Data center networking company Cornelis Networks raised $205 million in a round led by IAG Capital Partners and announced a new Active Compute Fabric architecture at the AI Infra Summit. The company also announced a strategic collaboration with Qualcomm to shape networking for rack-scale AI data centers.
The interesting part is not the money. It's the architectural claim: the network should be treated as a first-order design decision, not an afterthought bolted onto a compute cluster after the accelerators have been chosen.
What Active Compute Fabric actually does
Cornelis builds congestion-free networking systems for AI and HPC. Its CN5000 and CN6000 switches compete with Cisco and Arista. The Active Compute Fabric pushes programmable compute into the network itself, so the fabric can process data while it moves instead of just forwarding packets.
Chief Marketing Officer Brandon Draeger gave SiliconANGLE some concrete examples of what this looks like. The fabric can assemble KV cache data for disaggregated inference. It can coordinate expert dispatch for mixture-of-experts models. It can speed up collective operations like AllReduce and compress gradients in transit. All of this happens before data reaches its destination.
The result, according to Cornelis's pre-production simulations, is up to 50% less network traffic and less time spent by GPUs waiting on synchronization. That second part matters more than the first. An H100 or a GB200 sitting idle because it's waiting for an AllReduce to finish is an expensive way to burn electricity.
The architecture is built on Ethernet and UALink for scale-up, and Ultra Ethernet for scale-out. Open standards mean it can slot into existing compute architectures without a forklift upgrade. That is a real differentiator against proprietary interconnects.
Why this matters
There's a number in the Cornelis announcement that deserves attention. The network is about 15% of the cost of an AI system, but it determines how much value you get out of the other 85%. That framing is aggressive and also correct in most deployments I've seen. You can spend $40,000 on a GPU and $2,000 on the NIC and switch port connecting it, and then watch the GPU sit at 40% utilization because the fabric can't keep up with gradient synchronization.
Draeger framed it this way: "AI infrastructure is reaching a point where faster endpoints alone are not enough. The fabric has to become an active part of the compute system."
I'm skeptical of the marketing language, but the underlying point holds. The assumption that the network is passive plumbing and the compute is where the work happens breaks down at scale. When you're training a 70B parameter model across 4,096 GPUs, the communication pattern becomes the dominant cost in the training loop.
Qualcomm is collaborating with Cornelis on joint technology evaluation for rack-scale inference across scale-up and scale-out environments. Qualcomm has been pushing into AI data center silicon, and having a networking partner that can co-design the fabric with the accelerator is a sensible move. Whether it produces anything meaningfully better than what Broadcom or NVIDIA offer in their respective walled gardens is an open question.
AWS adds dynamic flows to WhatsApp
Also this week, AWS End User Messaging added support for dynamic WhatsApp Flows. This is not going to move markets, but it is a useful capability if you run customer-facing workflows on WhatsApp.
The pitch is straightforward. Instead of kicking a user out of WhatsApp to a web form, you can build interactive flows that run inside the chat. Appointment booking, lead gen, sign-up, surveys, transactions. Each screen in a flow can call your own HTTPS endpoint in real time, so you can show live availability or personalize each step.
You define a flow with a JSON schema, then deliver it with a message template. You build and manage everything through the AWS End User Messaging Social console or API. Meta provides starter templates with text inputs, date pickers, dropdowns, and buttons, or you can build from scratch.
For engineering teams, the relevant detail is the HTTPS endpoint callout. That means the flow is not static. You can do real-time inventory lookups, fetch account balances, or check calendar availability. This turns WhatsApp into something closer to a lightweight app platform for conversational use cases, which is where a lot of consumer traffic already lives in Latin America, India, and parts of Europe.
The catch: it's still WhatsApp. You're building on Meta's platform, subject to Meta's template approval process and Meta's pricing for business-initiated conversations. The AWS integration makes it easier to manage from your existing infrastructure, but it doesn't change the underlying dependency.
Crusoe opens a second Tulsa facility
Crusoe opened a second manufacturing facility in Tulsa. The company describes itself as an AI factory company and builds AI-optimized data centers powered by large-scale energy sources. The new facility expands production capacity for the hardware that goes into those data centers. Specifics on square footage and production capacity were not in the announcement.
The Crusoe move is part of a broader trend of AI infrastructure companies vertically integrating manufacturing. When you're building data centers at the scale Crusoe is targeting, controlling your own production capacity for the physical plant equipment removes a supply chain dependency. Whether that is cheaper than buying from established vendors depends on utilization rates, which we don't have visibility into.
Takeaways
The Cornelis announcement is the one worth paying attention to if you run AI training or inference infrastructure. The question is not whether the network matters. It clearly does. The question is whether Cornelis can deliver on the promise of in-network compute without introducing the kind of complexity that makes debugging a distributed training run even more painful than it already is. Programmable fabric sounds great until the fabric itself becomes a source of subtle, hard-to-reproduce failures.
For teams evaluating this: watch for independent benchmarks on AllReduce offload and KV cache assembly. The 50% traffic reduction figure comes from Cornelis's own simulations, not from production deployments. That number needs verification from someone who isn't selling the product.
The AWS WhatsApp flows feature is a simpler decision. If you already use WhatsApp for customer communication and you have workflows that currently require a web bounce, this removes a step. Test it with a small flow before committing. The JSON schema approach is developer-friendly, but the real-time endpoint requirement means you need backend services that can handle the latency and concurrency of a chat flow. That is not free.
Crusoe's expansion is a supply chain play. Interesting if you follow AI infrastructure economics, not actionable unless you're buying capacity from them.
The through line across all three: the physical layer of AI keeps getting more attention. Networking, chat delivery, manufacturing. The model architecture gets the headlines, but the infrastructure around it is where a lot of the real cost and real differentiation now lives.