Why cheaper neocloud compute does not always win enterprise customers

Neoclouds can win AI compute without winning the enterprise account

A company can see substantial savings in AI infrastructure and still make a sound decision to keep its existing cloud provider. Different Day faced that choice and stayed with its hyperscalers, the large cloud platforms that supply computing alongside a broad range of managed services. Its decision shows why computing capacity is only part of the purchase: an enterprise also needs data access, identity controls, security, procurement agreements, and people who can operate the application.

Neoclouds specialize in usage-based access to graphics processing units, or GPUs, and infrastructure for AI training and inference, the process of running a trained model. CoreWeave, Lambda, Nebius, Runpod, and Vultr offer alternatives to Amazon Web Services, Google Cloud Platform, and Microsoft Azure. Nii Osae, CEO and founder of AI-services company Mindbeam AI, says neoclouds combine NVIDIA GPUs with high-bandwidth storage and networking so research laboratories and enterprises can train and run models reliably. Mindbeam AI sells services into this market and benefits when customers can get infrastructure for AI projects.

Demand for that infrastructure is substantial. Synergy Research Group reports that neocloud revenue exceeded $25 billion in 2025, while Gartner forecasts that neoclouds could capture 20% of a $267 billion AI cloud market by 2030. Global consultancy McKinsey & Company counts more than 100 neoclouds worldwide, with 10 or 15 operating at “meaningful scale” in the United States. These figures describe a growing infrastructure business, but some demand comes from hyperscalers themselves, so revenue can grow while hyperscalers keep the enterprise accounts.

Specialized infrastructure improves the economics of AI workloads

For teams waiting to train a large model, neoclouds offer lower capacity costs and faster access. Lauri Kien Kotcher, CEO and co-founder of AI application-development company Different Day, says neocloud capacity can cost approximately one-third of equivalent hyperscaler GPU capacity and become available within days, compared with months quoted for high-density AI infrastructure at hyperscalers. ComputerWeekly also reports savings of 60% to 70% for neocloud GPU instances compared with traditional hyperscaler instances. Buyers have a reason to evaluate specialized capacity, although the hourly GPU charge covers only part of a workload’s cost.

Access to NVIDIA helps explain the pricing advantage. Michael Byrne, vice president of data center solution architecture at digital-services and infrastructure-solutions company Presidio, says many neoclouds have close relationships with the chipmaker that help keep GPU-hour costs low. Presidio sells infrastructure solutions and related services, giving it a commercial interest in customers’ infrastructure decisions. Byrne says those supplier relationships become especially valuable during GPU shortages, when securing capacity determines how soon a project can begin.

Once capacity is available, the connections among GPUs affect how much work it can do. Hardeep Singh, senior principal analyst at Gartner, describes neoclouds as offering NVIDIA-oriented infrastructure, high-performance GPU clusters, flexible contracts, InfiniBand networking, and direct hardware access. InfiniBand is a high-speed connection technology that lets distributed GPUs exchange data with little delay. A training run spread across thousands of GPUs depends on those exchanges to keep the processors working together.

Coordinating work at that scale requires a different approach from earlier cloud designs. Rohan Gupta, vice president of cloud, security, and devops at digital product engineering and technology-services company R Systems, contrasts the internet of 2010 with large AI training jobs. Earlier systems served millions of small virtual machines over Ethernet networks whose combined demand could exceed available bandwidth; a training job can place thousands of communicating GPUs under one customer’s control. R Systems sells technology services affected by these architecture choices, and Gupta argues that coordinated training requires different resource allocation and networking from many separate applications.

Direct hardware access addresses another part of the performance problem. John Q. Martin, technology partner and alliances manager at database devops software company Redgate Software, says bare-metal access, direct use of physical servers, removes the overhead of a hypervisor, the software layer that manages virtual machines. Redgate supplies software for database development and operations and also buys cloud capacity for its own AI back end. Martin says removing that layer can make more GPU performance available for large-scale AI and machine-learning training, generative AI platforms, and high-performance computing, provided the workload keeps the hardware busy.

Keeping GPUs busy also requires storage that can feed them quickly enough. Gupta says neoclouds adopted parallel filesystems from VAST, WEKA, and DDN early because large training jobs need coordinated, high-bandwidth data access across a cluster. A parallel filesystem lets multiple machines read and write through shared storage simultaneously. Gupta argues that this arrangement serves large training jobs better than a design centered on object storage, which stores data as individually addressed objects.

The result of those infrastructure choices is what David Johnson, director of product marketing at cloud-storage company Backblaze, calls “goodput”: the share of capacity that becomes useful training or inference work after failures, restarts, and input/output stalls, when processors wait for data transfers. Backblaze sells storage used alongside AI compute and benefits when customers invest in the data infrastructure needed to keep GPUs productive. Johnson’s measure brings storage and networking into the cost comparison, because a cheap GPU-hour loses value when the job spends it waiting or repeating work.

These requirements continue when AI moves from training into applications. Kevin Cochrane, CMO at Vultr, says specialized AI services and infrastructure can extend existing cloud-native applications, built to run in cloud environments, with AI agents, software that uses models to carry out tasks. Vultr sells the infrastructure he advocates and competes with hyperscalers for that business. Haseeb Budhani, CEO and co-founder of cloud and AI infrastructure-management platform Rafay Systems, adds that teams need fast storage, low-delay networking, secure access, shared governance, and usage metering. Rafay benefits from demand for managing those resources, and Budhani expects inference to receive most enterprise AI spending.

Enterprise operations can outweigh cheaper compute

The need to operate a complete application helps explain Redgate’s purchasing decision. Although Martin recognizes the performance value of bare-metal GPUs, Redgate chose a hyperscaler for its AI back end. “At our current stage, the managed services, ecosystem integrations, and ability to scale incrementally made more sense than committing to dedicated GPU infrastructure,” he says. Those services let Redgate increase usage within an established operating environment, giving them a business value that outweighed dedicated capacity at that stage.

Different Day also kept its hyperscalers despite the cost and provisioning advantages Kotcher identified. Its decision reflected continuity with existing technology, security certifications, identity controls, audit records, and procurement contracts. Kotcher argues that most neoclouds have yet to reach comparable maturity in those areas. For her company, moving compute would also mean preserving the controls and agreements that determine who can use the service, how activity is recorded, and how the organization buys it.

Moving compute can also separate an application from its data. Martin says that when GPUs run on a neocloud while data remains with one or several hyperscalers, transfers can incur egress charges, the fees providers charge when data leaves their environments. He also identifies fragmented security and identity management, divided operating responsibility, and additional overhead. The engineering team must keep access policies and operating processes aligned across providers, so capacity savings have to cover both transfer costs and the work of maintaining those connections.

Bare-metal access can transfer further operating responsibility to that team. Martin says it can come with fewer ready-made provisions for automatic scaling, observability, the ability to understand system behavior from operational information, and enterprise service-level agreements, or SLAs, which define service commitments. Direct access therefore improves control over execution while changing what the buyer must manage. Its value depends partly on whether the organization has the people and processes to handle that work.

The responsibilities grow when several teams share the service. Budhani calls enterprise maturity the largest trade-off and points to the years hyperscalers have spent developing governance, security, identity, auditing, billing, and operational controls. These functions let teams use a common platform within organizational rules. For Rafay, the need to coordinate such controls also creates demand for the infrastructure-management capabilities it sells.

Shared controls become more consequential as AI enters a larger product or business process. Scott Sanders, corporate vice president of engineering at code-quality and security company Sonar, says hyperscalers retain advantages in service breadth and integration. Sonar sells tools used in software development, so its business operates within the workflows affected by those integration choices. Sanders argues that hyperscalers provide continuity when AI workloads sit inside a wider enterprise architecture, where changing compute providers can affect connections to surrounding systems.

For essential or regulated applications, the provider’s ability to sustain those connections becomes another purchasing requirement. Nigel Gibbons, director and senior advisor at cybersecurity consultancy NCC Group, identifies portability difficulties, additional operating work, and dependence on specialized providers as risks. NCC Group sells cybersecurity advice, giving it a commercial interest in how organizations assess and manage such risks. Gibbons says buyers must weigh performance and cost against resilience, support depth, compliance coverage, and the provider’s ability to remain viable over time.

Existing deployment patterns show why these decisions often involve several environments. Redgate’s 2026 State of the Database Landscape found that 43% of respondents run hybrid database estates, meaning their databases span different deployment environments. The finding concerns database deployment rather than neocloud purchasing, but it shows the setting into which an AI compute decision can arrive. Buyers may already have data access and operating arrangements spread across locations before they add another provider.

AI hosting choices reveal a further range of options. A10’s The State of AI Infrastructure Report 2025 found that 35% of respondents primarily hosted AI in public-cloud environments, while 42% favored a balanced hybrid approach combining deployment environments. These percentages address a different question and respondent population from Redgate’s survey. They place neocloud purchasing within a broader choice that also includes on-premises and hybrid strategies.

Hyperscalers also buy neocloud capacity

The distinction between buying compute and moving an enterprise account becomes especially clear when the buyer is a hyperscaler. “The largest neocloud customers right now are the hyperscalers themselves,” Gupta says. In that arrangement, a neocloud supplies infrastructure while the hyperscaler can retain the relationship with the organization using the AI service. The providers therefore compete for some customers while buying capacity from one another.

Microsoft’s relationship with CoreWeave illustrates that overlap. One report attributes 67% of CoreWeave’s 2025 revenue to Microsoft, and OpenAI is also identified as one of CoreWeave’s leading clients. CoreWeave’s 2025 annual SEC filing states: “We recognized an aggregate of approximately 77% of our revenue from our top two customers for the year ended December 31, 2024.” The filing leaves those customers unnamed, while showing how much revenue depended on a small group of buyers.

That dependence creates commercial risk alongside demand. CoreWeave’s outstanding debts and reliance on a few large customers have raised questions about its long-term sustainability. Large buyers can keep infrastructure occupied while also gaining substantial influence over a supplier’s business. A hyperscaler customer adds a further complication because it can compete for the enterprise business the neocloud wants to win.

IT consultancy Futuriom examines these dependencies in AI, GPU Clouds, and Neoclouds in the Age of Inference. Its report describes interconnected relationships among NVIDIA, OpenAI, and major neoclouds including CoreWeave. Futuriom argues that hyperscaler investment and outsourcing of GPU infrastructure reduce exposure on hyperscaler balance sheets. In that account, hyperscalers obtain capacity while other providers carry some of the infrastructure investment and associated risk.

The commitments extend to the storage suppliers that support GPU services. Backblaze announced a five-year, $335 million agreement with CoreWeave, illustrating both the storage investment around AI compute and Backblaze’s direct commercial stake in neocloud growth. Such agreements help a neocloud assemble the resources needed to deliver its service. They can grow the infrastructure market while leaving enterprise applications and purchasing relationships with hyperscalers.

Winning broader enterprise business requires continuing investment

Selling infrastructure to hyperscalers leaves room for neoclouds to develop a broader enterprise offering. Writing for Forbes, Scott Raynovich, founder and chief analyst of Futuriom, predicts that neoclouds will expand beyond GPU rental into worldwide infrastructure, application services, data privacy, and security. Those capabilities would address requirements that influenced Redgate and Different Day’s decisions. A provider supplying them could take responsibility for more of an application’s operating environment.

Taking on that responsibility requires investment beyond acquiring GPUs. Martin says a competitive AI-native software offering demands substantial capital, engineering expertise, and development time. A neocloud must fund that work while maintaining relationships with large customers that may also compete for the broader enterprise account. Its infrastructure business can support expansion, but those customer relationships also shape how it competes.

The equipment itself adds a continuing funding requirement. Futuriom says GPUs can depreciate over four to five years, limiting how long an initial hardware investment can sustain an infrastructure-led advantage. Providers developing software and enterprise services must therefore manage assets that lose value relatively quickly. Demand for current GPUs can fund growth while successive hardware cycles require further spending.

Hyperscaler investment in custom chips puts additional pressure on that hardware model. Gupta points to AWS’s Trainium, Google’s Ironwood, and Microsoft’s Maia, arguing that they give hyperscalers cost and scale advantages over a model centered on NVIDIA capacity. Close alignment with NVIDIA supports neocloud access and optimization today, while alternative chips can change the economics of workloads that can use them. So buyers need to compare the useful work delivered by the available hardware with the services and operating responsibilities attached to it.

Those differences may produce a lasting division of workloads. Johnson sees leading neoclouds serving AI workloads for which hyperscaler economics become unfavorable while complementing the wider cloud portfolio. Gupta forecasts that neoclouds will retain frontier-model training, the development of leading-edge models, and NVIDIA-linked inference, while hyperscalers retain the data plane, meaning the systems that store, move, and serve application data, along with regulated workloads, custom-chip inference, and enterprise commercial relationships. Kotcher also expects neocloud success to remain bounded; whether providers extend beyond those roles will depend on their ability to fund and operate the surrounding enterprise services.

Key executive takeaways

• Measure useful compute per dollar: Infrastructure buyers evaluating neoclouds need to measure completed training or inference work, including time lost to failures and data-transfer delays. Lower GPU prices and faster provisioning deliver value when storage, networking, and utilization keep the hardware productive.
• Price the full operating environment: Platform and security teams need to include data egress, identity controls, compliance, support, and additional operating work in provider comparisons. Redgate and Different Day retained hyperscalers because existing services and controls outweighed specialized compute advantages.
• Assess customer concentration: Procurement teams evaluating neocloud providers need to examine their dependence on major customers, debt, and long-term support capacity. Hyperscalers supply substantial demand while also competing for enterprise accounts, creating growth opportunities and commercial risks for neoclouds.
• Match providers to workload requirements: CTOs can evaluate neoclouds for demanding training and NVIDIA-based inference while retaining hyperscalers where integrated services and governance carry greater value. Broader neocloud adoption depends on sustained investment in enterprise capabilities alongside hardware renewal.