Neocloud providers are the fastest-growing category of cloud infrastructure on record, and yet most enterprise architecture teams still leave them out of their AI plans. That is the puzzle David Linthicum sets out in “Neoclouds and the enterprises that need them”, a column InfoWorld published on 9 October 2026. The revenue is real, the growth is extreme, and the buyers who should care most are not hearing a coherent story about where these GPU clouds fit.
Linthicum’s diagnosis is blunt. The vendors grew by selling capacity to other technology companies, often in circular deals, and they have not learned to explain themselves to the architects who design enterprise systems. If that does not change, the enterprise AI budget goes to the hyperscalers, the managed service providers and on-premises GPU racks instead.
This article tests that argument against the numbers. We go back to the Synergy Research and ABI Research figures the column quotes, check the arithmetic, read CoreWeave’s latest quarter for evidence of enterprise demand, and set out where these providers genuinely belong in an enterprise AI architecture. We finish with a risk register, a buyer’s checklist and the questions to ask before signing anything.
Table of contents
- What a Neocloud Is, and What It Is Not
- The Market Numbers, Checked
- Why Enterprises Leave the Neocloud Off the Shortlist
- CoreWeave’s Quarter Shows Both Sides of the Story
- Where a Neocloud Fits in the Enterprise Architecture
- The Enterprise Case: Cost, Capacity and Sovereignty
- The Risks Enterprises Take On With GPU Specialists
- How to Evaluate a Neocloud Provider: A Buyer’s Checklist
- What the Vendors Need to Fix
- What This Means for Your AI Infrastructure Plan
- Neocloud Questions Answered
- References
What a Neocloud Is, and What It Is Not
A neocloud is a cloud provider built almost entirely around one job: running AI training and inference on dense clusters of GPUs. Instead of the hundreds of managed services a hyperscaler sells, it offers GPU compute, high-bandwidth memory and fast interconnect, plus the cooling and power to keep them running.
GPU compute, memory and interconnect, and little else
Linthicum describes the model as “a specialized cloud that focuses entirely on serving up AI-based systems”. Synergy Research Group, whose data opens the column, frames it more technically. Its founder Jeremy Duke calls the category “an architectural response” to AI workloads that “impose far more rigid constraints – particularly around parallelism, locality and the concentration of compute” than generalised cloud elasticity was designed for. The column repeats that framing almost word for word.
The practical difference shows up in what you can buy. A hyperscaler account gives you databases, identity, queues, analytics, serverless functions and a GPU catalogue among thousands of other products. A typical GPU cloud account gives you clusters, schedulers, storage close to the GPUs and an API for spinning them up.
The providers in the category
Synergy names CoreWeave, Crusoe, Core Scientific, Lambda, Nebius and Nscale as the leaders, with CoreWeave “standing out as the most direct challenger to traditional hyperscale cloud providers”. We have covered several of them this year: Nscale’s $45 billion power and chips agreement with Anthropic, Crusoe’s $3.9 billion raise for data centres and modular AI factories, GMI Cloud’s $668 million Series B and Lambda’s plan to raise $4 billion ahead of an IPO.
A long tail sits behind those names. Synergy points to “newer entrants and transitioning crypto infrastructure providers”, and Data Center Dynamics notes that former crypto miners now see AI infrastructure as the more profitable use of their power contracts and sites.
Months, not years
Speed of construction is part of the pitch. Linthicum notes that these clusters “can be brought online in months rather than the three to five years a traditional hyperscale data center build requires”. For an enterprise that needs capacity this year, that gap matters more than any benchmark.
| Option | What you buy | Strength | Weakness |
|---|---|---|---|
| Hyperscaler (AWS, Azure, Google Cloud) | Thousands of managed services, GPUs among them | Integration, compliance tooling, existing contracts | GPU allocation, opaque pricing, egress fees |
| Neocloud | Dense GPU clusters, fast interconnect, AI-tuned storage | Capacity, price per GPU-hour, speed of build | Thin platform layer, younger balance sheets |
| Managed service provider | People and processes running your estate | Operational cover, one accountable supplier | GPU capacity normally comes from a cloud partner |
| On-premises GPUs | Your own servers in your own or colocated racks | Control, data locality, fixed cost once paid | Capital outlay, power, cooling, hiring |
Linthicum places the GPU specialists between the first and third rows: “a purpose-built layer for the specific AI workloads that neither hyperscalers nor MSPs serve efficiently”. That is a reasonable map. The difficulty, as the rest of the column argues, is that the vendors have not drawn it for their buyers.
The Market Numbers, Checked
The column opens with the figures that make the category impossible to ignore. We traced each one to its source and checked the sums.
Synergy Research: $25 billion in 2025, $400 billion by 2031
Synergy’s forecast release says neocloud revenue reached $9 billion in the fourth quarter of 2025, up 223% year on year, and exceeded $25 billion for the full year. It forecasts the market will approach $400 billion by 2031, “representing a sustained 58% compound annual growth rate”.
The arithmetic holds. Going from $25 billion to $400 billion is a 16-fold increase over six years, and the sixth root of 16 is about 1.587, which is a compound growth rate of roughly 58.7% a year. The 223% figure also implies that fourth-quarter revenue in 2024 was about $2.8 billion ($9 billion divided by 3.23).
ABI Research: $250 billion of GPU-as-a-service by 2030
ABI Research measures something narrower: revenue that neocloud providers earn from GPU-as-a-service (GPUaaS). Its November 2025 market data puts that at $42 billion in 2025, rising to “nearly US$250 billion” by 2030, a 43% compound annual growth rate. Again the sum works: $250 billion is 5.95 times $42 billion, and the fifth root of 5.95 is about 1.43.
ABI adds two details the column only touches. Inference revenue for these providers will grow 88% a year against 15% for training, and inference will account for 80% of the market by 2030. North America takes 88% of this GPUaaS revenue in 2026, falling to 72% by 2030 as sovereign cloud projects elsewhere come online.
The forecasts moved a long way in a short time
Here is something the column does not mention. An earlier ABI Research Highlight, built on its third-quarter 2025 report, forecast GPUaaS revenue from these providers rising “from US$24 billion in 2024 to more than US$65 billion by 2030”, a 17% annual growth rate. A few months later the 2030 figure was nearly $250 billion. That is an upward revision of about 3.8 times ($250 billion divided by $65 billion) to the same analyst firm’s view of the same year.
The two firms also disagree on the base year, because they count different things. ABI’s $42 billion of 2025 GPUaaS revenue is larger than Synergy’s whole-market figure of $25 billion for the same year. Neither number is wrong. They use different definitions, so they cannot be added, averaged or compared directly.
Bar widths are each figure divided by Synergy’s $400 billion forecast. The point of the chart is not that one forecast is right. It is that the published outlook for this market has swung by a factor of almost four inside a year, which should make any buyer cautious about vendor pitches built on “the market will be worth” slides.
Data centres: from 558 to more than 2,200
The column also quotes ABI’s facility count. ABI’s 2026 trends post forecasts more than 2,200 neocloud-operated data centres by 2035, “way up from just 558 facilities in 2025”. The firm’s April 2026 chart data puts the count at 696 in 2026 and 2,247 in 2035. That is roughly four times the 2025 base (2,247 divided by 558 is 4.03).
Bar widths are each count divided by 2,247. Growth is uneven: ABI expects North America to expand fastest, Europe to grow “steadily despite energy and permitting constraints”, and Asia-Pacific to go from 35 to 65 sites. One detail in ABI’s data stands out for enterprise buyers. By 2035 it expects ordinary, non-AI workloads to become the larger share of active capacity, as companies move traditional compute onto GPU clouds for latency, cost and supply reasons.
Why Enterprises Leave the Neocloud Off the Shortlist
If the numbers are that large, why are enterprises not asking for neoclouds by name? Linthicum’s answer is that the vendors are “confusing the very buyers they need most”.
A messaging problem, not a product problem
Enterprise and cloud architects are trying to work out where neoclouds sit relative to hyperscalers, managed service providers and on-premises kit, and “they’re not getting a coherent answer”. Faced with technology that does not look mature, enterprises “move on”, Linthicum writes: they default to the hyperscalers, hand the problem to an MSP or build their own GPU capacity.
He calls this “a missed opportunity of historic proportions for the neoclouds, and it’s one of their own making”. The phrase that matters for buyers is the one about architects: “Enterprise architects don’t buy petaflops; they buy a defined place in a reference architecture, a security and governance story, and a credible explanation of how this layer will interact with everything else they run.”
Circular deals flattered the growth figures
The second half of the argument is about where the revenue came from. “During the past two or three years, most of the neoclouds grew by selling to other technology companies in big side-to-side deals,” Linthicum writes, summarising the pattern as: “You buy $3 billion of my GPUs, and I’ll commit $3 billion to your neocloud service over the next five years.”
The concentration figures behind that are public. Microsoft represented 62% of CoreWeave’s total revenue in 2024, and Nvidia is a key customer for both Lambda and CoreWeave. ABI Research makes the same point and draws the same warning: without enterprise customers, neoclouds risk staying “GPU brokers for hyperscalers and chipmakers, trapped in a commodity position”. We looked at the financing side of this loop in our piece on Nscale’s IPO and Wall Street’s appetite for concentrated AI bets.
The 2009 parallel
Linthicum has seen this before. When public cloud started to take off in 2009 and 2010, the hyperscalers “were horrible at selling to enterprises”. They could not explain where their technology sat in the enterprise stack and did not have enough solutions architects. “It took years and billions in enterprise-facing investment to fix that.” His prescription for the newcomers is the same: more solutions architects, fewer tactical benchmarks.
CoreWeave's Quarter Shows Both Sides of the Story
CoreWeave is the largest listed pure play in the category, so its results are the best public evidence of whether enterprises are arriving. Its second-quarter 2026 release shows strong demand and an expensive way of meeting it.
| CoreWeave, quarter to 30 June | 2026 | 2025 |
|---|---|---|
| Revenue | $2,575m | $1,212m |
| Operating income (loss) | ($49m) | $19m |
| Interest expense, net | $640m | $267m |
| Net loss | ($626m) | ($290m) |
| Adjusted EBITDA (margin) | $1,510m (59%) | $753m (62%) |
| Revenue backlog | About $104bn | Not compared in release |
| Active power / contracted power | 1.5 GW / about 3.7 GW | Not compared in release |
Demand is not the problem
Revenue more than doubled, up 112% year on year. The backlog of about $104 billion excludes “more than $25 billion of net new customer commitments added in early Q3”. Divide the backlog by the quarter’s revenue and you get about 40 quarters, or roughly ten years of work at the current run rate, before counting the new commitments.
Financing is the pressure point
The cost of building that capacity is visible too. Net interest expense of $640 million equals about 25% of the quarter’s revenue (640 divided by 2,575), and the company swung from a small operating profit to an operating loss. A neocloud funded heavily by debt can be a perfectly good supplier, but its financial health becomes part of your risk assessment in a way that Microsoft’s or Google’s does not.
Bar widths are each figure divided by revenue of $2,575 million. Adjusted EBITDA leaves out depreciation and interest, the two biggest costs of owning debt-financed GPUs. That is how the same quarter can show $1,510 million of adjusted EBITDA and a $626 million net loss.
Enterprise names are starting to appear
The same release lists customer wins “across AI labs, hyperscalers, and enterprises”, naming Bentley Systems, Caterpillar, Grammarly, Isomorphic Labs and Sunday Robotics, and expanded work with Cognition, Databricks, Hudson River Trading, Periodic Labs, Rescale and Runway ML. Chief executive Michael Intrator said “customer demand is accelerating, as enterprise adoption broadens”.
The more telling detail for architects is the product list. CoreWeave launched CoreWeave Interconnect, private fibre linking directly to other hyperscale platforms “beginning with Google Cloud”, plus tools for running its SUNK scheduler and its data accelerator across clouds. That is a neocloud drawing its own place in the reference architecture: next to the hyperscaler, connected to it, not instead of it.
Where a Neocloud Fits in the Enterprise Architecture
Linthicum’s most useful line is “all architecture is personal”. Training a frontier model, serving real-time inference for a healthcare company bound by HIPAA and running regional inference with sovereign data controls for a European manufacturer are three different architectures. Each needs its own answer to whether a neocloud can serve it.
Workloads that suit GPU specialists
The strongest fits are the ones where GPU density, price per GPU-hour and time to capacity dominate. Large fine-tuning runs, batch inference over big document stores, model evaluation sweeps and high-volume inference behind a product all qualify. So does any project that is stuck because the hyperscaler account team cannot allocate the GPUs it needs this quarter.
Workloads that should stay where they are
Linthicum’s August column, “Neoclouds become AI’s new power brokers”, draws the line clearly: “General-purpose workloads, data platforms, application modernization, and enterprise integration may remain with the big cloud providers.” Moving your data warehouse to a neocloud so that it sits next to the GPUs usually trades one problem for several.
| Workload | Usual best home | Neocloud fit | Why |
|---|---|---|---|
| Fine-tuning an open-weight model | Neocloud | Strong | Short, GPU-bound jobs priced by the hour |
| High-volume inference behind a product | Neocloud or hyperscaler | Strong | Steady demand suits reserved capacity at a lower rate |
| Regional inference on regulated data | Sovereign neocloud or in-country region | Good, if residency is contractual | Local operators can guarantee where data sits |
| Retrieval-augmented chat over company documents | Hyperscaler | Partial | The data, identity and search live elsewhere |
| Data warehouse and analytics | Hyperscaler or existing platform | Weak | Few managed data services, egress both ways |
| Experimental AI features | Hyperscaler credits or APIs | Weak | Too small and uncertain to justify a commitment |
The interconnect is the architecture
The table hides the hardest design question: how the GPU cluster reaches your data. Every workload marked “strong” still needs training data, prompts and results to move between the GPU provider and wherever your systems of record live. Private interconnects such as CoreWeave’s are the right answer; pulling terabytes across the public internet is not. Settle that link before you compare GPU prices. If you are planning this as part of a wider cloud strategy, treat the GPU provider as one more region to connect, not a separate island.
The Enterprise Case: Cost, Capacity and Sovereignty
ABI Research’s “Three Gripes” research highlight lists the reasons enterprises look beyond the hyperscalers. Each one is worth testing against your own workloads.
Price transparency and egress
The first gripe is “opaque and unpredictable pricing models”. ABI says some organisations pay “US$3,500 to US$7,000 to transfer 50 Terabytes (TB) of data between cloud environments”. Taking 50 TB as 50,000 GB, that works out at $0.07 to $0.14 per gigabyte. Neoclouds such as Lambda and Civo publish simpler hourly and usage-based prices, which makes cost forecasting easier. Our guide to controlling AI token costs covers the API side of the same budgeting problem.
Capacity when the big clouds are short
Linthicum’s August column makes the point that many customers “are not buying compute because it’s cheaper than owning it. They buy compute because they cannot get the hardware any other way.” Shortages of advanced GPUs, DRAM and high-bandwidth memory are the binding constraint, and a provider that has secured supply is selling availability as much as price.
Sovereignty and regional control
The second gripe is compliance. ABI notes that the US CLOUD Act “introduces legal conflicts” with the EU’s General Data Protection Regulation, and that partnerships such as Vultr with Digital Realty and Lambda with Aligned Data Centers are “driving sovereign cloud deployments”. For a European enterprise, a regional operator that contractually guarantees in-country processing can be simpler to defend to a regulator than a US hyperscaler’s sovereign region. That still needs the same cybersecurity due diligence as any other supplier holding your data.
| Enterprise gripe (ABI) | What a neocloud offers | What to verify |
|---|---|---|
| Opaque pricing and lock-in | Flat-rate or usage-based GPU pricing | Egress, storage and support charges, not just GPU-hours |
| Compliance and sovereignty | In-country sites and local operating partners | Who can access the data, under which jurisdiction |
| One-size-fits-all platforms | GPU architectures tuned to specific workloads | Whether the tuning matches your workload, not a benchmark |
The Risks Enterprises Take On With GPU Specialists
None of this makes a neocloud a safe default. ABI reminds readers that hyperscalers “possess preferential treatment when it comes to GPU access”, which creates “pricing and procurement risks for neoclouds”, and that profitability is hard when providers compete on price.
Supplier concentration and financial health
A provider whose revenue depends on one or two giant customers, and whose growth is funded by debt, carries risks your procurement process should weigh. If its largest customer walks away or its lenders tighten terms, your capacity commitments are exposed. Ask for financial statements, check who owns the data centre buildings and the GPUs, and read the change-of-control clauses. The same vendor management checks apply to any young infrastructure supplier.
A thin platform layer
ABI’s buyer guide warns against providers where a “fragmented or DIY setup” is required beyond raw compute. Many of these providers hand you GPUs and leave the machine learning platform, observability, identity integration and security tooling to you. That work costs engineering time that rarely appears in the GPU-hour comparison.
Repeating the first cloud wave’s mistakes
Linthicum’s sharpest warning is in the August column. A decade ago enterprises “lifted and shifted applications as quickly as possible, signed commitments, and celebrated migration numbers”, then spent years on “cost overruns, poor workload placement, security gaps, operational confusion, and technical debt”. With AI infrastructure, he argues, mistakes “could cost 10 to 20 times more than a properly designed equivalent solution”.
| Risk | What it looks like | Mitigation |
|---|---|---|
| Provider failure or acquisition | Capacity withdrawn or terms changed mid-contract | Change-of-control and termination rights, data return clauses |
| Over-commitment | Paying for reserved GPUs you cannot use | Start on demand, commit only after measured use |
| Hidden data movement costs | Egress and storage bills larger than GPU spend | Private interconnect, keep data and compute close |
| Security and access gaps | Weaker identity, logging or key management than your main cloud | Security questionnaire, certifications, your own key control |
| Architectural sprawl | Another platform with its own tools and skills | Reference architecture first, provider second |
How to Evaluate a Neocloud Provider: A Buyer's Checklist
ABI’s 2026 guide for CIOs and CTOs sets out six criteria: a resilient hardware and supply chain strategy, large distributed GPU cluster capability, an advanced AI optimisation stack, integrated cloud, data and AI services, software co-development with partners, and high-performance interconnect. They are a good start. We would add five questions that come from the enterprise side of the table.
Five questions ABI’s list leaves out
- Exit: how do we get our data, models and logs out, at what cost and in what format?
- Residency: which named sites will process our data, and can that change without our consent?
- Service levels: what are the availability and response commitments, and what are the credits if they are missed?
- Financial resilience: who funds the hardware, and what happens to our capacity if the provider is sold or refinanced?
- Integration: how does the cluster connect to our existing cloud, identity provider and monitoring?
Run a paid pilot before any commitment
The simplest risk control is a short paid pilot on a real workload: one fine-tuning job or one inference service, measured on cost per useful result rather than raw throughput. It also tests the provider’s support team, which is where Linthicum’s missing solutions architects will either appear or not. A sound AI strategy starts from that question of business value, not from capacity.
Define the requirement before the provider
Linthicum’s August advice is the right order of work: “define requirements first, model the economics second, and select providers third.” He also asks the uncomfortable first question, “whether the workload needs AI at all”, noting that “traditional analytics, rules engines, search systems, automation platforms, and better application design may solve many of these problems without expensive AI infrastructure.”
What the Vendors Need to Fix
The column is addressed to vendors as much as buyers. Its advice is to “hire more solutions architects who know what they’re doing” and people who can “define technology configurations at a strategic level, not just throwing out grandiose tactical capabilities and tactical benchmarks.”
Reference architectures, not benchmarks
“Benchmarks don’t mean anything to anybody working on a real problem,” Linthicum writes. A vendor that publishes reference architectures for regulated inference, hybrid training with a hyperscaler data lake and sovereign deployments would answer the question architects are actually asking.
Platform depth through acquisitions
ABI notes that neoclouds have been buying software companies to build “full-stack capabilities”, citing CoreWeave’s purchase of Weights & Biases for model tracking and experiment management among several 2025 deals. Its analysts also urge providers to look at silicon beyond Nvidia, AMD and Intel, because relying on one chip supplier exposes them to “supply constraints, pricing volatility, and competitive risks”.
Vertical solutions
ABI’s strongest warning matches Linthicum’s: enterprise demand “may never materialize at scale unless neoclouds actively educate verticals and build tailored solutions”. Healthcare, financial services and manufacturing each need a packaged answer, including compliance evidence, not a GPU price list.
What This Means for Your AI Infrastructure Plan
For most organisations the right conclusion is neither “move to a neocloud” nor “ignore them”. It is to add the category to the options you assess, with a clear view of the workloads where it wins.
A practical sequence
- List your AI workloads and sort them by GPU intensity, data gravity and regulatory sensitivity.
- Keep data platforms and general applications where they are unless there is a specific reason to move.
- Price the GPU-heavy workloads on your current cloud, on two neoclouds and on owned hardware, including egress and staff time.
- Run a paid pilot on the strongest candidate, then commit in stages.
- Put the provider through the same IT governance and security review as any critical supplier.
Watch the energy question too
GPU density brings power and water costs with it, and the communities hosting these sites are starting to push back. Our analysis of AI data centre energy use and of Zankore’s AI factory and its water bill covers why sustainability reporting will increasingly reach your supplier questionnaires.
The irony Linthicum points to
The column ends on an irony worth remembering. If the vendors cannot explain themselves, the $400 billion opportunity “will be brought to neocloud competitors by the same category of customer the neoclouds never learned to sell to”. Enterprises do not need to wait for the vendors to fix that. They can draw the map themselves.
Neocloud Questions Answered
What is a neocloud?
A neocloud is a cloud provider specialising in GPU compute for AI training and inference, offering dense GPU clusters, high-bandwidth memory and fast networking instead of the broad service catalogue of AWS, Azure or Google Cloud. CoreWeave, Lambda, Nebius, Crusoe, Nscale and Core Scientific are the best-known examples.
Is a neocloud cheaper than a hyperscaler?
Often per GPU-hour, and pricing tends to be simpler. But the total cost depends on data movement, storage, support and the engineering work the provider leaves to you. Compare the full cost of a real workload, not a list price.
Should an enterprise use a neocloud for inference?
ABI expects inference to make up 80% of the neocloud market by 2030, so providers are building for it. It suits high-volume, steady inference where the data can sit close to the GPUs or travel over a private interconnect. Low-volume or experimental inference is usually cheaper through a model provider’s API.
Are neoclouds risky suppliers?
They can be. Many are young, heavily financed by debt and dependent on a few large customers. That is manageable with staged commitments, strong exit terms and a check on the provider’s finances, the same discipline you would apply to any critical supplier.
Why don’t enterprise architects shortlist neoclouds?
According to Linthicum, because the vendors have not explained where they fit relative to hyperscalers, MSPs and on-premises infrastructure, and because they sell on benchmarks rather than on a defined place in a reference architecture.
References
Neoclouds and the enterprises that need them (InfoWorld)
Neoclouds become AI’s new power brokers (InfoWorld)
Market forecast to approach $400B by 2031 (Synergy Research Group)
Revenue exceeds $25bn in 2025 (Data Center Dynamics)
GPU-as-a-service revenue forecast (ABI Research)
Data center count forecast, 2026 to 2035 (ABI Research)
Four trends for 2026 (ABI Research)
Three gripes driving enterprises from hyperscalers (ABI Research)
How to choose an AI-first provider: a six-point checklist (ABI Research)
CoreWeave reports strong second quarter 2026 results (CoreWeave)
SP 800-145, The NIST Definition of Cloud Computing (NIST)
Cloud services market investigation (Competition and Markets Authority)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
