Cornelis Networks raises $205M for AI-scale interconnect

On Monday Sep 14, 2026, Cornelis Networks made three coordinated announcements from Wayne, PA, ahead of the AI Infra Summit (Santa Clara, Sep 15–17):

1. $205 million in funding, led by IAG Capital Partners — the largest VC round of the year in the Philadelphia region per the Philadelphia Business Journal. Cornelis' FAQ states "approximately $205 million," to build out scale-up networking, develop the next-generation scale-up and scale-out roadmap, support CN6000 manufacturing and customer deployments, scale production, deepen customer partnerships and accelerate go-to-market. No valuation, series label or full investor syndicate was disclosed. IAG partner Joel Whitley: "Open-standard scale-up and scale-out networking for AI represents more than $55 billion of opportunity by 2030… the network will decide how much of the total AI build-out delivers real return."
2. Active Compute Fabric — an open architecture spanning scale-up and scale-out networking with programmable compute built into the fabric. Per the press release it combines lossless transport, in-fabric acceleration and programmable compute; uses open industry standards — UALink and ESUN for scale-up, Ultra Ethernet specifications for scale-out — and supports a broad range of accelerators. It marks Cornelis' entry into scale-up networking. CEO Lisa Spelman: "The fabric has to become an active part of the compute system."
3. Qualcomm Technologies collaboration — Tony Pialis (EVP & GM, Data Center, Qualcomm Technologies) joined Spelman's Sept 15 keynote, with both companies positioning the network as a "first-order design decision" for rack-scale AI. SiliconANGLE reported the two are "in advanced stages of joint technology evaluation focused on rack-scale inference across both scale-up and scale-out environments."

Product status at announcement: CN5000 (400G) is shipping; CN6000 (800G multi-protocol SuperNIC) is sampling with customers ahead of expanded availability in Q4 2026; a first look at the next-generation scale-up/scale-out roadmap was shown at the summit (booth #830).

The HPCwire headline framed the story as "Cornelis to Build Scale-Up Interconnect with UALink and ESUN," consistent with the press release's standards list.

• Before: Cornelis was a scale-out interconnect vendor built on the proprietary Omni-Path architecture (spun out of Intel in 2020). Its CN5000 400G platform shipped in 2025 with deployments at LLNL/NNSA CTS-2 "Lynx" and TACC Stampede3 upgrades; prior disclosed funding totaled roughly $85M–$140M depending on tracker (Series A $20M Sep 2020; $29M "Series B" capitalization Nov 2022 led by IAG; Series C $25M Mar 2024 led by IAG; plus convertibles). The company had no scale-up product and, per its own framing, was one of several alternatives flanking Nvidia's InfiniBand/NVLink complex.
• Change: Cornelis raised $205M (led by IAG Capital Partners, its lead since 2022) — more than doubling its cumulative disclosed funding — and simultaneously pivoted from a single-protocol scale-out vendor into an open-standards, compute-in-the-network, scale-up + scale-out player, with a visible Qualcomm collaboration ahead of Qualcomm's own Dragonfly rack-scale AI push.
• After: Cornelis is financed to take a run at both sides of the interconnect market — scale-out (Omni-Path, RoCEv2 Ethernet, Ultra Ethernet via CN6000) and scale-up (UALink/ESUN) — with programmable compute (RISC-V cores) embedded in NICs and switches to offload KV-cache handling, MoE routing and collectives. The company now competes simultaneously with Nvidia (InfiniBand/NVLink), Broadcom (whose Tomahawk switches it supports in Ethernet mode), Cisco and Arista, while partnering with Qualcomm and standing inside the UALink/Ultra Ethernet standards ecosystem.

• Active Compute Fabric (company description): combines lossless transport, in-fabric acceleration and programmable compute across scale-up and scale-out networks. The network can operate on data in transit — adapting to workloads, offloading collective operations, and performing functions on data as it moves. Open standards: UALink and ESUN (Ethernet for Scale-Up Networking) for scale-up; Ultra Ethernet specifications for scale-out; broad accelerator support.
• Compute placement (Network World interview): real compute — licensed third-party RISC-V cores — sits inside both the SuperNIC and the switch, going beyond a typical smart NIC/DPU that only offloads network overhead. Cornelis is fabless and integrates licensed RISC-V IP alongside its own ASIC designs. Target functions: KV-cache acceleration and mixture-of-experts (MoE) routing; inference is described as "absolutely a sweet spot" (Spelman).
• In-fabric operations (SiliconANGLE, per CMO Brandon Draeger): the fabric assembles KV-cache data for disaggregated inference, coordinates expert dispatch for MoE models, accelerates collectives like AllReduce and compresses gradients in transit — "The payload does not arrive the way it left" — so partial results from many endpoints are combined inside the fabric and a single reduced result lands at the destination. Cornelis claims this can cut overall network traffic by up to 50% (pre-production simulation — company claim).
• CN6000 multimode silicon (Network World): one chip, two wire modes — Ethernet/RoCEv2 (pairing with third-party switches such as Broadcom's Tomahawk line) or end-to-end Omni-Path with Cornelis switches. In Ethernet mode, traffic passes through an internal Omni-Path-based encapsulation layer that preserves credit-based flow control, congestion management and packet spraying, so Omni-Path's latency/flow benefits survive on an Ethernet wire.
• Economic thesis (company modeling, footnote in release): in a modeled 100,000-GPU system, ~50% of GPU hours are spent waiting for data — worth ~$1.68B/year in wasted capacity and 500 GWh (≈48,000 US homes), at $4.00/GPU-hour and 8,400 annual operating hours. Spelman: "We're giving you your GPUs back" — utilization typically 42–54%, with an active fabric targeting +5–10 points (company claim).

• Interconnect is the hidden bottleneck of multi-thousand-GPU clusters: as model and cluster sizes grow, communication overhead and synchronization idle expensive accelerators; Cornelis puts the strongest numbers on this yet ($1.68B/yr wasted in a 100k-GPU system, per its modeling — company claim, but consistent with published industry utilization studies).
• A serious, well-funded challenger to Nvidia's InfiniBand/NVLink grip: $205M gives an open-standards alternative staying power; the strategy explicitly targets accelerators "that don't want to be on the NVLink ecosystem" (Spelman). TechCrunch frames Cornelis as part of the wave "breaking apart Nvidia's market dominance, piece by piece."
• The scale-up market is reopening: NVLink has dominated scale-up (chip-to-chip within racks); UALink (AMD/Broadcom-led consortium) and ESUN are the open counter-moves. Cornelis entering scale-up via those standards — while still supporting Omni-Path/Ethernet scale-out — is a multi-protocol hedge in a standards war ("VHS versus Betamax," Spelman).
• Compute-in-the-network is a genuine architectural shift: moving KV-cache assembly, MoE routing and collectives from GPU cycles into the fabric is a different lever than adding bandwidth — with inference (the fastest-growing, most cost-sensitive workload) as the near-term payoff.

• A funded, credible NVLink alternative for scale-up: enterprises and sovereign-AI builders wanting NVLink-class rack-scale performance without Nvidia lock-in now have a financed vendor path built on UALink/ESUN.
• Ethernet-native AI scale-out with Omni-Path-like behavior: CN6000 lets operators use standard Broadcom Tomahawk Ethernet fabrics while retaining credit-based flow control and packet spraying — reducing the all-or-nothing switch-vendor commitment of Gen-1 Omni-Path.
• In-fabric acceleration for inference economics: KV-cache assembly/disaggregated inference, MoE expert dispatch and in-transit gradient compression promise higher token throughput and lower latency without more GPUs (subject to simulation-based claims being validated in shipping silicon).
• Qualcomm-anchored rack-scale inference systems: with Qualcomm's Dragonfly AI300 built to scale up via UALink and ESUN (Qualcomm's own materials), the Cornelis–Qualcomm evaluation could yield a joint reference architecture for inference racks — the same pattern Cornelis already runs with AMD (EPYC/MI400) and NextSilicon (Maverick-2).

1. Is the $205M a Series C/D or a growth round, and at what valuation? (Not disclosed.)
2. Which accelerators have actually validated against Active Compute Fabric, and what are the UALink/ESUN deployment timelines?
3. Does the Qualcomm collaboration extend beyond validation to a named reference architecture or design win for Dragonfly racks?
4. When do shipping CN6000/scale-up benchmarks (third-party, reproducible) validate the simulation-based numbers?
5. Which hyperscaler or enterprise-scale customers beyond LLNL/TACC commit to Active Compute Fabric scale-up? ("Hundreds of data centers" is unquantified in deployments vs. trials.)
6. How does Cornelis' RISC-V-in-fabric approach compare head-to-head with DPU/SmartNIC offload (BlueField, AMD Pensando) for KV-cache and MoE workloads?
7. Will Ultra Ethernet Consortium and UALink/ESUN convergence (or conflict) reshape the multi-protocol support Cornelis is betting on?

• Q4 2026 (guided): CN6000 expanded availability; first Active Compute Fabric scale-up roadmap items begin sampling; watch for named customers and reference architectures (AMD EPYC/MI400, NextSilicon Maverick-2, Qualcomm Dragonfly).
• Near term (window + 1–3 months): expect follow-on announcements from AI Infra Summit (Sep 15–17) — keynote content, first-look roadmap details, potential design wins; HPCwire/SC26 (Nov 15–20, Chicago) will likely surface technical depth and possibly third-party results.
• 2027–2028: UALink/ESUN silicon and Qualcomm Dragonfly AI300 (sampling mid-2027) converge — the defining window for whether open scale-up becomes a real NVLink competitor; watch for a Series D/growth round or strategic partnership in that timeframe.
• Watch items: whether any hyperscaler publicly commits to Active Compute Fabric; whether Broadcom deepens or resists the Tomahawk-pairing relationship; whether Nvidia responds with Ethernet-mode NVLink pricing pressure.

Cornelis' $205M raise is the strongest signal yet that the "open standard vs. NVLink" battle is now fully funded on both sides. The amount is verified (Sep 14, in-window, led by IAG Capital Partners) and the real news is the pivot: a proprietary Omni-Path scale-out vendor is reinventing itself as a multi-protocol, compute-in-the-network player spanning scale-up and scale-out — betting that the network, not the GPU, is where the next 50% of AI efficiency is hiding. That thesis is compelling and the timing (with Qualcomm, UALink and ESUN all converging on rack-scale inference) is auspicious; but every headline number is still vendor simulation. The story to tell: interconnect is becoming compute, Nvidia's rivals now have real money behind the alternative, and the 2026–2028 standards war over scale-up will decide how much of that materializes.