Heavy Reading AI Cluster Networking Operator Survey

Artificial intelligence (AI) data centers are significantly more complex than conventional data centers, requiring higher speed network connections and multiple network architectures. Hyperscale data center operator requirements have driven most networking technology developments for almost 10 years, and the introduction of AI processing to data centers has dramatically increased the bandwidth and latency challenges. AI data center operators continue to invest in new data centers and are focused on improving performance, reliability, and outcomes while meeting constraints on budget and power consumption.

Following many years of development, AI has quickly come to the fore with the introduction of large-scale generative AI services, including ChatGPT, Microsoft Copilot, and Google Gemini. AI technology continues to develop quickly, with DeepSeek and xAI Grok 3 introducing solutions using reinforcement learning and enhanced reasoning, a shift toward agentic, and companies building advanced large language models (LLMs) that require huge clusters of AI accelerators. Connecting these AI accelerators and the closely linked AI data center servers within AI clusters to the rest of the campus puts huge demands on the networking infrastructure.

Ethernet, InfiniBand, and point-to-point copper links such as NVLink are widely used in data centers but are being pushed to their limits. To meet these demands, AI data center operators are optimizing their data center networks using new tools, including network emulation and looking at new network technologies, such as Ultra Ethernet. As data center operators optimize their existing infrastructure and plan new infrastructure deployments, they must consider the ongoing developments in AI processing and the latest networking capabilities, as well as effectively evaluate real-world performance.

Recognizing the growing importance of AI data centers and the requirements to efficiently connect large AI accelerator clusters and servers, Heavy Reading (now part of Omdia) launched the AI Cluster Networking Market Leadership Program to investigate opportunities and identify challenges. During March and April 2025, Heavy Reading (now part of Omdia) conducted a global survey that received 103 responses from individuals working for service providers.

This report presents the results across all three survey topics:

• Key challenges and strategies for AI data center growth
• Shaping AI data center network evolution
• Optimizing AI data center