Cornelis Networks, the Intel spinoff that builds AI-data-center networking gear, raised a $205 million round led by IAG Capital Partners and unveiled Active Compute Fabric, a new architecture that puts programmable compute directly inside the switches and network cards carrying traffic between GPUs, rather than treating the network as a passive pipe. The company also announced a strategic collaboration with Qualcomm on rack-scale AI infrastructure design, though neither company has announced a joint commercial product -- the work so far is described as joint technology evaluation.
The bet targets a specific and expensive kind of waste. CEO Lisa Spelman says purchased GPU capacity across the industry sits idle roughly 50% of the time, waiting on the network rather than the chips themselves. Active Compute Fabric embeds programmable RISC-V cores directly into the fabric -- switches and network cards -- so operations like key-value cache acceleration and mixture-of-experts routing happen in transit, on data already moving between accelerators, instead of round-tripping back to a GPU. "The payload does not arrive the way it left," said chief marketing officer Brandon Draeger. "The work happens once, in the path the data was already taking."
The architecture runs on open standards -- UALink, ESUN, and Ultra Ethernet -- rather than a proprietary interconnect, positioning Cornelis as a vendor-agnostic alternative to Nvidia's NVLink and InfiniBand stack, which locks customers into Nvidia GPUs end to end. Cornelis's CN5000 switching platform, whose first generation began shipping in 2025, ships in its newer form today; the multi-protocol CN6000 -- an 800-gigabit SuperNIC supporting both Ethernet (RoCEv2) and Cornelis's own Omni-Path protocol -- is sampling with customers ahead of a wider release in the fourth quarter of 2026.
What the $205 million confirms, and what it doesn't
- $205M · Cornelis Networks
- New funding round led by IAG Capital Partners
Includes: Scale-up product development, CN6000 manufacturing, and customer deployments, per the company
Excludes: A stated valuation, a named round designation, and a complete investor list -- none disclosed - ~50% · GPU idle-time claim
- Spelman's stated rationale for the product
Includes: The CEO's own characterization of industry-wide GPU utilization, as stated to reporters
Excludes: Any published methodology, benchmark, or third-party measurement backing the figure
Cornelis is one of several networking challengers -- alongside Broadcom's and Marvell's custom silicon and Qualcomm's own rack-scale ambitions -- betting the AI buildout doesn't stay locked to Nvidia's proprietary interconnect stack. Cornelis itself was spun out of Intel in 2020 and has built its business on the scale-out side of that fight; Active Compute Fabric is its first move into scale-up, the tighter, faster fabric linking GPUs inside a single rack.
- Cornelis Networks raised $205 million and unveiled Active Compute Fabric, a compute-in-network design.
- CEO Lisa Spelman says purchased GPU capacity sits idle roughly half the time industry-wide.
- The design uses open standards, not Nvidia's proprietary NVLink/InfiniBand interconnect stack.
- A Qualcomm collaboration on rack-scale AI infrastructure was announced with no product yet.
- Caveat: the round's valuation, designation, and full investor list remain undisclosed.