How neoclouds can win AI compute without replacing hyperscalers

Neocloud revenue exceeded $25 billion in 2025, according to Synergy Research Group, and Gartner predicts neoclouds could capture 20% of the $267 billion AI cloud market by 2030. That makes neoclouds economically significant. For enterprise leaders, the separate question is how much of the wider cloud relationship will move with AI compute.

For CIOs and CTOs, the distinction matters because AI compute is only one part of enterprise cloud architecture. Data, applications, identity, security, governance, procurement, managed services, and operating processes can remain with a hyperscaler when GPU workloads run elsewhere. A neocloud can win AI infrastructure business while a hyperscaler retains other parts of the enterprise relationship.

Neocloud growth reflects specialized AI demand

More than 100 neoclouds operate today, although McKinsey & Company says only 10 or 15 operate at a “meaningful scale” in the United States. Providers include CoreWeave, Lambda, Nebius, RunPod, and Vultr.

Nii Osae, CEO and founder of Mindbeam AI, an AI services company with a commercial interest in AI infrastructure adoption, describes neoclouds as “key components of the AI infrastructure ecosystem,” providing vertically integrated GPUs alongside high-bandwidth storage and networking for training and inference.

Kevin Cochrane, CMO at Vultr, argues that neoclouds can extend cloud-native compute with specialized infrastructure and AI services for agentic applications. Vultr is itself a neocloud provider and benefits commercially from demand for this model.

These claims define a narrower competitive boundary than the whole enterprise cloud account. A specialized provider can supply AI compute while other infrastructure and controls remain elsewhere.

Neoclouds are designed around dense AI workloads

Rohan Gupta, vice president of cloud, security, and devops at R Systems, a digital product engineering and technology services company with commercial exposure to cloud and infrastructure projects, describes the architectural distinction this way: “Hyperscalers were designed for the internet of 2010: millions of small, multi-tenant virtual machines (VMs) over oversubscribed Ethernet. Neoclouds were designed for a workload that didn’t exist back then: a single tenant, thousands of GPUs, all talking to each other at the same time.”

Hardeep Singh, senior principal analyst at Gartner, says neoclouds combine NVIDIA-native infrastructure with InfiniBand networking, bare-metal access, flexible contracts, and lower costs. Lauri Kien Kotcher, CEO and co-founder of Different Day, an AI application development company with commercial exposure to AI infrastructure choices, says neoclouds can run at roughly a third of hyperscaler cost for the same GPU capacity and provision capacity in days rather than the months hyperscalers may quote for high-density AI infrastructure. ComputerWeekly puts potential savings for neocloud GPU instances at 60% to 70% compared with traditional hyperscaler GPU instances.

Michael Byrne, vice president of data center solution architecture at Presidio, a digital services and infrastructure solutions provider with commercial exposure to infrastructure spending, says many neoclouds have developed strong NVIDIA relationships and favorable raw GPU-hour economics.

The architecture around the GPU matters too. John Q. Martin, technology partner and alliances manager at Redgate Software, a database devops software company, says bare-metal access removes hypervisor overhead, while high-bandwidth interconnections provide the communication needed for distributed training.

Gupta says neoclouds adopted parallel filesystem architectures from VAST, WEKA, and DDN early because large-scale training puts parallel demands on storage. Accelerator utilization therefore depends partly on how quickly storage can feed data to the GPUs.

David Johnson, director of product marketing at Backblaze, a cloud storage provider that sells infrastructure used with AI workloads, uses “goodput” for the share of AI infrastructure capacity that produces useful training or inference after accounting for failures, restarts, and I/O stalls. Backblaze recently announced a five-year $335 million deal with CoreWeave, giving the company a direct commercial relationship with a major neocloud.

Haseeb Budhani, CEO and co-founder of Rafay Systems, provider of a cloud and AI infrastructure management platform that can benefit from demand for managing AI infrastructure, points to requirements beyond accelerators. He says enterprises also need high-speed storage, low-latency networking, secure access, governance, and usage metering across teams. He expects these requirements to matter particularly for inference, where he says most enterprise AI spending will land.

Enterprise placement depends on the surrounding system

Redgate shows how these wider requirements can change a placement decision. “We went with a hyperscaler for our AI back end,” Martin says. At Redgate’s current stage, Martin says managed services, ecosystem integrations, and incremental scaling outweighed a commitment to dedicated GPU infrastructure.

Redgate’s 2026 State of the Database Landscape found that 43% of respondents operate hybrid database estates. Martin identifies data-egress costs, fragmented security and identity management, and split operational overhead as potential consequences when AI compute runs on a neocloud while data remains elsewhere. He also says bare-metal GPU access can have fewer built-in controls around autoscaling, observability, and enterprise service-level agreements than hyperscalers provide.

Kotcher describes a related decision at Different Day. He says staying with a major cloud account preserves security certifications, identity controls, audit logs, and procurement contracts already associated with that environment. Different Day therefore stayed with hyperscalers despite Kotcher’s assessment that neocloud GPU economics can be significantly better.

Budhani calls the wider issue “enterprise maturity.” He says hyperscalers have spent years building governance, security, identity, auditing, billing, and operational controls that enterprises have incorporated into their architecture and processes.

Scott Sanders, corporate vice president of engineering at Sonar, a code quality and security company, says hyperscalers retain an advantage in breadth and integration when AI workloads sit inside a larger enterprise environment.

Nigel Gibbons, director and senior advisor at NCC Group, a cybersecurity consulting firm with commercial exposure to enterprise security and resilience work, identifies operational overhead, reliance on niche providers, and portability challenges as further concerns. For regulated or mission-critical systems, he says buyers need to consider support depth, compliance coverage, resilience, and long-term vendor viability.

A10’s The State of AI Infrastructure Report 2025 found that 35% of respondents use primarily public cloud infrastructure, while 42% use a balanced hybrid approach. Hybrid arrangements let organizations place different parts of an AI system in different environments when their technical and operating requirements differ.

Neocloud demand can come through hyperscalers

The relationship becomes more complex when hyperscalers buy capacity from neoclouds. Gupta says, “The largest neocloud customers right now are the hyperscalers themselves.”

CoreWeave provides a concrete example of customer concentration. In its 2025 annual SEC filing, CoreWeave disclosed: “We recognized an aggregate of approximately 77% of our revenue from our top two customers for the year ended December 31, 2024.”

Neocloud revenue can therefore reach the market through more than one route. Gupta says some capacity goes to hyperscalers themselves rather than to enterprises contracting directly with the specialist provider.

IT consultancy Futuriom describes such arrangements in AI, GPU Clouds, and Neoclouds in the Age of Inference as a response to dependencies involving NVIDIA, OpenAI, and major providers such as CoreWeave. Futuriom argues that investment in and outsourcing of GPU infrastructure can reduce risks on hyperscaler balance sheets.

Customer concentration creates a strategic issue for providers as well. Martin says building an AI-native software stack requires large amounts of capital, engineering talent, and time. Some of the biggest customers, he adds, may be hyperscalers developing competing capabilities. Long-term provider viability therefore matters to enterprise workload placement.

Competition is expanding above the GPU layer

Futuriom estimates that GPUs can depreciate in as little as four to five years. Providers that own large GPU fleets must therefore balance utilization of existing hardware against investment in newer generations.

Writing for Forbes, Futuriom founder and chief analyst Scott Raynovich argues that these companies can expand beyond GPU rental into global infrastructure, application services, data privacy, and security. Futuriom has a commercial interest in analysis and consulting around infrastructure markets, so this is a market thesis rather than a settled outcome.

Hyperscalers are investing in specialized AI infrastructure too. Gupta points to AWS Trainium, Google Ironwood, and Microsoft Maia as technologies that can give hyperscalers cost and scale advantages, particularly against providers built heavily around NVIDIA hardware.

The competitive boundary is widening. Raynovich says neocloud providers are adding services around specialized compute, while Gupta says hyperscalers are developing specialized silicon of their own. Enterprise buyers need to compare the economics and operating requirements of the complete workload rather than GPU rental rates alone.

Plan workload placement around separation costs

For CIOs and CTOs, the procurement boundary can sit inside a workload. Training runs or inference services with high GPU density, heavy interconnect traffic, and clean interfaces to enterprise systems fit the technical characteristics that Osae, Gupta, Singh, Martin, and other speakers associate with neocloud infrastructure. Workloads tightly connected to existing databases, identity, governance, compliance, application services, and operating processes bring additional integration requirements.

Kotcher describes the potential outcome as a “bounded victory,” in which neoclouds win a real market without taking the broader one. Johnson similarly says leading neoclouds are building a complementary layer for AI workloads where hyperscaler economics become unattractive. Backblaze’s commercial relationship with CoreWeave gives Johnson a stake in growth in this infrastructure model.

Gupta expects the boundary to settle around frontier training and NVIDIA-tied inference for neoclouds, with hyperscalers retaining the data plane, regulated workloads, custom-silicon inference, and enterprise commercial relationship. For an enterprise buyer, the practical question is which AI compute can be separated cleanly enough for specialized infrastructure to deliver an advantage after accounting for data movement, security, governance, resilience, and operations.

Key executive takeaways

• Specialized AI demand is driving neocloud growth: Neoclouds can win AI compute without displacing hyperscalers across the wider enterprise cloud estate. CIOs should treat specialized compute and the broader cloud relationship as separate sourcing decisions.
• AI-focused architecture can improve compute economics: Bare-metal GPUs, high-bandwidth networking, and parallel storage can improve utilization and lower costs for dense training and inference workloads. Compare total workload economics rather than GPU-hour pricing alone.
• Enterprise integration can outweigh GPU savings: Data movement, identity, security, governance, observability, and procurement requirements can make hyperscalers a better fit for tightly integrated workloads. Factor these separation costs into placement decisions.
• Hyperscalers can benefit from neocloud growth: Hyperscalers can source capacity from neoclouds, making the two models both competitors and partners. Enterprises should assess provider concentration and long-term viability alongside performance and price.
• Competition is moving beyond GPU rental: Neoclouds are expanding into services while hyperscalers invest in specialized AI infrastructure and custom silicon. Buyers should evaluate each provider’s broader platform and roadmap.
• Workload boundaries should drive placement: Neoclouds are strongest where GPU-intensive workloads can be separated cleanly from enterprise systems. CIOs should balance compute savings against data movement, security, governance, resilience, and operational complexity.