Nvidia pivots to 'extreme co-design' as AI networking revenues triple
Nvidia executives and theCUBE Research suggest a shift toward systems-level co-design for AI factory networking to support production workflows like distributed inference and agentic AI. The analysis highlights that standardized Ethernet protocols, specifically Nvidia's Spectrum-X platform, are becoming critical for optimizing GPU utilization and operational efficiency in distributed AI environments.
Key Takeaways
- 95% of organizations now view networking as more critical to business objectives than two years ago, per theCUBE Research.
- Nvidia's Spectrum-X platform utilizes standard Ethernet protocols rather than proprietary designs to solve congestion and jitter in distributed AI systems.
- The 'extreme co-design' strategy integrates networking, compute, storage, and software to fix bottlenecks in agentic AI and RAG workflows.
- Networking performance now directly determines economic metrics including cost per token, power consumption, and GPU utilization rates.
Why It Matters
Distributed inference and agentic AI are moving networking from a data transport layer to a primary coordinator of the compute environment. For the streaming and video industry, where low-latency delivery and real-time AI processing are merging, the ability to architect networking as a single unit with compute is vital for scaling multi-step context-aware workflows. Nvidia’s emphasis on Ethernet-standard interoperability through Spectrum-X signals an aggressive play for data centers that previously relied on siloed infrastructure. Strategists should monitor if this unified architecture successfully prevents 'idle' GPU time, which currently represents a major overhead in large-scale video AI deployments.
Additional Context
In the months leading up to July 2026, Nvidia's networking business has transitioned from a support segment to a primary revenue engine. Per IDC reporting from June 2026, Nvidia became the world's number one vendor by revenue in data center Ethernet switching for the first time in Q1 2026, with switch revenues surging 192.7% year-over-year to $2.1 billion. This growth allowed the company to capture a 21.5% share of the lucrative data center segment. By May 2026, Nvidia’s broader data center networking revenue reached a record $14.8 billion, a 199% increase from the previous year, as disclosed in its fiscal 2027 first-quarter earnings. This shift reflects a deepening rivalry with Broadcom, which has also seen its AI semiconductor sales grow over 140% during the same period. Per 24/7 Wall St. in July 2026, Broadcom has countered Nvidia’s 'universal platform' approach by prioritizing custom AI accelerators and high-margin ASICs for hyperscalers like Google and Meta. While Nvidia leverages the dominance of its Blackwell GPU architecture, Broadcom is riding a wave of custom silicon demand that focuses on optimizing cost-per-token for specific production inference workloads. Meanwhile, the broader networking ecosystem continues to push for open alternatives to proprietary fabrics. The Ultra Ethernet Consortium (UEC), founded by rivals including AMD, Broadcom, Cisco, and Intel, held a technical summit in Denver in July 2026 to finalize specifications for high-performance AI networking. According to UEC documentation, the consortium aims to adapt the Ethernet stack to handle the specific multipathing and low-latency requirements of AI clusters without the 'locked-in' nature of traditional supercomputing interconnects. Nvidia’s insistence that its Spectrum-X platform remains 'open by definition' is a strategic response to this growing industry-wide movement toward standardized, vendor-neutral hardware.
Read full article at siliconangle.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source