We’re used to technology moving at breakneck speeds, but the blink-and-you’ll-miss-it pace of AI is restructuring the cloud landscape in ways we could have not imagined even two or three years ago. Demand for GPU-accelerated compute has surged and, to meet that demand, a new class of cloud provider has emerged to fill the gap left by hyperscalers – often constrained by capacity, cost or geography.
These platforms were fast to market, focused on performance rather than breadth – and highly specialised. In their early phase of life, they were loosely grouped under labels like “AI clouds” or “GPU clouds”, but neither phrasing fully captures the new generation of cloud they represent. This is a pattern borne out over the ages – practice first, definition second.
Vendors and operators were experimenting with software-based networking long before software-defined networking emerged, and compute was inching closer to users for a decade or more before the term “edge computing” became widely accepted. Now it’s the cloud’s turn. The term “neocloud” has now become embedded in the industry vernacular, to describe a distinct category of cloud infrastructure providers that are purpose-built for AI workloads, operating independently of hyperscale platforms and traditional enterprise hosting models. As consultancy firm McKinsey puts it: “Hyperscalers have secured the lion’s share of advanced-chip allocations to manage advancing AI workloads, leaving many AI start-ups, research labs and enterprises unable to access capacity at the speed they require” – and this is the very gap that neoclouds are looking to fill, but in order to do so sustainably they’ll need think about more than just raw compute.
Neoclouds might just have made it into our lexicon, but they’ve grown at extraordinary speed in the past couple of years, with some estimates pointing to year-on-year growth exceeding 200% during the height of the AI infrastructure boom, and projected revenues approaching $20bn in 2026 alone.
That momentum, however, is quite fragile. As hyperscalers expand supply and competition intensifies, the limits of a “bare-metal-first” strategy – in which hardware is king – are becoming increasingly apparent. Hardware alone doesn’t offer a long-term advantage, considering the rate at which it depreciates, which means that in order to be successful and avoid going head-to-head with dominant hyperscalers, they need to think beyond raw compute and instead consider data sovereignty, agility and optimised connectivity. This year, the defining characteristics of neoclouds will extend beyond GPUs to include network architecture, interconnection strategy and the ability to operate as a trusted part of a wider digital ecosystem.
The “bare metal” reality check
The first wave of neocloud growth was driven by an urgent shortage of AI-ready compute, but that window is narrowing fast. While the market expanded at breakneck speed, the underlying economics are becoming more challenging. Bare-metal GPU services are increasingly exposed to price pressure as hyperscalers expand capacity and competition intensifies, while the useful lifespan of accelerator hardware continues to shorten due to the sheer speed of innovation. That makes it difficult to sustain margins through hardware alone (Bare-Metal-as-a-Service, or BMaaS), particularly when neoclouds are smartly positioning themselves as alternative sources of raw compute.
As a result, many providers are shifting their focus toward AI inference workloads, which promise more predictable demand, longer-term enterprise engagement and slower hardware renewal cycles. However, inference – the real-time use of AI rather than the intensive training of models – places different demands on infrastructure, prioritising consistent performance, low latency and proximity to users and data sources above raw compute. This is where the fundamental limitation of a compute-first strategy becomes apparent, and it’s an area where neoclouds, with the right network architecture, can really shine.
Connectivity is the real differentiator
If neoclouds move toward inference-driven workloads – and many are – the limitations of traditional connectivity models become an immediate problem. Relying on best-effort public Internet transit (IP transit) offers little control over routing, latency or traffic paths, all of which are critical for delivering consistent AI performance at scale. This is where direct connectivity and peering at Internet Exchanges (IXs) make a real difference. By interconnecting directly with enterprise networks, cloud platforms and data sources, neoclouds can benefit from deterministic performance, lower and more predictable latency and far greater visibility into how data moves across their infrastructure. Peering also enables a level of control that directly aligns with emerging sovereignty requirements, allowing providers and customers to understand where traffic is exchanged and under which jurisdiction it flows.
We’ve entered the age of the neocloud, but their relevance will be short-lived unless they can succeed where hyperscalers failed. Instead of going to head to head with hyperscalers on raw compute, they need to become the “brains behind the network”, offering speed, control and visibility at the time it’s most needed in AI’s upward trajectory – and providing companies with much-needed control over their mission-critical data flows.
By Dr. Thomas King, CTO, DE-CIX
