Cornelis Networks announced $205 million in new funding this week, led by IAG Capital Partners, alongside a product it calls Active Compute Fabric. The company spun out of Intel in 2020 and sells the wiring between AI accelerators rather than the accelerators themselves. Every write-up reached for the same frame: a challenger taking aim at Nvidia’s dominance.
The frame is not wrong, and it skips the interesting part. The pitch rests on one number, and Cornelis produced it. In a modelled 100,000-GPU system, the company says roughly half of all GPU hours are spent idle waiting for data to arrive, which it converts into about $1.68 billion of wasted capacity per year and 500 gigawatt hours of electricity, enough to supply around 48,000 American homes. That figure is a simulation built from public data, not a measured result from a customer fleet, and nobody covering the raise said so.
The Layer Is the Right Target
Start with what Cornelis has correct, because it is substantial.
Nvidia’s moat has never been only the GPU. It is that the GPU arrives attached to NVLink inside the rack and InfiniBand between racks, and those pieces do not come apart. You cannot drop a Broadcom switch into an NVLink domain or put AMD silicon on the other end of one. Buying Nvidia compute has meant buying Nvidia’s interconnect, and that bundling is what turned a chip lead into a stack lead.
Attacking the fabric rather than the chip is therefore a more serious strategy than building another accelerator. It goes at the joint that holds the bundle together.
Active Compute Fabric is Cornelis’s answer: an open architecture combining lossless transport, in-fabric acceleration and programmable compute across both scale-up and scale-out networks, so the network processes data while it moves and absorbs collective operations instead of just carrying traffic. The CN5000 switch is shipping now. The CN6000 is sampling with customers, with wider availability expected in the fourth quarter.
“AI infrastructure is reaching a point where faster endpoints alone are not enough,” chief executive Lisa Spelman said in the announcement. “The fabric has to become an active part of the compute system.”
Where the $1.68 Billion Comes From
Now the number. Cornelis is explicit that the 100,000-GPU system is modelled, and the arithmetic is reasonable on its own terms: take a cluster of that size, assume roughly 50% idle GPU time, price the idle hours, and $1.68 billion is what falls out.
The assumption is doing all the work. Utilisation in large training clusters varies enormously by workload, by parallelism strategy, by how well the job is tuned, and by what the operator counts as idle. A recommender workload and a frontier pretraining run do not sit in the same place on that curve. A 50% figure is defensible as an illustration and is not a finding.
A vendor-modelled inefficiency is a sales artifact until a customer measures the same thing on their own hardware and publishes it.
This is not an accusation of bad faith. Every infrastructure company sizes its market this way, and IAG partner Joel Whitley put the same framing around the investment, calling open-standard scale-up and scale-out networking for AI a market worth more than $55 billion by 2030. The problem is that the press picked the number up as though it described the world rather than a scenario, and buyers reading that coverage will arrive at procurement with a figure that came from the seller.
BTN’s position is narrow and practical: any operator considering a fabric swap on this thesis should instrument its own fleet first and establish what its real idle share is before underwriting anything. If it turns out to be 20%, the payback case changes completely. If it turns out to be 50%, Cornelis has a very strong product and will be able to say so with someone else’s data.
Open Already Has a Consortium
There is a second thing the coverage glossed. Cornelis is selling an open alternative to Nvidia’s proprietary fabric, and open alternatives to Nvidia’s proprietary fabric are not scarce.
UALink, the scale-up interconnect standard, has more than 85 member companies including AMD, Intel, Google, Microsoft, Apple, Cisco, Meta, AWS and HPE. Nvidia is not among them. Ultra Ethernet covers the scale-out side. HPE’s Juniper QFX5252 is already shipping as the first production switch with native support for running UALink over Ethernet, and AMD’s MI400 generation brings UALink silicon into broader availability this half.
So Cornelis is raising money to sell its own fabric into a segment where a very large consortium is converging on shared specifications. That is a legitimate position, and vertically integrated products frequently beat committee output on performance, which is the whole lesson of NVLink. It is a different story from the one in the headlines, which implied a lone challenger against a monopolist.
Qualcomm Is Validating, Not Buying
The Qualcomm involvement has been read more strongly than the announcement supports. Qualcomm Technologies will validate the architecture toward future rack-scale AI data centre designs, and Tony Pialis, who runs its data centre business, appeared with Spelman at the AI Infra Summit to talk about utilisation and AI economics.
Validation toward future designs is a technical engagement. It is not a purchase commitment, a volume agreement, or a product on a roadmap with a date. Qualcomm has been building out data centre ambitions for a while, including its AI chip deal with ByteDance, and pairing with an interconnect vendor that is not Nvidia is consistent with that. Treat it as a signal of intent from a company that needs a non-Nvidia stack to exist, which is worth something, and not as revenue.
The Number That Actually Frames This
Set the raise against what the incumbent is spending on the same layer. Since March, Nvidia has put at least $6.5 billion into photonics and optical networking: roughly $2 billion each into Coherent, Lumentum and Marvell, up to $3.2 billion into Corning, plus participation in Ayar Labs’ Series E. The Marvell investment is tied directly to NVLink Fusion. Nvidia’s equity holdings across the sector now total around $99 billion, and it has been backstopping customer purchases at extraordinary scale.
Cornelis raised $205 million. Nvidia has spent about thirty-two times that in six months buying into the component supply chain that any competing fabric would also need to draw on, with purchase commitments attached.
That asymmetry does not make Cornelis’s technology wrong. It does mean the competitive question is not whether Active Compute Fabric works. It is whether an independent fabric vendor can secure optics and packaging capacity at a workable price while the incumbent is pre-buying it, and whether a hyperscaler will take the integration risk to find out. Those are the things to watch in the fourth quarter when the CN6000 reaches wider availability, and neither of them is settled by a modelled $1.68 billion.