Cornelis raises $205M and launches Active Compute Fabric for AI networking

Cornelis said it introduced Active Compute Fabric — an open networking architecture that puts programmable compute inside scale-up and scale-out fabrics for AI and high-performance computing — and announced about $205 million in funding to build the next generation of those products. Qualcomm Technologies will join Cornelis’ CEO on stage at AI Infra Summit to talk about networking’s role in rack-scale AI.

AI clusters waste money when GPUs sit idle waiting on the network. Same-day Cornelis primary pairs Active Compute Fabric with about $205 million for an open scale-up path versus locked stacks — treat the idle-hour waste figures as company modeling, not an independent audit.

Cornelis, a high-performance networking company, said via Business Wire on 14 Sep 2026 that it introduced Active Compute Fabric. A fabric, here, is the network that ties chips together. Scale-up means the fast links inside one rack. Scale-out means linking racks to one another. HPC means high-performance computing — the big science and engineering jobs that already run on these networks. That company release is the filing event. These are company claims. This desk did not plug in the switches.

The architecture marks Cornelis’ entry into scale-up networking, the company said, and extends technology it says is already running AI and HPC jobs in hundreds of data centers. File “hundreds of data centers” as Cornelis’. This desk did not count those rooms.

Cornelis also announced about $205 million in funding to expand into scale-up, build the next generation of those products, support CN6000 manufacturing and customer deployments, and push go-to-market work. TechCrunch’s Dominic-Madori Davis, writing the same day, said the round was led by IAG Capital Partners. File the lead as TechCrunch’s. The company release quotes IAG partner Joel Whitley; it does not print a full investor list. This desk is not inventing one, and it is not inventing a valuation.

Whitley said on the company release that open-standard scale-up and scale-out networking for AI is a large 2030 market, and that “the network will decide how much of the total AI build-out delivers real return.” He said IAG’s confidence sits on “a differentiated vision and a team that has demonstrated it can execute.” File the market-size talk and the quotes as his. This desk is not publishing a 2030 forecast.

CEO Lisa Spelman said faster endpoints alone are no longer enough. Quoted, via the company release: “The fabric has to become an active part of the compute system.” And: “Customers want complete rack-scale solutions without being locked into a single vendor or architecture.” File both as hers. This desk is not saying customers have already switched.

Active Compute Fabric, the company said, combines lossless transport — a network path meant not to drop packets — with in-fabric acceleration and programmable compute. That mix is meant to let the network operate on data while it is in motion, offload collective operations — the all-to-all math GPUs do when they sync — and raise how much of each accelerator is actually busy. File those jobs as Cornelis’ product claims. This desk did not time a training run.

On a modeled 100,000-GPU system, Cornelis says its own math on public data shows roughly half of GPU hours spent waiting for data — about $1.68 billion a year in wasted capacity and 500 GWh of power. Those figures are Cornelis estimates from pre-production simulation and modeling, spelled out in the release footnote. They are not an independent audit, and they are not a production benchmark this desk ran.

The architecture uses open industry standards, the company said: UALink and ESUN for scale-up, and Ultra Ethernet specifications for scale-out, with support for a broad range of accelerators. File that standards list as Cornelis’. This desk is not ranking those specs against every rival link.

Qualcomm Technologies will join the announcement at AI Infra Summit. Tony Pialis, executive vice president and general manager of data center at Qualcomm Technologies, is scheduled to join Spelman on her keynote to talk about the network’s role in efficient AI infrastructure. The company FAQ frames the work as a collaboration on validation toward future rack-scale AI data-center designs. Both companies, the release says, see today’s passive fabrics — networks that only move data — as a reason accelerators sit idle. File the shared-view and validation framing as the companies’. This desk is not inventing a closed chip co-design deal.

CN5000 is shipping today, the company said. CN6000 is sampling with customers, with expanded availability expected in the fourth quarter of 2026. Spelman keynotes at AI Infra Summit on 15 Sep 2026. Cornelis will show CN6000 and preview the next-generation scale-up and scale-out roadmap at booth #830. More is at cornelis.com/GoActive. File the ship, sample, and booth lines as Cornelis’.

TechCrunch’s same-day wrap says Cornelis spun out of Intel in 2020 and is pitching an open fabric that can sit under a mix of GPUs and accelerators, versus a locked full-stack path. File the spin-out year and the open-versus-locked frame as TechCrunch’s. Do not upgrade that to a claim that Cornelis already beats Nvidia networking in production benchmarks. This desk has no such bake-off.

CONFIRMED here: Cornelis’ 14 Sep 2026 Business Wire release — PRIMARY company source — plus the same-day FinancialContent wire reprint. REPORTED: TechCrunch’s same-day Dominic-Madori Davis wrap. The $1.68 billion / 500 GWh waste figures are company modeling, not an independent audit. NOT claimed: a valuation, a full investor list, that this desk tested the fabric, that Cornelis already beats Nvidia networking in production benchmarks, a stock tip, or investment advice. Distinct from the already-filed nvidia-cuda-q-logical, cognichip-aci-enterprise-chip-design, fujitsu-monaka-cpu-sovereign-ai, and dmatrix-nvlink-fusion-raptor.