Specialized GPU providers are filling a capacity gap, but their growth exposes a more complicated business: one built on power, capital, contracts and delivery.

On November 10, CoreWeave reported a revenue backlog of approximately $132.5 billion as of September 30. The figure was striking, but the company’s own filing attached an important condition: the backlog represented committed customer contracts, subject to delivery and service-availability requirements. On the same day, an amended Anthropic filing brought that qualification into sharper focus. It disclosed that Anthropic could cancel a reported $34.9 billion compute agreement with Lambda if delivery conditions were not met. The filing did not indicate that Anthropic intended to exercise the right, and neither party had issued a termination notice, according to MLQ.ai’s analysis. But the contract language made the point: a large order is only as secure as the infrastructure and delivery commitments behind it.

Those two developments capture the tension in the neocloud market. Demand for AI computing is enormous, and specialist providers are signing contracts that would have seemed extraordinary only a few years ago. Yet delivering the capacity requires power, land, buildings, networking equipment, GPUs and financing, often on timelines that do not move together. A backlog can signal demand without guaranteeing that the provider can turn every commitment into operating revenue.

Neoclouds are specialist cloud providers built primarily around accelerated computing, especially GPU capacity for AI training and inference. They have emerged alongside the hyperscalers, not because AWS, Microsoft Azure and Google Cloud lack GPU services, but because the scale and urgency of some AI workloads have exposed constraints in the conventional cloud model. The important question is no longer simply whether a neocloud can rent a GPU for less. It is what part of the infrastructure chain it controls, what risks it assumes, and whether its business can remain viable as supply expands.

A market created by a capacity shortage

The label is relatively new. Microsoft for Startups traces the shared use of the term to SemiAnalysis’s AI Neocloud Playbook and Anatomy, published in October 2024. The companies themselves are not all new, but the category gained visibility as demand for large GPU clusters outpaced the capacity available to many customers through established cloud providers.

That shortage has been more than a temporary inconvenience. Microsoft’s guide points to comments by CFO Amy Hood on the company’s April 2026 earnings call that Azure was expected to remain capacity constrained at least through the year. For customers trying to train or serve large models, waiting for capacity can be more costly than paying a premium to secure it elsewhere. Specialist providers have therefore benefited from a practical advantage: they can focus their capital and operations on a narrower set of workloads and bring purpose-built clusters to market.

Market forecasts suggest substantial growth, although their headline figures should be treated carefully. Gartner expects neocloud providers to capture 20% of a $267 billion AI cloud market by 2030. The World Economic Forum cites a $250 billion market forecast for the same year, but the scope behind that figure is not clearly aligned with Gartner’s definition. Estimates of the 2025 starting point also differ: one InfoWorld report, citing Synergy Research Group, puts neocloud revenue above $25 billion, while another InfoWorld article states $42 billion. These figures indicate the scale of expectations, not a single settled market measurement.

The number of providers can also create a misleading impression of market depth. InfoWorld, citing McKinsey, reports that more than 100 neoclouds operate in the United States, but only 10 or 15 are considered to be at meaningful scale. The distinction matters. A company with a few available GPU instances and a provider capable of delivering thousands of accelerators under a long-term contract are not competing on the same terms.

Why enterprises look beyond the hyperscalers

For many organizations, the hyperscaler remains the natural place to run AI. Data may already sit in AWS or Azure; identity, security controls, procurement agreements and operational teams are already established. Keeping workloads close to that environment can reduce integration work and simplify governance. Moving to a specialist provider is not automatically an improvement, particularly when an AI project is closely tied to existing applications and data services.

The case for a neocloud becomes stronger when the workload is dominated by large-scale GPU processing. Training a model across many machines places unusual demands on the network connecting those machines. Data must move rapidly between GPUs, and delays in communication can leave expensive accelerators waiting rather than computing. Specialist providers often build clusters with bare-metal or lightly virtualized access and high-speed fabrics such as InfiniBand or 400G/800G RoCE. A Zettabyte comparison describes configurations with four to eight GPUs per node, while Nutanix highlights the east-west bandwidth requirements of distributed AI training.

This is not to say that hyperscalers lack high-performance AI networking. They invest heavily in specialized fabrics of their own. The distinction is one of emphasis and product design: a broad cloud platform must serve databases, applications, storage, analytics and countless other services, while a GPU-first provider can concentrate on making a smaller number of compute-intensive workloads run efficiently. As Rohan Gupta of R Systems told InfoWorld, the traditional cloud was built around large numbers of smaller, multi-tenant workloads; neoclouds are designed around customers running very large GPU clusters.

Price is often part of the pitch. Industry commentators quoted by InfoWorld have claimed that neocloud compute can cost roughly one-third as much as comparable hyperscaler capacity, or 60% to 70% less. Those comparisons need context: GPU type, utilization, contract length, networking, storage, support and availability all affect the effective price. Providers also have commercial incentives to emphasize the gap. For many buyers, access to capacity at the required time may matter more than a headline hourly rate.

The market is better understood by what each provider controls

It is tempting to treat neoclouds as a single class of cheaper GPU suppliers. That misses the most important differences. Some companies own or develop the sites and power capacity; others operate large GPU clusters; others package compute into platforms that developers can use without managing infrastructure. Still others aggregate hardware from independent hosts. These models overlap, and companies may operate in more than one layer, but the distinctions help explain where the economics and risks sit.

The first layer is physical capacity. A provider needs a suitable site, access to electricity, cooling, buildings and network connectivity before it can install and operate a large GPU cluster. Companies such as Crusoe, Fluidstack and Firmus are associated with power- and infrastructure-led capacity development. In this part of the market, the scarce asset may be a deliverable site with sufficient power, not a software feature. The risks are construction delays, grid connections, lease terms, financing costs and the ability to secure tenants before the investment becomes burdensome.

The second layer is large-scale GPU cloud operations. CoreWeave, Lambda, Nebius, Nscale, IREN and Northern Data are examples of providers associated with substantial GPU capacity and dedicated infrastructure for AI workloads. Their proposition is straightforward: deliver large clusters, often with bare-metal access and fast interconnects, to AI labs, hyperscalers and enterprises that need more capacity than they can obtain elsewhere. The commercial challenge is concentration. A small number of customers and large contracts can support rapid expansion, but missed delivery dates or changes in customer plans can have an outsized effect.

A third group sits closer to the developer. Platforms such as Together AI and Modal emphasize model serving, fine-tuning and access to compute through developer-oriented services. Vultr also offers GPU cloud services within a broader cloud infrastructure business. These platforms can make specialist compute easier to consume: teams can focus on deploying models rather than assembling and operating clusters. But the abstraction does not remove the underlying dependency. Capacity, margins and service quality still rely on access to GPUs, power and upstream infrastructure.

GPU marketplaces and community-cloud models take a different approach. Vast.ai and Runpod connect users with GPU capacity from a wider pool of hosts and providers. This can offer lower-cost or more flexible access, particularly for developers, researchers and smaller teams that do not need a large reserved cluster. The trade-off is that hardware, availability, performance and support can vary by host or offering. Buyers need to understand whether capacity is dedicated, interruptible or backed by a service commitment that fits the workload.

A further group includes smaller independent and silicon-specific providers. TensorWave, Verda, Hyperstack, GMI Cloud, Hot Aisle and WhiteFiber illustrate the range of regional, specialist and alternative-hardware propositions in the market. Some target customers seeking alternatives to a single dominant supplier; others emphasize different accelerator architectures or geographic availability. The category is diverse, and public evidence about the scale and durability of individual providers is uneven. That makes financial, operational and supply-chain diligence especially important.

The dependency chain is the real competitive map

These layers are not independent. A developer platform needs dependable GPU supply. A GPU operator needs a completed facility, power, cooling and network capacity. A site developer needs financing and customers willing to commit to capacity that may take years to deliver. A delay at the physical layer can affect every business above it. The Lambda contract disclosure is a reminder that the availability of a building or a lease can become a condition in a multibillion-dollar compute agreement.

The customer relationships add another complication. The largest neocloud customers can include the hyperscalers that compete with them. InfoWorld reports that Microsoft accounted for a reported 67% of CoreWeave’s 2025 revenue, and cites Futuriom’s view that hyperscalers outsource some GPU capacity to reduce the capital and execution risk of building all of it themselves. That arrangement can be mutually useful: hyperscalers gain access to additional capacity, while neoclouds gain large customers and revenue commitments. But it also creates dependence on a small number of buyers that may have substantial bargaining power and the resources to build or source capacity elsewhere.

The financial exposure is becoming harder to ignore. In an October 14, 2026 release, Gartner warned that cost pressure and a lack of differentiated business models could force more than 80% of neocloud providers to cease independent operations by 2030. The release also said more than $3.5 trillion in AI infrastructure spending could be exposed to impairment. Gartner estimates spending on AI-optimized IaaS will rise from $66 billion in 2027 to $174 billion in 2030. Its webinar summary argues that a strategy of spending cash to acquire scarce compute is sustainable only while revenue growth keeps pace.

That forecast should not be read as a prediction that AI demand will disappear. It is a warning that demand alone does not guarantee attractive returns for every infrastructure provider. If GPU supply improves, customers may gain more choice and pricing power. If utilization falls, hardware becomes outdated or financing costs rise, providers with expensive assets and thin differentiation may struggle. Some businesses may be acquired, merged or absorbed into larger cloud and infrastructure groups rather than disappearing as services.

What enterprise buyers should examine

For enterprise technology leaders, the choice is not simply hyperscaler versus neocloud. It is a workload decision. A specialist provider may be attractive for large, relatively self-contained training jobs, predictable inference demand or projects that need capacity unavailable through an existing cloud agreement. A hyperscaler may remain the better fit when the workload depends heavily on managed data services, identity integration, compliance tooling or close proximity to applications already running there.

Before committing, buyers should test several issues. First, establish exactly what capacity is being promised: GPU model and quantity, interconnect, storage, location, start date and whether the allocation is dedicated. Second, examine the service commitment. What happens if the provider cannot deliver on time, loses access to a facility or cannot maintain the required availability? Third, assess portability. Can models, data and orchestration move to another provider without a costly rebuild? Fourth, understand the provider’s financial position and customer concentration. A low price is less compelling if the supplier’s ability to deliver depends on one facility, one financing arrangement or one major customer.

Data governance also needs attention. Location, access controls, retention, support arrangements and the handling of customer data should be reviewed against the organization’s regulatory and contractual obligations. Buyers should not assume that a specialist GPU provider offers the same breadth of controls, certifications or managed services as a hyperscaler. Those capabilities vary by provider and should be verified directly rather than inferred from the presence of enterprise customers.

The next phase will reward execution, not just access to GPUs

Neoclouds have a genuine opening because AI infrastructure demand has grown faster than conventional supply can satisfy. Their specialization can deliver advantages in cluster design, access and price, particularly for workloads that use large numbers of GPUs together. But the category is not protected by demand alone. It is capital intensive, exposed to rapid hardware cycles and dependent on a chain of physical assets, contracts and customer commitments.

The providers best placed to endure will need more than access to scarce GPUs. They will need reliable power, facilities that come online when promised, high utilization, durable customer relationships and a reason for buyers to stay when capacity becomes easier to find. For enterprises, the opportunity is real, but so is the need to look beyond advertised GPU rates. In this market, the most important question may not be who has the fastest chip today. It may be who can still deliver the promised compute when the contract comes due.

How the neocloud market is organized

These categories describe different positions in the supply chain, not mutually exclusive company types. Some providers span more than one layer.

Market position What the provider supplies Typical buyer Examples Main consideration
Power and site development Land, power access, facilities, cooling and the infrastructure needed before GPU capacity can be installed. Hyperscalers, AI labs and large compute buyers. Crusoe, Fluidstack, Firmus Construction, grid access, leases, financing and tenant commitments.
Large-scale GPU clouds Dedicated GPU clusters, bare-metal access and high-speed interconnects for training and inference. AI labs, hyperscalers buying overflow capacity and large enterprises. CoreWeave, Lambda, Nebius, Nscale, IREN, Northern Data Large contracts and customer concentration make delivery and utilization critical.
Inference and developer platforms Model serving, fine-tuning and developer-facing access to GPU compute. AI product teams, startups and enterprises running models. Together AI, Modal, Vultr Ease of use depends on reliable underlying GPU supply and sustainable margins.
GPU marketplaces and community clouds GPU capacity aggregated from multiple hosts, often with flexible or lower-cost access. Developers, researchers and early-stage companies. Vast.ai, Runpod Availability, performance, support and interruption risk can vary.
Independent and silicon-specific providers Regional capacity, alternative accelerator options and specialist cloud services. Teams seeking supplier diversity, regional capacity or non-NVIDIA hardware. TensorWave, Verda, Hyperstack, GMI Cloud, Hot Aisle, WhiteFiber Scale, financing, hardware roadmap and long-term service durability require diligence.

What to Expect

The next few years will likely bring consolidation as providers face the cost of infrastructure, financing, power availability, and increasingly competitive GPU markets. Some will be acquired or absorbed into larger platforms, while others will focus on specialized hardware, regional capacity, or developer services.

For enterprise buyers, this could mean more choice and better pricing, but also greater scrutiny of provider stability. GPU availability alone is not enough. Delivery commitments, infrastructure readiness, financing, and long-term service reliability will become increasingly important measures of a provider’s value.

Closing Thoughts

Neoclouds are unlikely to replace hyperscalers. Their opportunity is to address the specialized compute requirements that traditional cloud platforms cannot always meet quickly or economically. But the market’s long-term winners will be determined by more than GPU capacity and ambitious contracts. They will need reliable infrastructure, sustainable economics, and the ability to deliver on what they promise.

The real test for neoclouds is no longer how much capacity they can announce, but how reliably they can turn it into a sustainable business.