AI growth exposes weaknesses in neocloud infrastructure

Research firm Omdia has warned that many neocloud providers are not prepared for the networking demands of artificial intelligence, urging enterprises to look beyond raw computing power when selecting suppliers.

Neoclouds are specialised, AI-first cloud providers that offer high-density GPU infrastructure and GPU-as-a-Service designed specifically for training and running large-scale AI models. 

Unlike general-purpose hyperscalers, neoclouds focus on optimised performance, high-speed networking, and flexible pricing for compute-intensive AI workloads.

The findings are based on an audit covering 50 neoclouds, which revealed that while providers have scaled compute capacity, network infrastructure is emerging as a critical constraint.

The report highlights that AI performance increasingly depends on secure, low-latency data movement across distributed systems and geographical regions.

“Network infrastructure will make or break neoclouds,” said Omdia research director for telco B2B Camille Mendler.

“Low-latency, resilient and secure connectivity from backbone to edge is table stakes for success, not least because sovereignty spans where AI workloads move,” she said.

The audit identified significant gaps in in-house networking and security expertise, with 43 per cent of neoclouds actively seeking engineers and specialists to address urgent competency shortages.

It also raised concerns over accountability and contractual obligations, noting that more than one third of providers minimise liability in areas such as network uptime, security, and data sovereignty.

The study pointed to weaknesses in cloud on-ramps and interconnection strategies, with more than half of neoclouds not using internet peering exchanges to ensure consistent performance.

Limitations in IP asset ownership were also highlighted, with 46 per cent of neoclouds controlling only small IPv4 address blocks, potentially restricting scalability and routing flexibility.

In addition, the report warned of risks to IP transit resilience, as one in five neoclouds relies on a single transit provider, creating potential single points of failure.

Omdia noted that networking capabilities vary significantly across neoclouds, reflecting their diverse origins in sectors such as bitcoin mining, content distribution, and web hosting.

The report found that neocloud networking strategies are currently in flux globally, as providers increasingly seek to partner, acquire, or build infrastructure to meet growing demand.

A group of 15 providers, led by Tier 1 backbone operators Arelion, Cogent, and Lumen, was found to control 47 per cent of all identified neocloud IP transit relationships.

Neoclouds were also shown to utilise 64 terabits per second of aggregated port capacity across 191 internet exchanges worldwide.

Among these, Frankfurt’s De-CIX exchange accounts for 10 per cent of total port capacity, underscoring its central role in global connectivity.

The report further identified Equinix and Digital Realty as key interconnection hubs, linking the largest number of neocloud providers through their global data centre networks.Omdia further stressed that enterprises adopting AI must carefully assess network readiness, resilience, and scalability, as these factors are becoming decisive in determining performance and reliability.