The Fabric Learns to Compute: Cornelis Unveils Active Compute Fabric, a $205M Raise and a Qualcomm Alliance

As the AI Infra Summit opens this week at the Santa Clara Convention Center — 8,000 engineers, eight technical tracks, three days of talk about the plumbing underneath large language models — one of the more strategically interesting announcements didn’t come from NVIDIA. It came from Cornelis Networks, the Pennsylvania-based networking company spun out of Intel in 2020, and it amounts to a direct assault on one of the least glamorous but most expensive problems in AI infrastructure: GPUs sitting idle while they wait for data to arrive.

Cornelis introduced Active Compute Fabric, an open architecture that spans both scale-up (inside-the-rack) and scale-out (across-the-datacenter) networking with programmable compute built directly into the fabric. Alongside the architecture, the company announced $205 million in new funding and a collaboration with Qualcomm Technologies, whose data center chief will join Cornelis CEO Lisa Spelman on stage during her September 15 keynote at the summit.

Half your GPU fleet is waiting

The economic case Cornelis is making starts with a number that should unsettle anyone paying for AI compute. Drawing on public industry data, the company models that in a 100,000-GPU system, roughly half of all GPU hours are spent waiting for data — stalled on communication and synchronization rather than doing useful math. Priced at $4.00 per GPU-hour across 8,400 annual operating hours, that works out to about $1.68 billion a year in wasted capacity for a single hypothetical cluster, plus some 500 GWh of electricity — enough to power nearly 48,000 U.S. homes.

The point is not that any one operator is literally burning $1.68 billion; it’s that communication overhead has quietly become a first-order cost in frontier AI. As models and clusters grow, collective operations like all-reduce synchronize thousands of accelerators, and every nanosecond a GPU spends blocked on the network is a nanosecond of sunk capital doing nothing. Compute, storage, and memory have all become more workload-aware over the years. Networking, Cornelis argues, has mostly remained a dumb pipe.

“AI infrastructure is reaching a point where faster endpoints alone are not enough,” Spelman said in the announcement. “The fabric has to become an active part of the compute system.”

What Active Compute Fabric actually does

Active Compute Fabric combines three things: lossless transport (the congestion-free pedigree Cornelis inherited from Omni-Path), in-fabric acceleration, and programmable compute. In practice, that means the network can operate on data as it moves through the system — offloading collective operations, adapting to changing workloads, and taking on new functions as AI algorithms and software evolve, rather than waiting for the next switch silicon refresh.

The claimed result is higher utilization of the accelerators customers already own, because work that would otherwise stall the GPU runs inside the fabric itself. Crucially for an industry increasingly nervous about lock-in, the architecture is built on open industry standards: UALink and ESUN for scale-up, Ultra Ethernet specifications for scale-out, with support for a broad range of accelerators — not just one vendor’s.

That standards-first framing is the whole ballgame. Scale-up networking — the ultra-fast interconnect that binds GPUs inside a rack — is today dominated by NVIDIA’s proprietary NVLink, which the company has increasingly bundled into complete rack-scale systems like Vera Rubin NVL72. UALink is the open-specification counterweight backed by much of the rest of the industry. Cornelis entering scale-up with a UALink-based story is a bet that customers building rack-scale AI systems want accelerator choice without giving up fabric performance.

The Qualcomm angle

The Qualcomm collaboration is the detail worth watching. Tony Pialis, EVP and GM of Data Center at Qualcomm Technologies, will share the keynote stage with Spelman — and his quoted framing reads like a roadmap hint: “As AI systems continue to scale, moving data efficiently across the rack becomes just as important as the compute itself. Improving utilization and AI economics will require a more integrated approach across compute, memory, and networking.”

Qualcomm has been pushing back into the data center with its own AI accelerators, and a chip vendor that lacks an NVLink-class interconnect story needs open fabric partners. Cornelis says the two companies will collaborate on validation toward future rack-scale AI data center designs. If Qualcomm’s data center GPUs end up shipping alongside Cornelis fabrics, the alliance becomes one of the first credible open-stack alternatives to NVIDIA’s integrated rack — and a test of whether “open” can match “co-designed” on real deployments.

The money and the market math

The $205 million raise — with IAG Capital Partners publicly championing the round — is earmarked for scaling production, deepening customer partnerships, and bringing the next generation of scale-up and scale-out products to market. IAG partner Joel Whitley sizes the opportunity bluntly: open-standard scale-up and scale-out networking for AI represents “more than $55 billion of opportunity by 2030,” and “the network will decide how much of the total AI build-out delivers real return.”

On the product side, Cornelis is transitioning between generations. The CN5000 — 400G Omni-Path, the technology already running in hundreds of data centers — is shipping today. The CN6000, described as the industry’s first 800G multi-protocol SuperNIC, supports Omni-Path, Ethernet RoCEv2, and Ultra Ethernet on PCIe 6.0; it is sampling with customers now, with expanded availability expected in Q4 2026. The company is also giving the summit a first look at its next-generation scale-up and scale-out roadmap at booth #830.

From Intel castoff to thorn in NVIDIA’s side

The back story explains why this announcement carries more weight than a typical Series-round press release. Cornelis is the corporate descendant of Intel’s Omni-Path fabric business — a technology Intel nearly killed before spinning it out in 2020 with backing from Intel Capital and others. Spelman, a nearly 19-year Intel veteran who once ran the Xeon business, took over as CEO in July 2024 and has repositioned the company from HPC niche player to AI-contender, riding the same open-Ethernet wave as the Ultra Ethernet Consortium.

Analysis: the network becomes a design decision

Strip away the branding and the announcement marks a shift in how AI data centers get designed. For a decade, the network was an afterthought — you bought GPUs and then connected them. In the rack-scale era, with 100,000-GPU clusters and gigawatt campuses, the fabric determines how much of the total build-out produces real tokens versus idle heat. NVIDIA internalized this long ago with NVLink and NVSwitch; the open ecosystem is now racing to offer the same coherence without the exclusivity.

Cornelis’s wager — programmable compute inside an open, standards-based fabric — is technically distinctive, but it faces a familiar challenge: NVIDIA’s moat was never any single product, it’s the ability to sell compute, network, storage, and software as one co-designed system. An open fabric only wins if the pieces around it integrate as smoothly. The Qualcomm alliance suggests the open camp understands this. Whether validation projects turn into production racks — and whether in-fabric computing delivers measurable GPU-utilization gains beyond simulations — is what to watch as CN6000 reaches general availability in Q4.

For an industry simultaneously debating whether AI is moving too fast and struggling to make its existing compute pay for itself, a technology that promises to reclaim half of the world’s most expensive idle silicon is a welcome kind of news.

Performance figures cited are Cornelis projections based on pre-production simulation and modeling of a 100,000-GPU system; actual results will vary.