AI Cluster Networking Report AI Grows Faster Than the Networks: Ultra Ethernet and Emulation Become Important

Related Vendors

Keysight Headquarters

Keysight Technologies Deutschland GmbH

dataTec AG

AI is sweeping through industries, but infrastructure is lagging behind. A study by Keysight and Heavy Reading shows: 89% are continuing to invest, 59% see tight budgets as the main bottleneck. The focus is shifting from expansion to optimization.

The survey by Heavy Reading commissioned by Keysight ran from mid-March to mid-April 2025 and includes 103 responses from telecom and cloud companies. The majority come from large organizations with billion-dollar budgets. The result: The AI boom is driving the demand for faster, more stable data transmission, while computing power is scaling faster than the fabrics.

At the same time, more than half report three major hurdles: budget constraints (59%), infrastructure limits (55%), and a lack of skilled workers (51%). This shifts priorities. 62% aim to get more performance out of existing infrastructure without immediately purchasing new hardware.

From Expansion to Optimization

Operators are making several adjustments: higher link speeds (61%), precise config tuning (53%), and phased GPU upgrades (50%). Changes are first implemented in the lab cluster (43%), optics are systematically qualified (34%), and switches are compared in a bake-off (27%). In the end, it's not the fastest port that matters but the reliable completion time: fewer packet losses and shorter tail latencies.

Emulation is becoming an essential tool. 95% deem it crucial because it mirrors real topologies and traffic patterns, delivers reproducible results, and detects errors before rollout. Currently, 35% use emulation productively, and 60% plan to implement it. Closely linked to this: measuring instead of guessing—completion time, p99/p999 latency, drop rates, and energy per job are decisive.

Ethernet Gains the Test Bench, InfiniBand Remains in Play

In interconnects, 400G is currently the standard (55% in use), with 800G following (34%) and 1.6 Tbit/s starting in tests (22%). In parallel, 58% are evaluating the Ultra Ethernet concept, which envisions a seamless Ethernet stack with very low latency and lossless performance as its goal. These points represent target scenarios, not universally achieved characteristics. InfiniBand/NVLink remains an option where tightly coupled GPU collectives and deterministic latency matter.

"AI data centers are reaching a tipping point where performance and scalability alone are no longer sufficient. Operators need deeper insights, stricter validations, and smarter infrastructure decisions," says Ram Periakaruppan, Vice President and General Manager, Network Applications & Security Group at Keysight. "This study confirms what we observe in practice: success in the age of AI depends on optimizing every single layer of the network. Keysight is proud to support this transformation with solutions that enable providers to emulate, validate, and future-proof their AI infrastructure on a large scale."

The sample is telco- and hyperscaler-heavy, primarily reflecting the perspective of large operators. The finding remains clear: AI is growing faster than the networks themselves. The industry is shifting its focus from expansion to optimization, with emulation as a reality check and Ethernet roadmaps heading toward 800G and 1.6 Tbit/s. (mc)

Secure and Compliant Authentication in Laboratories

Secure & Contactless User Authentication at GMP/GLP