Cornelis Raises $205M for Active Compute Fabric

An AI cluster can hold enough accelerators to make a finance team sweat and still spend costly hours waiting for data to move. Cornelis has raised $205M to attack that imbalance by pushing networking beyond transport and into active computation across the rack and the cluster.

The Wayne, Pennsylvania company announced the financing on September 14, 2026 alongside Active Compute Fabric, an open architecture for scale-up and scale-out networking. Cornelis says the capital will help bring its next generation of products to market, scale production, deepen customer partnerships, and accelerate go-to-market execution.

The financing arrives as AI infrastructure design shifts from counting chips to examining how well an entire system keeps those chips working. Cornelis is betting that customers will treat the network as part of the compute system, and that open standards can create a credible alternative to vertically controlled rack-scale architectures.

What Cornelis Announced

Active Compute Fabric combines lossless transport, in-fabric acceleration, and programmable compute. Instead of only carrying data between endpoints, the architecture is designed to operate on data in transit, adapt to changing workloads, and offload collective operations that otherwise consume accelerator time.

Cornelis says the architecture will use UALink and ESUN for scale-up networking and Ultra Ethernet specifications for scale-out. Its CN5000 platform is shipping today, while the 800G CN6000 is sampling with customers ahead of expanded availability expected in Q4 2026. The company will use the new funding to support CN6000 manufacturing, customer deployments, production growth, and the product roadmap behind its scale-up entry.

The announcement also places Qualcomm Technologies inside the product story. Tony Pialis, Qualcomm Technologies' EVP and GM of Data Center, is joining Cornelis CEO Lisa Spelman at AI Infra Summit to discuss networking for efficient rack-scale AI systems. Qualcomm is presented as a technical collaborator, not as an investor in the financing.

Why the Network Is Becoming Compute

Accelerator economics get ugly when utilization falls. Cornelis argues that communication overhead, synchronization, and collective operations leave expensive GPUs waiting, which means the effective cost of a training or inference job depends on more than the number and speed of the chips installed.

The company's own model of a 100,000-GPU system estimates that roughly half of GPU hours may be unproductive, representing $1.68B a year in wasted capacity and 500 GWh of power. Cornelis labels those figures as projections based on public data and pre-production modeling, so they should be read as the company's economic case for its architecture, not as audited customer outcomes.

The commercial point is still important. If networking can perform useful work while data moves, an infrastructure buyer may improve accelerator utilization without relying exclusively on a single vendor's end-to-end stack. That is the opening Cornelis is financing: make openness, programmability, and workload awareness part of the buying decision for rack-scale AI.

The People Behind the Expansion

Lisa Spelman leads Cornelis as CEO after more than two decades at Intel, including responsibility for Xeon Products and Solutions. CTO Charles Archer oversees the company's hardware and software technology stack, bringing experience from IBM, Intel, Akuna Capital, and Jump Trading.

Cornelis' three co-founders remain in operating roles. Philip A. Murphy Jr. is president and COO, Gunnar K. Gunnarsson leads solutions delivery and support, and Vladimir K. Tamarkin leads platform engineering. That continuity matters because the company is taking technology descended from Intel's Omni-Path business into a broader market and a different part of the system architecture.

Joel Whitley, a partner at IAG Capital Partners, framed the opportunity around more than $55B in open-standard scale-up and scale-out networking by 2030. The official announcement quotes IAG's investor perspective but does not disclose the full 2026 investor roster or formally identify a lead investor.

What the $205M Figure Does and Does Not Say

Cornelis describes the transaction as $205M in funding. It does not disclose a round label, valuation, full investor list, exact close date, or detailed mix of securities. Those omissions matter because a large financing can combine primary equity, debt, convertible instruments, or other capital without fitting a familiar venture-round label.

A July 20 SEC Form D/A reports $112M sold within a $245M offering that began on November 25, 2024. The filing includes equity, debt, warrants or options, and convertible promissory notes. The public record does not explain how that filing reconciles with the September headline amount, so the cleanest account is the company's own: Cornelis has announced $205M in funding, with the structure undisclosed.

The distinction avoids importing old facts into a new transaction. Cornelis' 2022 Series B was a $29M capitalization led by IAG Capital Partners with a named syndicate. That history establishes a long relationship with IAG, but it does not prove the same investors participated in 2026.

What the Capital Has to Accomplish

Cornelis now has to turn a broad architectural claim into products that customers can qualify, deploy, and support at scale. CN6000 sampling gives the company a near-term proving ground, while Active Compute Fabric extends the challenge into scale-up systems where latency, software integration, reliability, and accelerator compatibility become purchase criteria rather than presentation slides.

The $205M gives Cornelis room to finance silicon, manufacturing, customer validation, and market expansion at the same time. It also raises the operational stakes. Open standards matter only when buyers can assemble reliable systems around them, and programmable fabric becomes valuable only when software and workloads consistently use it.

Cornelis is entering that test with experienced leadership, shipping scale-out products, a next-generation platform in customer sampling, and a market looking for more choice. The coming deployments will show whether its network can move from an overlooked component into the part of the AI system that keeps the rest of the investment working.