The short answer
Cornelis Networks raised $205 million and shipped a network that does math, not just moves bytes — aimed squarely at the GPUs that sit idle waiting for data. Announced September 14, 2026 and led by IAG Capital Partners, the round backs Active Compute Fabric, which embeds programmable RISC-V cores into network cards and switches. Cornelis, spun out of Intel in 2020, says AI clusters commonly run GPUs at only 42–54% utilization, and that handling work inside the network can lift that by five to ten points.
The bigger signal: the interconnect is now a first-class part of AI performance and cost, and an open, GPU-agnostic fabric is emerging as an alternative to Nvidia's tightly integrated stack. For teams building on cloud and DevOps, that reframes a question most had outsourced — how your accelerators talk to each other increasingly decides how much you actually get from them.
What Cornelis announced
Cornelis Networks, a Pennsylvania company spun out of Intel in 2020, used TechCrunch Disrupt to disclose a $205 million funding round led by IAG Capital Partners and to launch what it calls Active Compute Fabric. The money and the product point at the same target: the growing gap between how many GPUs organizations buy and how much useful work those GPUs actually do. Cornelis says accelerators in real AI clusters frequently run at only 42–54% utilization, sitting idle while they wait for data to arrive from other nodes.
Rather than treat the network as dumb plumbing that only shuttles bytes between endpoints, Active Compute Fabric puts programmable compute — RISC-V cores licensed from third parties — directly into the network interface cards and switches. That lets some operations happen while data is in flight. “We're giving you your GPUs back,” CEO Lisa Spelman said of the approach, arguing that a network with compute in it can “drive that GPU utilization up five points, 10 points.” On a cluster where each accelerator costs tens of thousands of dollars, a few points of recovered utilization is real money and real throughput.
The launch builds on hardware Cornelis already ships. Its first-generation CN5000 platform shipped in 2025, and the new CN6000 is an 800-gigabit SuperNIC that runs in two modes — the company's proprietary, HPC-derived Omni-Path and standard RoCEv2 Ethernet — and can pair with Broadcom Tomahawk switches or Cornelis's own. That multimode, GPU-agnostic design is the strategic core: it is a bet that buyers want an open interconnect they can use with a range of accelerators, not a single vendor's end-to-end stack.
How in-network compute works
The idea behind in-network computing is old in high-performance computing but newly urgent for AI. In a normal cluster, every byte a GPU needs travels across the network to the accelerator, gets processed, and travels back. When thousands of GPUs train one model, they spend a large share of their time on collective operations — all-reduce, all-gather, and the like — where every node must exchange and combine partial results before the next step. While that exchange happens, the GPUs are stalled. That stall is a big part of why measured utilization lands in the 40s and 50s rather than near 100%.
Active Compute Fabric attacks the stall by doing some of that work inside the network. Because the NICs and switches now carry programmable cores, operations such as combining partial results, moving key-value cache between nodes for inference, and routing tokens for mixture-of-experts models can be handled in transit instead of round-tripping to a GPU. Cornelis frames inference — where memory movement and routing dominate — as the primary opportunity, though the same offloads help large training runs. The net effect the company claims is fewer idle cycles: keep the accelerators fed, and effective utilization rises.
The competitive framing matters as much as the mechanism. Nvidia sells a vertically integrated stack — GPUs, NVLink for scale-up inside a rack, and InfiniBand or Spectrum-X Ethernet for scale-out — all optimized to its own software, which is exactly what makes the full stack easy to adopt and hard to leave. Cornelis is pitching the opposite: an open fabric aligned with emerging standards such as UALink and ESUN, so organizations can mix accelerators and avoid single-vendor lock-in. It joins a wave of infrastructure companies, from custom-silicon efforts to open Ethernet consortia, trying to pry apart Nvidia's dominance one layer at a time.
What it means for US & EU software teams
For most teams, the direct lesson is not “buy Cornelis.” It is that the network is now a first-class variable in AI performance and cost, and treating it as an afterthought quietly wastes budget. If you own or lease a cluster, the difference between 45% and 65% utilization is the difference between needing one more expensive node and not. Before you approve more hardware, it is worth confirming whether your GPUs are actually compute-bound or simply starved by communication and data movement — a distinction that changes what you should spend on.
If you consume GPUs through a cloud provider, you do not pick the fabric, but you inherit its economics. In a capacity crunch, anything that squeezes more useful work out of the same silicon flows through to per-token prices and instance availability. It also means your own architecture choices — data locality, how communication-heavy your training pattern is, which instance family and interconnect you select — increasingly determine real throughput. The GPU model on the invoice is no longer the whole story, and vendors that publish interconnect and collective-performance details deserve more of your attention than they used to get.
There is a portability angle too, and it cuts the way EU and US buyers should like. An open, standards-based fabric is a hedge against lock-in: the more of your stack that rides on UALink, Ethernet, and vendor-neutral abstractions, the more freedom you keep to switch accelerators as price, supply, and sovereignty requirements shift. That flexibility is easiest to preserve when it is designed in from the start — in how you write collective code, choose frameworks, and structure deployments — rather than retrofitted after a single vendor's stack has quietly become load-bearing.
What to do now
- Measure real GPU utilization. Instrument your training and inference jobs and find out whether accelerators are compute-bound or stalled on communication and I/O. You cannot fix — or justify spending on — a bottleneck you have not measured.
- Fix software before buying hardware. Better batching, data pipelines, and communication patterns often recover more utilization than new nodes, at a fraction of the cost. Exhaust those first.
- Ask about the interconnect. When evaluating cloud instances or an on-prem cluster, ask about bandwidth, collective performance, and any in-network offload — not just GPU type and count.
- Favor open standards where you can. Prefer accelerators, fabrics, and frameworks that align with open interconnect standards so you keep the freedom to switch vendors later.
- Track the fabric layer, not just the models. Interconnect and in-network compute are becoming decisive for AI economics. Fold them into your infrastructure planning the way you already track GPU roadmaps.
Frequently asked questions
What did Cornelis Networks announce on September 14, 2026?
A $205 million funding round led by IAG Capital Partners, disclosed at TechCrunch Disrupt, alongside the launch of Active Compute Fabric — a networking architecture that embeds programmable RISC-V compute cores into network cards and switches so some work happens as data moves through the network rather than only on GPUs. Cornelis was spun out of Intel in 2020 and already ships its first-generation CN5000 products.
What is in-network computing and why does it raise GPU utilization?
It runs certain operations on programmable silicon inside NICs and switches while data is in transit, instead of round-tripping every byte to a GPU. Cornelis says AI GPUs often run at only 42–54% utilization because they wait for data from other nodes. Handling tasks such as collective communication, mixture-of-experts routing, and key-value cache movement in the fabric keeps GPUs fed and, per CEO Lisa Spelman, can lift utilization by five to ten points — a direct cost and throughput gain on a large cluster.
How is Cornelis different from Nvidia's networking?
Nvidia sells an integrated stack — GPUs, NVLink for scale-up, and InfiniBand or Spectrum Ethernet for scale-out — tuned to its own software. Cornelis is GPU-agnostic and open: its multimode CN6000 800-gigabit SuperNIC runs both proprietary Omni-Path and standard RoCEv2 Ethernet, works with Broadcom Tomahawk switches, and targets open scale-up standards such as UALink and ESUN. Nvidia's optimized software makes its full stack easier to adopt; an open fabric gives buyers hardware choice and less lock-in.
Does this matter if we rent GPUs in the cloud?
Yes, indirectly. You do not choose your provider's fabric, but its economics reach you. If in-network compute lifts effective GPU utilization, providers get more useful work from the same silicon during a capacity crunch, which affects per-token prices and instance availability. It also validates a broader shift: the network is now an active part of AI performance, so data locality, collective-heavy training patterns, and provider selection increasingly determine real throughput — not just the GPU model on the invoice.
What should engineering teams do with this now?
Treat the network as a first-class performance variable. Profile where jobs actually spend time and confirm whether GPUs are starved by communication rather than compute. When evaluating instances or clusters, ask about interconnect bandwidth, collective performance, and in-network offload, not just GPU type and count. Keep the stack portable by favoring open standards. For most teams the highest-leverage wins are still software: better batching, data pipelines, and communication patterns that keep expensive accelerators busy.
Sources
TechCrunch — AI infrastructure company Cornelis raises $205M to chip away at Nvidia's dominance (September 14, 2026)
Network World — Cornelis lands $205M to make AI networks compute, not just connect