Neoclouds have grown fast by offering cost-effective GPU services for AI training, but price pressure, hyperscaler dominance, and rapid hardware depreciation are forcing these players to rethink their strategy.
The future isn’t just about raw compute – it’s about sovereignty, agility, and connectivity. Discover why, according to Dr. Thomas King, peering and network control are becoming the secret weapons for neoclouds to evolve from commodity providers into enterprise-grade leaders in AI inference.
Neoclouds – smaller, independent GPU as a Service operators that provide infrastructure for training large language models – have been emerging over the last couple of years as new players in the cloud landscape, boosted by the massive demand for AI infrastructure. These young players are growing rapidly: Synergy Research Group estimates the sector has grown over 200% in the past year and forecasts 69% per year until 2030. Already in 2026, research and advisory firm Forrester forecasts that AI-focused neocloud providers will generate US $20 billion in revenue. Despite their rapid growth – these newcomers are facing significant challenges: Price erosion in the Bare Metal as a Service sector and increasingly short refresh cycles (as new, higher-powered AI servers regularly appear on the market) are driving their margins ever lower. In turn, AI chip generations are depreciating three to five times faster than traditional hardware.
Neoclouds are finding themselves in a double bind in the maturing AI infrastructure industry. Unable to compete head-to-head with the major global hyperscalers, they are instead dependent on these large players as major customers. Hyperscalers procure their bare metal services, value add, and sell on their own services at a premium. In the world of AI, this relegates neoclouds to the status of commodity providers, with prices more than 50% cheaper than hyperscalers, according to Uptime Institute figures from 2025. With the big three hyperscalers – Amazon Web Services (AWS), Microsoft Azure, and Google Cloud – capturing more than 60% of the global cloud market in 2025, there’s not a lot of room for smaller operators to increase their margins. So neoclouds must evolve fast in 2026 in order to survive.
Transforming the business model of neoclouds
Sovereign cloud services – and in particular, those offering increasingly sought-after sovereign AI – present a promising option for GPU as a Service providers to carve out a niche for themselves. Here, their independence from the hyperscalers becomes a unique selling proposition. As sovereign cloud providers, they will need to delve deeper into the world of networking: offering control of data processing locations and ensuring jurisdiction-aware interconnection and data pathways, while attending to data security and protection in conformance with local laws.
Data sovereignty is a growing priority for companies the world over. And it’s not simply about what enterprises demand: In 2026, data sovereignty is set to also become essential from a compliance perspective, with the EU’s proposed Cloud and AI Development Act (CADA) potentially doing for digital sovereignty what the General Data Protection Regulation (GDPR) did for personal data nearly a decade ago.
Additionally, neoclouds have the opportunity to move their business model away from bare metal for AI training towards supporting enterprise customers with AI inference. This transformation is already afoot: While in 2026, inference workloads are set to make up around one third of the niche’s revenues, technology solutions consultancy ABI Research forecasts that by 2028, inference will have closed the gap and become the dominant revenue source.
Becoming AI as a Service providers for both training and inference has a range of benefits. Firstly, greater margins as a full-service provider. Secondly, long-term relationships with enterprise customers – with inference offering income stability to counter the piecemeal nature of training. Then there’s the fact that there is less pressure to upgrade all equipment to the newest CPUs or TPUs every six months, as inference is more demanding of agility than it is of processing power. Add that to the competitive advantage of offering sovereign cloud services for enterprises, and this is a promising combination.
The criticality of a mature connectivity model for future success
To successfully make this transition, these new kids on the block will need both geographical reach and technical reachability. Their presence in each market can be strengthened by working together with local colocation data center operators, expanding their geographical presence to and within relevant markets. However, they will need to build a more mature connectivity model. A model that can interconnect colocation facilities and their own data centers, give them access to the data they need, and deliver the inference where and when customers want it.
With the move towards AI inference, transporting data by the truckload is no longer an option. As Inference as a Service providers, neoclouds will need up-to-the-moment, real-time access to data from a wide range of sources, depending on the use case: at the very least from the hyperscalers, customer networks, access networks, and IoT clusters. While bandwidth continues to play a role for AI training, latency, resilience, and predictability now emerge as business-critical connectivity criteria.
Enterprises seeking sovereign services demand control, observability, security, predictability, and flexibility when it comes to networking. To fulfill these requirements, service providers must understand their connections to relevant networks and essential data sources and have the agility to adjust these whenever necessary. They must know which pathway data is flowing along, and how long it will take to get to its destination. They will need to understand which networks, which data centers, and which jurisdictions are involved. And in ensuring the sovereignty of their data flows, they simultaneously ready themselves for becoming AI inference service providers. Connectivity is central to both transformations.
In the same way that many end-users are still using the public Internet to access cloud resources, many neocloud operators are still using IP transit to access their data, wherever it is on the Internet. Unfortunately, IP transit does not offer the control needed for sensitive enterprise use-cases. Put simply, the Internet was conceived as a best-effort technology, with no performance guarantees and no control over data pathways. Data flows according to the policies of each network operator, and it is not possible to observe the route, which networks the data flows through, or at which data centers data is exchanged between networks. IP transit, a perfectly acceptable method of outsourcing Internet access for use cases where latency, performance, security, and data sovereignty are not critical, is not fit for purpose for use-cases where they are.
Peering and AI-IX: Why neoclouds should be present at Internet Exchanges
What these young clouds should be doing instead is peering. Peering at Internet Exchanges (IXs) brings network infrastructure under the control of the individual network (or cloud) operator. It provides a mature network connectivity solution with fine-grained control and transparency over not only traffic flows and which networks data is exchanged with, but also in which data centers the exchange takes place. It reduces the number of network hops, thus improving the latency of the connection and the performance of applications like AI inference. It also enables better filtering of traffic and protection against DDoS attacks.
Using distributed and neutral IXs (neutral in terms of both data center operators and carriers) – like those operated by DE-CIX – enables these emerging full-service providers to implement a multi-provider approach. This has the benefit of significantly increasing resilience against localized outages on any part of the infrastructure value chain, as well as offering a commercial advantage in negotiations with said infrastructure providers. Peering thus presents neoclouds with the opportunity to gain sovereignty over their own network infrastructure. An asset they can pass on to their customers.
In addition, with DE-CIX’s AI-IX functionality for AI inference use-cases, neoclouds can make themselves more accessible to their enterprise customers. By becoming a DE-CIX DirectCLOUD partner, they enable enterprises to connect directly with their network – alongside those of the hyperscalers – over the DE-CIX platform and give their customers the ability to control data flows within their multi and hybrid cloud environments. A plus for reachability. A plus for simplicity in connecting customers’ networks in a holistic, efficient, and resilient way. And another component in a mature connectivity solution.
Capitalizing on sovereignty to build new revenue streams with an AI-native stack
As neoclouds evolve towards an enterprise-facing AI-native software stack in 2026, they will be able to build new revenue streams and become less reliant on hyperscale customers. It’s clear that they need to move away from the bare metal business model they began with and build a more resilient and diversified customer base. One way forward is to leverage their sovereignty – in terms of their own ownership structure, as well as that of their customers’ data and dataflows – to differentiate themselves from their elder siblings.
From latency optimization to processing data within national or regional borders and ensuring transparency and conformance of dataflows: Peering at IXs is a small piece of the transformation puzzle, but it has a big impact. Because, according to ABI Research, by 2030, the market for neocloud-based inference workloads is projected to grow to more than US $150 billion, nearly 300% more than these players will earn from AI training.
