What is Cornelis Active Compute Fabric?

Active Compute Fabric (ACF) is a GPU-agnostic, open-architecture scale-out networking layer announced by Cornelis Networks on September 14, 2026 alongside a $205 million round led by IAG Capital Partners. Its pitch is that GPUs waste time waiting for data, so the fabric lets chips process and send at the same time; pre-production simulations claim up to 50% less network traffic in large AI clusters. The shipping CN5000 switch runs 400 Gbps; a multi-protocol 800 Gbps CN6000 is due in Q4 2026. Cornelis spun out of Intel in 2020 with the Omni-Path lineage and has a strategic collaboration with Qualcomm for rack-scale inference.

Is Active Compute Fabric a replacement for NVLink?

No. NVLink (NVLink 6 on Rubin NVL72) is a scale-up fabric that makes 72 GPUs in a rack behave like one memory-coherent device at multi-terabyte-per-second bandwidth. ACF, InfiniBand, Spectrum-X and Ultra Ethernet are scale-out fabrics that connect racks and nodes at 400-800 Gbps per port. ACF competes with InfiniBand and Ethernet for the scale-out tier; the scale-up alternative to NVLink is UALink, which AMD runs over Broadcom Tomahawk 6 switches in its Helios racks.

InfiniBand vs Ethernet for AI in 2026: which is winning?

Ethernet, on volume. NVIDIA's own Spectrum-X Ethernet is growing faster than its Quantum InfiniBand line, hyperscalers standardize on Ethernet, and the Ultra Ethernet Consortium's 1.0.x transport spec (1.0.3 circulating in August 2026) gives Ethernet InfiniBand-class RDMA semantics on a multi-vendor standard. InfiniBand keeps the lowest and most predictable latency and remains common in training clusters; analyses cite roughly 15% better performance at about 2.3x the TCO of Ethernet. Cornelis sits between them: not Ethernet, not InfiniBand, but open to any GPU.

Who should consider Cornelis instead of NVIDIA networking?

Operators building mixed or non-NVIDIA accelerator clusters (AMD, Qualcomm AI200/AI250, custom silicon) who want one fabric across all of them, and buyers who want a second source against NVIDIA's bundled GPU-plus-networking stack. NVIDIA-only shops rarely switch, because NVIDIA GPUs are tuned for NVIDIA's NCCL and Spectrum-X/InfiniBand software and the integration cost of another fabric is real. Wait for independent deployment results before believing the 50% traffic-reduction figure.

Quick Answer

Cornelis Active Compute Fabric vs InfiniBand vs NVLink 2026

Published: September 16, 2026

The short answer

Active Compute Fabric is a new open, GPU-agnostic scale-out network for AI clusters, competing with InfiniBand and Ethernet, not with NVLink. Cornelis Networks announced it on September 14, 2026 with a $205 million round, a shipping 400 Gbps switch, an 800 Gbps generation due in Q4 2026, and Qualcomm as a rack-scale inference partner. It is a credible second source for clusters that are not all-NVIDIA, and an unproven one until customers publish results. Verified September 16, 2026.

	Cornelis Active Compute Fabric	NVIDIA InfiniBand (Quantum)	NVIDIA NVLink 6 / NVSwitch	NVIDIA Spectrum-X Ethernet	Ultra Ethernet (UEC)
Tier	Scale-out (node to node, rack to rack)	Scale-out	Scale-up (GPU to GPU inside a rack)	Scale-out	Scale-out
Openness	Open architecture, GPU-agnostic	NVIDIA-controlled	NVIDIA-only (licensed to select partners via NVLink Fusion)	NVIDIA product on standard Ethernet	Multi-vendor spec (AMD, Broadcom, Cisco, Meta, Microsoft, etc.)
Port speed today	400 Gbps (CN5000); 800 Gbps CN6000 in Q4 2026	800 Gbps (Quantum-X800)	Multi-TB/s per GPU aggregate	800 Gbps (Spectrum-X800), Spectrum-6 for Rubin	800 Gbps switches (e.g., Broadcom Tomahawk 6)
Key claim	GPUs compute and communicate simultaneously; up to 50% less traffic (simulated)	Lowest, most predictable latency; mature at hyperscale	72 GPUs as one device; the reason NVL72 racks exist	Ethernet economics with InfiniBand-like congestion control; multiplane holds 90% bandwidth under partial failure	Standard RDMA transport for AI over Ethernet; spec 1.0.x public
Maturity	Shipping, pre-production performance claims	Very mature	Rubin generation pre-release; NVLink 5 shipping	Mature, fastest-growing NVIDIA networking line	Spec done; ecosystem building
Best fit	Mixed/non-NVIDIA accelerator clusters; second-source buyers	NVIDIA training clusters where latency dominates	Any NVIDIA rack-scale system	NVIDIA clusters standardizing on Ethernet	Hyperscalers and AMD/Broadcom-based clusters

Scale-up versus scale-out: the category error to avoid

Every AI cluster has two networks:

Scale-up connects the GPUs inside a rack (or a few racks) into one coherent domain. This is NVLink and NVSwitch on NVIDIA systems, UALink on AMD’s Helios racks. Bandwidth is measured in terabytes per second per GPU and latency in hundreds of nanoseconds. Nothing Cornelis sells lives here.
Scale-out connects those racks into a cluster of thousands or hundreds of thousands of GPUs. This is InfiniBand, Ethernet (Spectrum-X, Ultra Ethernet) and now Active Compute Fabric. Bandwidth is 400-800 Gbps per port.

Treating them as substitutes is the expensive mistake. Cornelis’s opportunity is only the scale-out tier, which is also where openness matters most because that is where the multi-vendor cluster is assembled.

What Cornelis is actually claiming

Active fabric. Conventional networks move bytes; Cornelis says its fabric lets chips “process and send information at the same time,” attacking idle GPU time spent waiting for collectives to finish. That is the same problem NVIDIA’s SHARP in-network reduction and Spectrum-X’s adaptive routing address, but Cornelis does it independent of the GPU vendor.
Up to 50% network traffic reduction in large clusters, from pre-production simulations. Treat as a vendor number until a deployment confirms it.
Roadmap: CN5000 at 400 Gbps shipping now; CN6000 at 800 Gbps, described as multi-protocol, in Q4 2026.
Qualcomm collaboration for rack-scale inference, which gives ACF a non-NVIDIA accelerator to pair with as Qualcomm’s AI200/AI250 racks reach customers (Qualcomm signed a 60B-scale supply deal with AWS in September 2026).
Lineage: Cornelis was spun out of Intel in 2020 with the Omni-Path interconnect business, so it has HPC-scale fabric experience; its previous products were mostly in supercomputers, not AI clouds.

Where InfiniBand still wins

Predictable latency and a decade of operational scar tissue at hyperscale. Training runs that synchronize gradients across tens of thousands of GPUs remain sensitive to tail latency, and InfiniBand’s credit-based flow control and mature NCCL integration are hard to match. The cost is TCO: industry analyses put InfiniBand around 15% faster at roughly 2.3x the cost of Ethernet, and its ecosystem is effectively one vendor.

Where Ethernet is winning

Volume and choice. NVIDIA’s Spectrum-X is its fastest-growing networking line, and at Hot Chips 2026 NVIDIA showed a multiplane Spectrum-X design that needs 1.7x fewer scale-out switches and holds 90% bandwidth under partial link loss where a traditional multi-tier fabric drops to zero. On the open side, the Ultra Ethernet Consortium’s transport spec is public in 1.0.x form (1.0.3 in August 2026 materials), so Broadcom, Cisco, AMD and the hyperscalers can build interoperable RDMA-class AI networks without NVIDIA. Note two distinctions: Spectrum-X is a vendor product, not the UEC standard, and a finished spec is not yet a mature multi-vendor NIC-and-switch ecosystem.

Where NVLink is untouchable, and where UALink fits

Inside the rack, NVLink 6 on Rubin NVL72 is why SemiAnalysis measured 2-7x more tokens per megawatt than GB300 on agentic inference; it is co-designed with the GPU, the Vera CPU, ConnectX-9, BlueField-4 and Spectrum-6. The open alternative is UALink, founded in May 2024 by AMD, Broadcom, Cisco, Google, HPE, Intel, Meta and Microsoft; AMD runs the UALink protocol over Broadcom Tomahawk 6 Ultra Ethernet switches inside Helios. Cornelis has no scale-up product and does not claim one.

Decision guide

You are building…	Scale-up	Scale-out
All-NVIDIA training cluster	NVLink	InfiniBand (latency) or Spectrum-X (cost)
All-NVIDIA inference fleet	NVLink	Spectrum-X
AMD Helios racks	UALink over Tomahawk 6	Ultra Ethernet or Cornelis ACF
Mixed NVIDIA + AMD + Qualcomm	Per-rack native	Cornelis ACF or Ultra Ethernet; pick on measured collectives, not slides
Sovereign/second-source mandate	Whatever the accelerator dictates	Cornelis ACF or UEC-based Ethernet

What to watch

A named customer publishing collective-operation benchmarks on ACF versus Spectrum-X.
CN6000 shipping in Q4 2026 at 800 Gbps, matching NVIDIA’s current port speed.
UEC-compliant NICs from multiple vendors, which would make ACF’s “open” argument compete with a standard rather than with NVIDIA.
Qualcomm rack deployments at AWS, the first place ACF could show up at scale.