Cornelis raised $205M — data-acceleration fabrics can reduce dependence on Nvidia

To the point: funding and a network product tackling GPU idle time

Cornelis raised $205M and announced Active Compute Fabric, a network technology aimed at addressing that "much GPU time is wasted waiting for data to arrive" (per TechCrunch). Shifting some data movement into the network layer is a practical path to improving effective GPU utilization.

How to prepare infrastructure to reduce reliance on a single GPU vendor

1. Profile workloads to quantify GPU idle time waiting for data and identify batch/transfer patterns.
2. Identify data-path operations suitable to move into the network (aggregation, batch slicing, compression).
3. Run a pilot with Active Compute Fabric or DPU/SmartNIC solutions and track GPU utilization and tail latency.
4. Define success metrics: increased effective GPU usage, lower operating cost per inference/training step, and network throughput gains.

Pilot results must be measurable: TechCrunch highlights that data transfer latency is a major cause of GPU idleness.

Selection criteria for network fabrics and DPUs and common implementation traps

Evaluate solutions for:

1. Compatibility with existing ML stacks and I/O protocols.
2. Programmability and integration (APIs, Kubernetes/ML orchestration hooks).
3. Measurable impact on GPU utilization and cost-related KPIs.

Common traps include skipping workload profiling, inadequate observability on network-accelerated paths, and locking into a single vendor layer. Industry coverage notes that DPUs/SmartNICs can offload data movement and relieve CPUs, supporting the rationale for this approach (per IEEE Spectrum).

Where risks lie and the legal obligations in the EU

Moving processing closer to the network layer requires GDPR compliance for personal data in the EU - controllers must ensure legal bases, data processing agreements with processors, and conduct a Data Protection Impact Assessment (DPIA) when processing is high-risk. Check provider jurisdiction, ensure encryption in transit and at rest, and document processing roles.

Summary

Network-based alternatives such as Active Compute Fabric present a viable way to raise GPU utilization and reduce dependence on a single GPU supplier. Practical steps are profiling, piloting, defined success metrics, and EU legal checks under GDPR. Funding and product details are from TechCrunch; the role of DPUs is corroborated by industry analysis (per IEEE Spectrum).

Lub System helps B2B companies implement AI, automation and IT solutions end-to-end - from strategy to deployment. See our services or get in touch to discuss your case.