Cornelis is expanding beyond scale-out networking with Active Compute Fabric, an open architecture that puts programmable compute directly into the network fabric connecting AI accelerators. The company also raised $205 million to fund the move, positioning its technology as an alternative to increasingly integrated AI infrastructure stacks dominated by NVIDIA and other major silicon and networking vendors.
AI infrastructure is running into a problem that faster GPUs alone cannot solve: moving data between accelerators can consume enough time and energy to leave expensive compute capacity waiting.
Cornelis wants to address that bottleneck by turning the network itself into a computational layer.
The company has introduced Active Compute Fabric, an architecture spanning scale-up and scale-out networking that combines lossless transport, in-fabric acceleration and programmable compute. The announcement marks Cornelis’ entry into scale-up networking, extending its existing focus on high-performance connectivity for AI and high-performance computing workloads.
At the same time, Cornelis announced $205 million in funding to expand production, develop its next-generation portfolio and accelerate customer deployments.
The central idea is different from the conventional data-center network. Instead of simply transporting packets between endpoints, Active Compute Fabric is designed to perform operations on data while it is moving through the system. Cornelis says this can include offloading collective operations, adapting network behavior to workloads and performing other functions that would otherwise consume accelerator resources.
That matters as AI clusters become larger and more tightly coupled. Training and inference workloads frequently require accelerators to synchronize, exchange data and coordinate collective operations. When those operations become the limiting factor, adding more GPU capacity does not necessarily translate into proportional performance gains.
Cornelis argues that moving some of this work into the fabric can improve utilization of accelerators that customers have already purchased.
The company’s model is built around open standards. For scale-up networking, Active Compute Fabric uses UALink and ESUN, while scale-out connectivity uses Ultra Ethernet specifications. That approach puts Cornelis on a different path from proprietary interconnect ecosystems such as NVIDIA’s NVLink, which has become a central component of NVIDIA’s rack-scale AI systems.
The competition is significant. NVIDIA increasingly designs GPUs, networking, switches and systems as an integrated platform, while Broadcom, AMD, Intel and cloud providers are developing competing components and architectures. Open interconnect standards such as UALink and Ultra Ethernet are consequently becoming strategically important for companies seeking alternatives to vertically integrated AI infrastructure.
Cornelis’ timing reflects the scale of investment flowing into the sector. Gartner forecasts worldwide AI spending will reach $2.59 trillion in 2026, with AI infrastructure accounting for more than $1.36 trillion of that total. Gartner also expects AI infrastructure—including AI-optimized servers, networking and related infrastructure—to remain the largest area of AI spending.
The economics extend beyond hardware costs. The International Energy Agency estimates that global data-center electricity consumption could reach about 945 terawatt-hours by 2030, more than doubling from 2024. AI-accelerated servers are expected to be one of the biggest contributors to that increase.
Cornelis says its own modeling illustrates the potential impact. In a hypothetical 100,000-GPU system, the company estimates roughly half of GPU hours could be spent waiting for data, representing approximately $1.68 billion in annual unused capacity and 500 GWh of electricity. Those numbers are projections based on pre-production simulations and published industry data, rather than independent measurements of an operating 100,000-GPU deployment.
The company is also using the financing to broaden its product roadmap. Its CN5000 is shipping today, while CN6000 is sampling with customers, with wider availability expected in the fourth quarter of 2026. Cornelis plans to demonstrate the CN6000 and outline its next-generation scale-up and scale-out products at AI Infra Summit.
Qualcomm is joining Cornelis at the event, reflecting growing interest in networking for rack-scale AI systems. Qualcomm’s data-center business is developing AI compute products of its own, including architectures aimed at inference workloads, making efficient communication between compute, memory and networking increasingly important.
The relationship also highlights where AI infrastructure is heading. Instead of treating networking as a separate layer beneath compute, vendors are increasingly designing accelerators, memory and interconnects as a coordinated system.
Cornelis is betting that an open, programmable fabric can capture part of that transition.
Its challenge will be proving that the architecture can deliver consistent gains across real-world models and deployments while competing with much larger ecosystems. NVIDIA has a substantial advantage in integrated AI infrastructure, software and developer adoption. Meanwhile, open standards still need broad hardware and software support before they can offer the same level of integration.
For customers, however, the appeal is straightforward: more choice in how AI clusters are assembled and potentially better utilization of expensive accelerators.
As AI infrastructure moves toward rack-scale architectures, networking is becoming less about bandwidth alone and more about what happens to data between the endpoints. Cornelis’ Active Compute Fabric is an attempt to make that network an active participant in computation—and the $205 million raise gives the company substantially more resources to pursue that strategy.
Market Landscape
AI infrastructure is rapidly becoming a systems-level market rather than a collection of standalone GPUs, servers and network components. Gartner projects worldwide AI spending of $2.59 trillion in 2026, including roughly $1.37 trillion in AI infrastructure spending.
That spending is increasing pressure on networking vendors to improve accelerator utilization, latency and energy efficiency. NVIDIA remains the dominant vertically integrated competitor, while AMD, Intel, Broadcom, Qualcomm and specialist networking companies are pursuing different parts of the AI infrastructure stack.
Open standards such as UALink and Ultra Ethernet could become particularly important as hyperscalers and enterprises seek alternatives to proprietary interconnect architectures. Cornelis is positioning Active Compute Fabric around that open ecosystem while adding programmable computation to the network itself.
Energy efficiency is another driver. The IEA expects data-center electricity consumption to roughly double to around 945 TWh by 2030, with AI-accelerated servers accounting for a significant share of incremental demand.
Top Insights
- Cornelis is entering scale-up networking with an architecture designed to make the network an active computational layer for AI workloads.
- The $205 million raise gives Cornelis capital to expand manufacturing, customer deployments and its next-generation AI networking roadmap.
- Active Compute Fabric uses UALink, ESUN and Ultra Ethernet to pursue an open alternative to proprietary AI interconnect ecosystems.
- Cornelis’ GPU-utilization and energy-savings figures are modeled projections, not independently verified results from production-scale deployments.
- The shift toward rack-scale AI is making networking, memory and compute increasingly interdependent infrastructure decisions.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
