Cornelis Networks, an Intel spinoff that builds high-speed networking hardware for AI data centers, announced in September 2026 that it has raised $205 million in new funding to scale up production of its Active Compute Fabric, a GPU-agnostic networking layer designed to connect thousands of AI accelerators inside a single cluster. The round underscores a growing bet among investors that the interconnects linking GPUs together, not just the chips themselves, represent one of the biggest remaining bottlenecks — and business opportunities — in AI infrastructure.
What Active Compute Fabric actually does
Modern AI training runs increasingly depend on linking tens or hundreds of thousands of GPUs so they can share data fast enough to train a single massive model without leaving expensive chips idle waiting on network traffic. Cornelis says its current CN5000 switch ships at 400 gigabits per second per port, with an 800 Gbps CN6000 generation already planned, putting it in direct competition with Nvidia’s own Quantum InfiniBand and Spectrum-X Ethernet networking product lines. Crucially, Cornelis markets its fabric as vendor-agnostic, meaning it can connect GPUs from Nvidia, AMD, or other accelerator makers within the same cluster, a pitch aimed squarely at data center operators wary of being locked into a single chip vendor’s proprietary networking stack.
Why networking has become AI’s next chokepoint
As frontier AI labs push toward training runs that span hundreds of thousands of accelerators, the network fabric connecting those chips has emerged as a critical performance variable, since even the fastest GPUs sit idle if data cannot move between them quickly enough. Nvidia’s dominance in AI computing extends well beyond its GPUs into its NVLink and InfiniBand networking technology, which the company has bundled increasingly tightly with its chip sales, giving it outsized influence over how AI clusters are architected end to end. Cornelis and a handful of other independent networking vendors are betting that hyperscalers and neoclouds — the specialized AI cloud providers that have proliferated in 2026 — want an alternative that lets them mix and match hardware vendors rather than committing entirely to Nvidia’s full stack.
An Intel spinoff finding its own path
Cornelis traces its technology back to Intel’s Omni-Path Architecture, which Intel discontinued as a commercial product line in 2019 before the underlying team and intellectual property were spun out into an independent company. That history matters because it means Cornelis is not a startup building networking silicon from scratch, but rather a team with more than a decade of high-performance computing networking experience now applying it to the AI boom, at a moment when demand for alternatives to Nvidia’s networking stack has never been higher. The $205 million raise will reportedly go toward scaling manufacturing capacity for the CN5000 and accelerating development of the follow-on CN6000 generation.
Investors betting on multi-vendor AI clusters
Backers of the round argue that as AI data center spending climbs into the hundreds of billions of dollars industry-wide, even a small percentage shift away from single-vendor networking stacks represents a large addressable market. Proponents of open, GPU-agnostic fabrics say hyperscalers have explicitly told suppliers they want more negotiating leverage and supply-chain redundancy after two years of GPU shortages tied to Nvidia’s production constraints, making an alternative networking layer more attractive even if it requires additional integration work.
Skeptics say Nvidia’s ecosystem lock-in is hard to break
Critics of the vendor-agnostic pitch note that Nvidia’s software stack, including its CUDA programming environment and NVLink interconnect, is deeply optimized to work together, and that switching to a third-party fabric can introduce performance overhead or compatibility headaches that many data center operators are unwilling to accept for mission-critical training runs. Nvidia has also continued to invest heavily in its own networking business, following its earlier acquisition of Mellanox, making it a moving target that independent challengers must continuously catch up to rather than a static incumbent. Whether Cornelis and similar vendors can win meaningful market share depends heavily on whether major cloud providers are willing to run mixed-vendor clusters at scale in production, something that remains more common in research settings than in commercial hyperscale deployments today.
What comes next
Cornelis has not disclosed a specific timeline for the CN6000’s commercial launch beyond describing it as a next-generation product still in development. The company’s near-term test will be whether the fresh capital translates into design wins with major cloud providers or AI labs willing to deploy Active Compute Fabric at meaningful scale, rather than in pilot deployments. Investors and rival networking vendors will be watching closely for any announced customer deployments in the coming months as a signal of whether GPU-agnostic AI networking can carve out real market share from Nvidia’s dominant stack.
Photo: panumas nikhomkhai / PEXELS via Pexels
