Cornelis raises $205 million for a scale-up fabric that ships inside other vendors' racks

The scale-up business Cornelis has just entered may never produce a Cornelis box. CRN reported that the move may not result in channel-ready products directly from the startup, and that the company will instead put its Active Compute Fabric architecture into the hands of firms building rack-scale platforms, Qualcomm among them ^[11]. Scale-out, the existing business, links servers across a large cluster; scale-up links processors inside a single rack-scale system ^[24].

The fabric combines programmable compute with lossless data transport and in-fabric acceleration, moving some processing off the accelerators and into the network itself ^[12]. It supports Ultra Ethernet and UALink, both consortium efforts aimed at Nvidia's proprietary interconnect stack ^[15]. The switch shipping now is the CN5000 at 400 Gbps. The CN6000 handles multiple protocols at 800 Gbps and is expected to be broadly available in Q4 2026 ^[13], double the throughput ^[14], six years after Intel spun Omni-Path out into Cornelis ^[20][23].

Two accounts of the Qualcomm relationship are on the record, and one claims more than the other. CRN wrote that Cornelis did not share details of what it called a "collaboration", and described the scale-up move as having Qualcomm's potential backing ^[4][1]. Cryptobriefing.com called it a strategic partnership that the two companies plan to showcase at the AI Infra Summit ^[5]. Both put Tony Pialis, general manager of Qualcomm's data center business, alongside Cornelis CEO Lisa Spelman this week ^[10]. Pialis said in a statement that "Giving customers more choice and flexibility across the infrastructure stack will be critical as AI moves toward rack-scale architecture" ^[9]. Qualcomm plans to debut its first rack-scale platform this year with its AI200 accelerator, with successor products in the years after ^[6].

The demand argument is a procurement one. An executive at a large U.S. systems integrator, who asked CRN not to be identified, said "If you're not married to one architecture, there are cost implications as well. There are availability implications as well, and the dependencies implication as well. So from a purely business point of view, this is great" ^[19]. Resellers cannot sell it yet; Qualcomm has not announced a partner program for its data center business ^[17]. Spelman said "I think there's a world in which some of these channel players become mechanisms for rack delivery and servicing of that enterprise" ^[22].

I would treat the $205 million as development money for an architecture with one publicly named prospective host. Cornelis did not say how the round splits between scale-up development, CN6000 manufacturing and customer deployments ^[26]. The open-alliance reading fails if Qualcomm's rack arrives with NVLink inside: Nvidia has spent more than a year promoting NVLink to rival accelerator makers as the de facto scale-up method, and that campaign has been winning their support ^[16][25]. What Cornelis has in production is supercomputing work, including the Lynx machine at Lawrence Livermore National Laboratory, which runs on its networking ^[21].