{"slug": "cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up", "title": "Cornelis targets inference accelerator utilization with $205M expansion into scale-up", "summary": "Cornelis raised $205 million and is expanding into the scale-up networking market with its Active Compute Fabric, which adds programmable compute to the network to improve AI accelerator utilization by offloading collective operations. CEO Lisa Spelman said the offering supports open standards including UALink, Ethernet for Scale-Up Networking (ESUN), and Ultra Ethernet, and that the shift from training to inference \"puts us on the cusp of an enterprise renaissance.\" Cornelis cited figures that idling hardware equates to about $1.68 billion a year in wasted capacity and 500 gigawatt-hours of power, enough to supply nearly 48,000 U.S. homes; its CN5000 SuperNIC has started shipping while the CN6000, its first to support Ultra Ethernet as well as Omni-Path, is sampling with customers ahead of expanded availability expected later this year.", "body_md": "Networking disruptor Cornelis, buoyed by $205 million in funding, is expanding into the scale-up space to boost fabric acceleration for inference.\n\nThe startup’s Active Compute Fabric offering adds programmable compute into the network, improving accelerator utilization by offloading collective operations.\n\nCornelis positions it as a cost-saver on top of performance improvements, citing figures that idling hardware equates to some $1.68 billion a year in wasted capacity and 500 gigawatt-hours of power, “enough to supply nearly 48,000 U.S. homes.”\n\nIn conversation with *SDxCentral*, CEO Lisa Spelman said not to think of it as “just network overhead offload” and that it ultimately means data “doesn't ever really get a rest.”\n\n“Your data is being computed on when it's in the AI accelerator … but now it can also be computed while it's in the network, while it's in transport, and you can actually determine there's enough compute in the NICs (network interface cards) and switches that you may not even have to send it all the way through the entire network,” Spelman said.\n\nSpelman added that the shift from training to inference “puts us on the cusp of an enterprise renaissance.”\n\n“Every enterprise in the world will start to look at infrastructure differently and more strategically; where does that workload placement go,\" Spelman said. \"For scale-out and for scale-up, what we're talking about has the opportunity to span across a wide variety of customers.”\n\nA key aspect of the startup’s Active Compute Fabric offering is its support for multiple open standards. The likes of UALink and Ethernet for Scale-Up Networking (ESUN) for scale-up and Ultra Ethernet for scale-out count among the architectures it covers, as the startup continues to [expand beyond its traditional home of Omni-Path](https://www.sdxcentral.com/analysis/cornelis-networks-sees-supercomputing-success-as-ethernet-vision-crystalizes/).\n\nSpelman likened the ever-increasing consortia to that of VHS versus Betamax or Blu-Ray and HD DVD, and said that not forcing compute providers to choose opens up more deployment possibilities.\n\n“We're giving options, and in times when standards are uncertain, and it's still being shaken out, that dual path and multimodal capability can be extremely valuable,” the CEO said.\n\nThe startup’s latest funding will go toward expanding its scale-up efforts and bring its next generation of products to market. Its CN5000 SuperNIC has started shipping while the CN6000 – its first to support Ultra Ethernet as well as Omni-Path – is sampling with customers ahead of expanded availability expected later this year.\n\n“Broadcom and Nvidia are both amazing companies that have done wonderful things. What we have is we have the freedom of being a startup and not having to [be] encumbered by the past,” Spelman said. “We have a chance to kind of step back from these problems and look at them and not need to take incremental steps toward them.”", "url": "https://wpnews.pro/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up", "canonical_source": "https://www.sdxcentral.com/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-scale-up/", "published_at": "2026-09-15 10:55:29+00:00", "updated_at": "2026-09-15 11:43:18.083969+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-chips", "ai-startups"], "entities": ["Cornelis", "Lisa Spelman", "Active Compute Fabric", "CN5000 SuperNIC", "CN6000", "Ultra Ethernet", "UALink", "Omni-Path"], "alternates": {"html": "https://wpnews.pro/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up", "markdown": "https://wpnews.pro/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up.md", "text": "https://wpnews.pro/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up.txt", "jsonld": "https://wpnews.pro/news/cornelis-targets-inference-accelerator-utilization-with-205m-expansion-into-up.jsonld"}}