For years, the enterprise infrastructure decision was relatively easy to map, even when the decision itself wasn't easy. An organisation could own and operate infrastructure itself. It could place hardware in a colocation facility. Or it could move workloads into the cloud and buy capacity from one of the large hyperscalers.
There were variations within each model, but the basic categories were familiar. AI is making those lines harder to draw. A growing group of neoclouds and specialist AI infrastructure providers is building cloud capacity specifically around accelerated computing.
Rather than trying to support every kind of enterprise workload, these providers are concentrating on the GPUs, high-performance networking and other infrastructure needed to train and run AI models. And they're growing quickly. Synergy Research Group found that neocloud revenues exceeded $25 billion in 2025, after reaching $9 billion in the fourth quarter alone.
It expects the market to approach $400 billion by 2031. More recent figures suggest the infrastructure behind that growth is still expanding. Dell'Oro Group reported in September 2026 that neocloud and AI model builder capital expenditure grew faster than any other customer segment during the second quarter.
For enterprises, this creates another way to source the infrastructure behind AI. It also makes choosing where those workloads should run considerably more complicated.
AI Infrastructure No Longer Fits Neatly Into The Traditional Cloud Market
A neocloud is essentially a cloud provider built around AI and other high-performance computing workloads. Instead of offering the huge catalogue of general-purpose services associated with a hyperscaler, its infrastructure is designed around the particular demands of accelerated computing. The category is becoming difficult to dismiss as a niche part of the cloud market.
Dell'Oro has added a separate AI-specialised cloud segment to its infrastructure forecasts, covering neocloud providers and AI model builders. It expects this segment to grow at nearly 60 per cent CAGR through 2030. Gartner now has a Cloud AI Infrastructure market that evaluates traditional hyperscalers alongside specialist providers such as CoreWeave, Crusoe, Lambda, Nebius and Nscale.
But this isn't creating a simple contest between old cloud and new cloud. Neoclouds may compete with hyperscalers for some workloads while supplying capacity to them for others. Model builders can consume infrastructure from several providers. Specialist providers can operate infrastructure inside third-party data centres.
And meanwhile the hardware itself may come from the same small group of accelerator suppliers. The result is less a new layer sitting neatly between enterprises and hyperscalers, and more a cloud market becoming increasingly specialised around AI. That gives infrastructure teams more choice. It doesn't necessarily make the choice easier.
More Infrastructure Choice Changes The Sourcing Decision
The temptation is to treat neoclouds as an alternative to hyperscale cloud. For some workloads, they may be. But that framing assumes every AI workload needs roughly the same thing from its infrastructure. They don't. Training a large model can demand enormous amounts of tightly connected compute for a concentrated period.
Fine-tuning may need far less. Production inference has another set of priorities because models are serving requests from applications, employees or customers, often alongside data and services that already live elsewhere. That changes what "best infrastructure" means. A specialist provider may offer attractive access to high-performance GPU clusters.
But an enterprise running inference close to data already stored in a hyperscale cloud could find that integration, latency and data movement are more important than access to the most specialised cluster. Gartner expects this distinction to become more important as inference grows.
Its 2026 AI-optimised infrastructure forecast argues that demand will increasingly be shaped by proximity to enterprise data and integrated platforms rather than GPU access alone. It expects hyperscalers to retain an advantage here, while neoclouds remain particularly relevant for specialised training and high-performance workloads.
So AI infrastructure sourcing doesn't have to become an either-or decision. Enterprises may own infrastructure where control or predictable utilisation justifies it, use hyperscale cloud where AI needs to sit close to existing applications and data, and turn to specialist providers where performance, capacity or particular hardware requirements make them a better fit.
The workload needs to lead the decision. The provider category comes afterwards.
The GPU Isn't The Infrastructure Decision
GPU availability is easy to compare. So is the hourly price attached to it. Neither tells an enterprise very much about the quality of the infrastructure it is actually buying. A GPU only becomes useful compute when the surrounding system can keep it working effectively. Storage has to feed data quickly enough.
Networks need to connect large clusters without creating bottlenecks. The software layer needs to make the capacity manageable. Then the provider has to operate the whole environment reliably enough for production workloads. NVIDIA's own requirements for AI cloud providers make the scale of this problem unusually clear.
Its latest requirements extend beyond hardware reference designs to the full stack of infrastructure and operational capabilities needed to provide AI compute. NVIDIA says it has published them partly as an industry reference for the capabilities a large GPU consumer should expect.
That gives procurement teams a more useful starting point than asking which provider has the right GPU. The same applies to price. The State of FinOps 2026 report found that 98 per cent of respondents now manage AI spending, up from 31 per cent two years ago.
Yet practitioners continue to struggle with visibility because AI pricing can vary widely between providers and services. A cheap GPU can become expensive infrastructure if utilisation is poor, data movement is costly or the organisation needs more engineering work to integrate and manage the environment.
Once neoclouds are treated as infrastructure providers rather than GPU suppliers, the due diligence starts to look very different.
Enterprise Due Diligence Has To Extend Beyond Performance
Performance still counts. Enterprises are buying specialist infrastructure partly because they need it to perform well. But the real question is whether that performance can become a dependable part of the wider technology estate. That means evaluating both the infrastructure being offered and the organisation expected to keep providing it.
Does the infrastructure fit the workload?
Start with what the workload actually needs. Training, fine-tuning and inference place different demands on compute, storage, networking and data. Some workloads need enormous clusters for relatively concentrated periods. Others need consistent capacity close to applications and enterprise data.
The provider should therefore be able to demonstrate more than access to the desired accelerator. Infrastructure teams need evidence that the wider architecture can deliver the performance, availability and integration the workload requires in production. This also means understanding what the organisation will need to change around it.
A technically impressive platform becomes less attractive if using it introduces disproportionate operational complexity elsewhere.
What will the workload actually cost?
Hourly GPU pricing can provide a useful comparison point, but it isn't a total cost. The real AI infrastructure cost includes how efficiently the organisation can use the capacity it buys. Storage, networking, data transfer and reserved capacity can change the economics. So can the engineering effort required to operate another infrastructure environment.
Commitment models deserve particular attention. Capacity that looks economical at high utilisation can become far less attractive if workloads are unpredictable and the enterprise is paying for resources it can't use. Infrastructure and FinOps teams therefore need a common view of the workload before comparing providers.
Otherwise, organisations risk comparing individual prices rather than the cost of producing the same business outcome.
Where will the workload run and who controls it?
A provider can have excellent infrastructure in entirely the wrong place. Geographic footprint affects latency, resilience and connectivity, but AI adds another consideration: control over where data and workloads are processed.
Gartner identifies sovereignty as one area where some neocloud providers are differentiating themselves, particularly through infrastructure that keeps data, operations or governance within defined jurisdictions. That can make specialist infrastructure attractive where local control is important. But "available in Europe" or "sovereign cloud" isn't enough on its own.
Enterprises still need to understand where the infrastructure physically operates, how data moves through it, what legal entities control it, what security controls apply and what happens if a region or facility becomes unavailable. Geography only becomes an advantage when it matches the workload's actual requirements.
Can the provider sustain the dependency?
This may be the least familiar part of the evaluation. Building AI infrastructure requires enormous amounts of capital, and specialist providers are finding increasingly sophisticated ways to finance it. Lambda, for example, closed a $926 million secured loan in August 2026 to fund GPU infrastructure for a committed customer.
The financing is secured against the GPU servers, related infrastructure and the cash flows they generate. The scale can become much larger. CoreWeave reported $35.6 billion in total indebtedness as of 30 June 2026. Its filing also showed significant customer concentration, with three customers accounting for 36, 26 and 10 per cent of quarterly revenue respectively.
Those figures shouldn't be read as a judgement on either provider. They illustrate how different the economics behind specialist AI infrastructure can be. An enterprise depending on a provider for important workloads needs to understand the business behind the capacity:
- How is expansion financed?
- How concentrated is demand?
- How dependent is the provider on particular hardware suppliers?
- Does it own infrastructure or depend heavily on other operators?
- And can it continue investing at the pace its commitments require?
At that point, counterparty risk starts becoming part of infrastructure risk.
How difficult would it be to leave?
Every sourcing decision looks different when viewed from the exit. Moving a workload can mean transferring large datasets, rebuilding integrations, changing orchestration tools and finding equivalent capacity somewhere else. Long-term commitments can make that technically possible but commercially painful.
The time to understand those constraints isn't when the relationship stops working. Infrastructure teams need to know how portable workloads are before committing them. That means examining data egress, contractual terms, interoperability, alternative capacity and how tightly the workload depends on provider-specific services.
Not every workload needs to be instantly portable. Some performance or cost benefits may justify deeper dependence. But that dependency should be a conscious architecture decision rather than something discovered later.
Neoclouds Should Be Evaluated As Infrastructure Partners
None of this makes neoclouds inherently riskier or safer than hyperscalers, colocation or owned infrastructure. It makes them another infrastructure model with a different balance of strengths and dependencies. For enterprises struggling to access specialised compute, a neocloud may provide capacity that would be difficult or slow to build internally.
For demanding training workloads, an environment designed around accelerated computing may offer advantages over general-purpose infrastructure. Sovereign providers could also create options in markets where local control is becoming more important.
Other workloads may remain better suited to hyperscale cloud because their data, applications and operational tooling already live there. Predictable workloads could make owned infrastructure attractive. Many enterprises will probably end up using some combination of all three.
So neocloud isn't a procurement recommendation. It's a provider category. The more useful questions sit underneath the label:
- What infrastructure is the enterprise actually getting?
- Does it suit the workload?
- What will it cost to operate?
- Where will data and compute reside?
- Who is the organisation becoming dependent on?
- And how easily can that dependency change later?
Those questions can survive whatever terminology the market eventually settles on.
Final Thoughts: AI Infrastructure Sourcing Starts With The Workload
Enterprise infrastructure used to fit into relatively familiar boxes. AI hasn't removed those choices. It has added new ones and blurred some of the boundaries between them. That will probably continue as hyperscalers, neoclouds, model builders, colocation providers and infrastructure investors build, buy and share more of the physical capacity behind AI.
For infrastructure leaders, trying to decide which provider category will ultimately "win" isn't especially useful. The more practical job is understanding which infrastructure dependencies make sense for which workloads. Start there and the provider conversation becomes clearer. A specialist AI cloud doesn't need to replace the hyperscaler to be valuable.
Nor does access to the newest GPU automatically make it the right home for an enterprise workload. What counts is whether the infrastructure, economics, control and provider relationship fit what the organisation is actually trying to run. As those choices continue to evolve, EM360Tech will keep examining how new infrastructure models are changing the way enterprises build, source and operate the technology underneath AI.
Comments ( 0 )