Cornelis Raises $205M to Attack Nvidia’s Networking Moat

The AI infrastructure race has spent billions making chips faster. Cornelis thinks the bigger opportunity is stopping those expensive chips from waiting around.

Cornelis Networks has raised $205 million in new funding led by IAG Capital Partners while unveiling an expanded networking architecture aimed at AI and high-performance computing systems.

The company, which emerged from Intel’s Omni-Path business in 2020, is positioning its Active Compute Fabric as an open alternative to tightly integrated networking stacks such as Nvidia’s.

The funding arrives as investors search for opportunities in every bottleneck surrounding AI infrastructure.

First it was GPUs.

Then electricity.

Now networking is becoming another multibillion-dollar battlefield.

An expensive GPU that is waiting is still wasted money

Modern AI systems rarely run on a single accelerator.

Thousands of chips may need to communicate while training models or processing inference workloads.

That creates a problem.

The GPUs can be incredibly fast individually, but the overall system still slows down if data cannot move between them efficiently.

When an accelerator sits idle waiting for information from another part of the cluster, the data-center owner is effectively paying for computing power it cannot use.

Cornelis argues that networking should therefore do more than transport packets.

Its Active Compute Fabric is designed to add processing capabilities directly into the network so certain communication and coordination tasks can happen while data is moving.

The goal is simple:

Keep the accelerators working.

Nvidia’s moat extends far beyond GPUs

Nvidia dominates AI partly because it does not sell a GPU in isolation.

Its ecosystem includes networking, software, libraries and tools designed to work together.

For customers, that integration is convenient.

It can also create lock-in.

Once a company builds around the Nvidia stack, switching one component to another supplier can introduce complexity.

Cornelis is betting that some customers increasingly want an open architecture that works across different accelerators and networking standards.

TechCrunch describes the company as part of a growing group attempting to break Nvidia’s dominance one infrastructure layer at a time.

That may be easier than trying to beat Nvidia directly at the GPU.

Open infrastructure is becoming investable again

For years, software startups attracted venture capital because software scales quickly and cheaply.

AI changed the calculation.

Investors are now financing chips, power systems, cooling, data centers and networking equipment because those physical layers are becoming strategic constraints.

Cornelis’ $205 million round reflects that shift.

Its product is far more capital intensive than another SaaS dashboard.

But networking systems that improve utilization across billions of dollars of computing infrastructure can potentially create enormous economic value.

If a data center can get meaningfully more output from the GPUs it already owns, infrastructure efficiency becomes almost equivalent to acquiring additional compute.

Qualcomm adds another interesting layer

Cornelis also announced a collaboration with Qualcomm around AI infrastructure as part of its latest expansion.

That matters because the AI accelerator market is gradually becoming more diverse.

Nvidia remains dominant, but AMD, Qualcomm, custom cloud chips and other accelerators are all competing for workloads.

A heterogeneous market strengthens the argument for networking that does not depend on one accelerator vendor.

If every data center contained only Nvidia hardware, Nvidia networking would remain the obvious default.

If future clusters mix different types of compute, open networking becomes more strategically valuable.

Networking could become AI’s next infrastructure war

The early AI boom rewarded companies capable of supplying scarce GPUs.

The next stage increasingly focuses on improving the efficiency of those GPUs.

That creates markets around networking, memory, storage and inference optimization.

Each layer represents an opportunity to reduce the total cost of AI.

And that matters because AI infrastructure spending cannot grow infinitely.

Companies eventually have to make the hardware they already bought more productive.

Cornelis is effectively betting that the network can help do that.

The challenge is Nvidia’s ecosystem advantage

Open architecture sounds attractive.

Replacing an established infrastructure stack is much harder.

Data-center operators value reliability.

Developers value mature software.

Enterprise buyers value support and compatibility.

Nvidia has spent years building those advantages.

Cornelis will therefore need to prove that openness produces enough performance, flexibility or cost benefit to justify adopting a newer alternative.

The $205 million gives it more resources to try.

What happens next?

AI infrastructure investment is broadening.

The companies raising large rounds are no longer only building the model or the GPU.

They are fixing whatever prevents the whole system from running efficiently.

Cornelis’ funding suggests investors increasingly believe networking is one of those critical bottlenecks.

The GPU race made Nvidia one of the most important technology companies in the world.

The next fight is over everything connecting those GPUs together.

And investors are starting to fund challengers accordingly.