A Startup Just Raised $205 Million to Fix the Part of AI Chips Nvidia Would Rather You Not Think About

TL;DR: Cornelis Networks, a 2020 spinoff from Intel, raised $205 million led by IAG Capital Partners to scale its Active Compute Fabric, a networking technology that keeps GPUs busy instead of waiting on data. It’s an open alternative to Nvidia’s tightly bundled networking stack, and it’s already shipping.

Here’s a number that doesn’t get talked about enough: a meaningful chunk of the time an AI training cluster spends “working” is actually spent waiting. GPUs sit idle while data crawls over the network between them. That waiting is expensive, and it’s the exact problem Cornelis Networks says it’s solving, which is apparently a good enough pitch to raise $205 million.

The round was led by IAG Capital Partners. Cornelis isn’t a new name dressed up for a funding announcement, it spun out of Intel back in 2020 and has spent the years since building a product called Active Compute Fabric.

What the fabric actually does

Active Compute Fabric lets chips process and transmit information at the same time instead of one after the other. In practice, that means less idle GPU time during large training runs, which matters enormously when a single high-end GPU can cost tens of thousands of dollars and a cluster might contain thousands of them.

The bigger pitch, though, is openness. Nvidia’s networking stack is built to work best with Nvidia’s own GPUs and Nvidia’s own software. Cornelis is positioning its fabric to work across a mix of GPU and accelerator hardware from different vendors, so a customer isn’t locked into buying everything from one company just to get good performance. That’s the kind of argument that plays well with hyperscalers who are quietly trying to diversify away from a single supplier.

Cornelis is already shipping its current generation and has a next one in the pipeline for later this year. This isn’t a lab demo looking for its first customer, it’s an existing product looking for scale.

Part of a bigger pattern

Cornelis is one of several companies chasing pieces of the AI infrastructure stack that Nvidia doesn’t fully own outright: networking, cooling, power delivery, memory. None of them individually threaten Nvidia’s position at the center of AI compute. Together, they represent a slow erosion of the idea that buying AI infrastructure has to mean buying an entire Nvidia-shaped stack.

That’s a meaningfully different bet than trying to out-build Nvidia’s GPUs directly, which is a game most competitors have already lost. Going after the plumbing instead of the chip itself is a narrower, more winnable fight.

Bottom Line

$205 million is a rounding error next to what Nvidia spends in a quarter, but it’s real money chasing a real inefficiency. If idle GPU time is genuinely costing customers as much as infrastructure vendors claim, Cornelis doesn’t need to beat Nvidia at anything. It just needs enough customers tired of paying full price for hardware that spends part of its life waiting around.