Cornelis secures $205M, partners with Qualcomm to challenge NVLink

Cornelis Networks raised $205 million in a round led by IAG Capital Partners and announced at the AI Infra Summit the Active Compute Fabric, an open network architecture that combines programmable compute, lossless transport, and in-fabric acceleration for AI clusters. The announcement was accompanied by a strategic partnership with Qualcomm, marking the company's first significant alliance outside of Intel's legacy Omni-Path ecosystem.

Under the leadership of Lisa Spelman, former Intel, Cornelis traces its origins back to the HPC division spun off from Intel in 2020. The company had been operating in the supercomputing niche. The current funding round and the new product mark a shift toward hyperscale AI data centers, a territory currently dominated by Nvidia's InfiniBand and NVLink duo.

What is Active Compute Fabric

The current CN5000 model operates at 400 Gbps per port. The next generation, the CN6000, achieves 800 Gbps and is set to enter production in fiscal year 2027. The technical difference from the competitor comes from what Cornelis refers to as in-fabric acceleration: the switch itself executes some of the collective operations typical of distributed training, such as all-reduce and all-gather, relieving GPUs from communication rounds and freeing cycles for computation. There is also flow-based bandwidth reservation, which reduces the long tail of latency that degrades efficiency in training with thousands of accelerators.

Qualcomm's interest is operational. Qualcomm's AI200 and AI250 accelerator program, announced in May, anticipated the use of open rack-scale networking. The partnership with Cornelis materializes this layer. Cristiano Amon, CEO of Qualcomm, has repeatedly stated in earnings calls that the strategy to compete with Nvidia is not to go head-to-head in pure GPU performance but to build complete racks with open standards that capture datacenter shares from hyperscalers resisting lock-in.

Where this resonates outside the United States

In the United States, the target of the move is clear. Oracle, Meta, and a subset of neoclouds (CoreWeave, Lambda, Nebius) purchase Nvidia GPUs but avoid getting locked into the manufacturer's network layer. An open fabric with in-network acceleration provides leverage in upcoming purchasing rounds.

In India, the IndiaAI Compute Portal program, which subsidizes access to GPUs for local startups, prices cluster scale by the hour and is sensitive to any efficiency improvements that reduce training time. An open fabric opens doors for operators like Yotta Data Services and Tata Communications, which already operate in hybrid pods and are negotiating a benchmark contract with the Indian government.

In Japan, where SoftBank and NTT are investing in national AI datacenters under the banner of Sovereign AI, the open fabric fits into the strategy of reducing dependence on a single American vendor. NTT Data announced in July its intention to standardize the new Data Center Innovation Hub cluster in Yokohama on an open platform.

In Brazil, where MEC announced in July the construction of CBPF-Adamas with a projected capacity of 3 exaFLOPS, the discussion between Nvidia SuperPod and open fabric alternatives remains unresolved. The unit price per port could sway the decision, especially since funding comes from BNDES under purchasing conditions that favor interoperability.

What Cornelis lacks

It's worth reading the funding round with sobriety. $205 million covers manufacturing expansion and go-to-market efforts, not three years of cash burn against an incumbent that commands over 90% of the fabric market in AI. Cornelis needs to convert pilots into benchmark contracts before Nvidia counters with the next generation of NVLink. Analysts at Dell'Oro Group estimate the scale-up fabric market will grow from $6 billion in 2026 to $14 billion in 2028, which allows room for at least two relevant competitors, but the timing is tight.

There is a legitimate counterpoint. Ian Cutress from More Than Moore Research wrote in a note to clients on Friday that open fabric initiatives have failed before due to a lack of critical software mass: communication libraries (NCCL, RCCL), training stacks, and profiling tools remain optimized for Nvidia's ecosystem, and porting them to a new mesh requires a commitment of engineering resources that only hyperscalers can afford. On the other side, Naveen Rao, former Intel Habana, argued in a panel at the AI Infra Summit that the total cost of ownership in open racks is beginning to converge around 2027, a timeline that Cornelis seems to have internalized in the roadmap for the CN6000.

The dual partnership with Qualcomm and the association with a path of open standards are the only ways for the startup to stay alive while NVLink continues to advance within Nvidia's nest. The next indicator to watch is the list of anchor clients that Cornelis will announce by the end of the fourth quarter. Without at least one named hyperscaler, today's funding round turns into a financial bridge, not a structural market change.