, and software stack for job scheduling, telemetry, and congestion management.

In practice, Cornelis is promising faster time-to-train, higher GPU utilization, and lower total cost per token for both training frontier models and serving large-scale inference.

Why GPU Idle Time Is the Enemy

The technical pitch taps into one of the biggest pain points in AI infrastructure today: wasted GPU time.

Nvidia H100s, H200s, and Blackwell-class GPUs can cost several dollars per hour to run, yet in large distributed jobs they often sit idle 20% to 40% of the time waiting for communication, synchronization, or data delivery across the network. As models grow to trillions of parameters and clusters scale past 50,000 to 100,000 GPUs, that inefficiency becomes enormously expensive.

Cornelis argues that brute-force bandwidth is no longer enough. Without smarter in-network collective acceleration, load balancing, and adaptive routing, adding faster GPUs yields diminishing returns. Active Compute Fabric is designed to keep data flowing deterministically, avoiding hot spots and stragglers that force entire clusters to wait for the slowest node.

If it delivers, even a 10 to 15 percentage point improvement in utilization could save hyperscalers hundreds of millions of dollars in capex and power.

Taking Direct Aim at Nvidia

Nvidia's dominance in AI compute is not just about GPUs. Through its Mellanox acquisition, it controls InfiniBand, the de facto standard for high-end AI training, plus NVLink and NVSwitch for scale-up inside racks, and its newer Spectrum-X Ethernet platform.

That vertical integration - GPU plus networking plus software like NCCL and DOCA - has made Nvidia incredibly hard to displace. Competitors including AMD with Ultra Ethernet, Broadcom, Marvell, and startups like Enfabrica and Celestial AI are all trying to break that lock-in with open Ethernet-based alternatives.

Cornelis stands out because it already ships proven end-to-end fabrics and owns its silicon IP, rather than relying solely on merchant chips. Its message to cloud providers is clear: an open, high-performance fabric that works with Nvidia GPUs today, AMD GPUs, and custom accelerators tomorrow, without being locked to a single vendor's roadmap and pricing.

The $205M raise gives Cornelis firepower to invest in silicon iterations, interoperability testing, and the customer support organizations hyperscalers demand before committing to a second source at scale.

What This Means for AI Compute Competition

For the broader market, Cornelis's funding is another sign that the AI networking layer has become the next major battleground.

Hyperscalers like Microsoft, Meta, Google, and Oracle, plus sovereign AI initiatives and neoclouds like CoreWeave and Lambda, are desperate to diversify supply, control costs, and optimize power efficiency. They want open fabrics that deliver InfiniBand-class performance with Ethernet economics and flexibility.

A well-funded Cornelis could accelerate the shift toward open standards like Ultra Ethernet, push Nvidia to innovate faster on Spectrum-X and Quantum-X, and give system integrators like Dell, HPE, and Supermicro more options for building AI factories.

Challenges remain. Competing with Nvidia requires not just better hardware, but mature software, deep integration with PyTorch, NCCL alternatives, and proven reliability at massive scale. Execution on deliveries over the next 12 to 18 months will be critical.

But with $205M in fresh capital and a fabric purpose-built to eliminate GPU waiting time, Cornelis has just become one of the startups to watch in the race to power the next generation of AI infrastructure.

AndroGuider Team

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