The hardware behind the future


Cornelis Raises $205 Million to Make the Network Do Some of the GPU’s Thinking

The Intel spinout’s Active Compute Fabric puts programmable compute inside the interconnect, and its idle-GPU arithmetic is worth checking before quoting.

Cornelis Raises $205 Million to Make the Network Do Some of the GPU's Thinking

Cornelis has raised $205 million and, in the same announcement, unveiled a networking design it calls Active Compute Fabric, a name I promise will make sense before your coffee goes cold. The Wayne, Pennsylvania company says the money funds its entry into a part of the AI networking market it has never sold into, and TechCrunch reports the round was led by IAG Capital Partners.

Start with the problem. When thousands of GPUs train a single model, each chip finishes its slice of the math and then waits for the others to send their results across the network before anyone can take the next step. Cornelis's own modeling puts that idle time at roughly half of all GPU hours in a 100,000-GPU system, worth about $1.68 billion a year in wasted capacity and 500 gigawatt hours of electricity.

Those are the company's numbers, not independent ones, so read the footnote. Cornelis assumes $4.00 per GPU hour, 8,400 operating hours a year and about 50 percent unproductive time. Multiply those together and $1.68 billion is exactly what falls out, so the figure is a model of a hypothetical cluster, not a measurement of a real one.

Active Compute Fabric is the proposed cure. A fabric, in networking, is the mesh of switches and cables that ties the chips together, and most fabrics are polite couriers: they carry data from one endpoint to another and never open the envelope. Cornelis wants its fabric to do some of the work in transit, taking over the collective operations in which every GPU shares its result with every other GPU, so that chore runs inside the network instead of stalling the accelerators.

"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 the release.

The other half of the announcement is the phrase scale-up, which the industry rarely explains. Scale-out networking connects many servers across a data center. Scale-up networking connects the accelerators inside one rack so tightly that they behave like one enormous chip. TechCrunch notes that Nvidia's chips are tuned for Nvidia's own software, which makes the full Nvidia stack the easy purchase. Cornelis pitches the opposite: an open fabric built on the UALink and ESUN standards for scale-up and the Ultra Ethernet specification for scale-out, so a buyer can mix accelerators from different vendors.

Qualcomm Technologies is lending its voice. Tony Pialis, who runs its data center business, will join Spelman on stage at AI Infra Summit, and the release says the two companies share the view that passive fabrics leave accelerators underutilized. That is philosophy rather than a product commitment; the release mentions validation work toward future rack-scale designs and nothing more specific.

On the money, Cornelis says the $205 million supports scaling production, deepening customer partnerships and its go-to-market plans, along with manufacturing of its CN6000 network card. IAG's Joel Whitley sizes open-standard AI networking at more than $55 billion by 2030, a number from the investor rather than any named market study. CN5000, the current scale-out product, is shipping today; CN6000 is sampling with customers ahead of expanded availability the company expects in the fourth quarter of 2026. The scale-up products cannot be ordered yet, and the performance claims for them are, by Cornelis's own footnote, projections from pre-production simulation.

So the bet is straightforward. Cornelis is wagering that the next wave of AI buyers would rather assemble racks from parts than buy one vendor's complete system, and that a network smart enough to do some of the thinking is what sells that freedom. The customers doing the validating will answer that, not the press release.

Leave a Reply