Michael H

Posted on Aug 19 Originally published at agent5.news

What Is a Neocloud? Why AI Chipmakers Are Becoming Cloud Providers

#ai #technology #aiinfrastructure #cloudcomputing

Originally published on agent5.news.

Something unusual is happening in the cloud computing industry. The companies that make the chips powering the AI revolution are now investing heavily in the companies that rent those chips back to AI developers. NVIDIA, the dominant GPU maker, has poured billions of dollars into cloud startups. Those startups, in turn, use the money to buy more NVIDIA chips. The resulting web of relationships has produced an entirely new category of infrastructure business called the neocloud, and understanding it is one of the clearest windows into how AI development actually gets paid for.

What Is a Neocloud?

A neocloud is an AI-first cloud infrastructure provider that specializes in GPU compute for machine learning training and inference workloads. Unlike traditional hyperscalers such as Amazon Web Services, Microsoft Azure, and Google Cloud, which offer hundreds of managed services ranging from databases to serverless functions, neoclouds keep their catalog deliberately narrow. Their primary product, and often their only product, is raw GPU compute delivered over fast networking. Dense GPU clusters, high-bandwidth interconnects such as InfiniBand and NVLink, and NVMe storage tuned for AI workloads are the defining features of the category.

The term itself is relatively new. It gained traction in late 2024 and spread through 2025 primarily through analyst research, particularly from firms like SemiAnalysis and McKinsey, as a way to distinguish GPU-first cloud companies from hyperscalers that simply added GPU instances to an existing general-purpose platform. Before the label stuck, these companies were typically called "GPU cloud" or "GPU-as-a-Service" providers. The category itself, however, is older than the name. CoreWeave launched in 2017 as a cryptocurrency-mining operation called Atlantic Crypto, then pivoted its GPU fleet toward rendering and AI compute starting around 2019. Lambda Labs was building machine learning hardware as far back as 2012 before shifting to cloud rental as research demand grew.

Why Neoclouds Exist: The Compute Gap

The neocloud category emerged from a specific market failure. When generative AI exploded in 2023, demand for high-end GPU capacity surged faster than established cloud providers could respond. Hyperscalers had secured much of the available advanced GPU supply for their own priorities, leaving startups, research labs, and enterprises scrambling. Pricing reached levels that were, for many organizations, prohibitively expensive.

Neoclouds stepped into that gap with a different value proposition: faster provisioning, simpler pricing structures, and configurations optimized for high-performance AI workloads at significantly lower cost than what hyperscalers were charging for the same silicon. Companies like CoreWeave and Lambda Labs had cultivated supplier relationships with NVIDIA before the boom and could provision hardware that hyperscalers were quoting customers months out. The cost advantage was substantial enough to shift serious workloads.

Hyperscalers built their platforms around general enterprise computing, and that history creates real friction for AI teams. When a machine learning team needs only to run a large training job or a batch inference pipeline, the complexity and pricing overhead of a hyperscaler ecosystem does not serve them. Neoclouds strip that overhead away. The tradeoff is that when a team needs a managed database, a message queue, a load balancer, and GPU access under a single identity and billing system, a hyperscaler remains the more practical choice.

Who the Major Players Are

The neocloud landscape has a handful of prominent names and a much longer tail of smaller operators.

CoreWeave has emerged as the largest and most prominent company in the category. It operates as a specialized NVIDIA partner and went public on Nasdaq in March 2025 under the ticker CRWV. Since its IPO, the stock has dramatically outperformed initial expectations. The company reported Q1 2026 revenue of approximately $2.08 billion, up 112 percent year over year, and holds a contracted revenue backlog of $99.4 billion as of March 31, 2026. Its customers include OpenAI, Microsoft, Meta, Google, and Anthropic. Microsoft alone accounted for a large share of CoreWeave's early revenue, creating a notable customer concentration risk that management has been working to diversify.

Lambda Labs has built a developer-first identity, maintaining a broad GPU fleet and emphasizing transparent pricing and strong Kubernetes-based orchestration tools. NVIDIA is an investor. The company also offers on-premises, private cloud GPU clusters with InfiniBand networking and storage, and has pursued leaseback arrangements with NVIDIA to maintain access to cutting-edge hardware.

Nebius, headquartered in Amsterdam, secured $700 million in funding in December 2024 followed by further financing in 2025, and has been building a global presence across North America, Europe, and the Middle East. It targets developers and enterprises seeking alternatives to both hyperscalers and CoreWeave's dominant position.

Other notable players include Crusoe, which focuses on sustainable GPU computing powered by stranded and renewable energy sources, RunPod, Together AI, and Vultr. The list is growing. ABI Research projected in 2025 that more than 2,200 neocloud-operated data centers could be in operation globally by 2035, up from roughly 558 facilities in 2025.

Why NVIDIA Is Becoming a Cloud Kingmaker

The most structurally interesting dynamic in the neocloud world is NVIDIA's role. The chipmaker has not just supplied hardware to neoclouds. It has actively invested in them, creating a web of financial relationships that analysts have described as circular financing.

NVIDIA first invested $100 million in CoreWeave back in 2023. It added to that position around the time of CoreWeave's IPO, then invested roughly $2 billion more for additional shares in a purchase disclosed in early 2026, bringing its disclosed stake to approximately 47.2 million shares. NVIDIA also invested $2 billion in Nebius in March 2026. It has separately backed Lambda Labs, Nscale, and Crusoe. NVIDIA's own regulatory filings acknowledge neocloud builders as a recognized customer category within its business.

The logic from NVIDIA's perspective is straightforward: investing in the companies most likely to buy enormous quantities of NVIDIA hardware for years to come is a way to lock in future chip demand. The neoclouds, for their part, use the investment capital to purchase more NVIDIA GPUs. Some portion of the money NVIDIA invests flows directly back to NVIDIA in the form of chip purchases. Critics have noted this loop explicitly. Supporters argue the arrangement reflects genuine mutual interest in expanding the AI infrastructure ecosystem.

NVIDIA has also provided more direct backstops. A signed take-or-pay capacity arrangement with CoreWeave, for instance, obligates NVIDIA to purchase GPU cloud capacity from CoreWeave if that capacity goes unsold, providing a financial floor under CoreWeave's utilization risk.

How Hyperscalers Became Neocloud Customers

One of the more counterintuitive facts about the neocloud market is that the traditional cloud giants have become among the biggest buyers of neocloud capacity. Microsoft has struck commitments worth tens of billions of dollars with CoreWeave, Nebius, and Nscale. Meta has signed multi-billion-dollar deals with CoreWeave and Nebius. OpenAI committed to over $22 billion in total capacity from CoreWeave through a series of expanding agreements.

The reason relates to balance sheet mechanics and speed. Building and operating data centers requires enormous capital expenditures. Purchasing capacity from a neocloud on a multi-year operating contract allows hyperscalers to secure AI compute without immediately loading their own balance sheets with the same level of hard assets. Neoclouds, meanwhile, carry the infrastructure risk and the debt load required to build it.

This arrangement also functions as an overflow valve. Hyperscalers face their own GPU allocation constraints and construction timelines. Neoclouds that already have clusters online and contracted relationships with NVIDIA can deliver capacity faster. As long as the economics of outsourcing compute are more attractive than building it in-house, the relationship persists.

The Real Economics: Capital Intensity and Concentration Risk

The business model of a neocloud looks straightforward on the surface: buy GPUs, rack them in dense clusters, rent access at a markup, and expand capacity as contracts grow. The underlying economics are considerably more demanding.

Neoclouds operate in a capital-intensive environment where hardware must be purchased, data centers powered, and networking installed long before revenue arrives. CoreWeave guided for capital expenditures of $31 billion to $35 billion in 2026, roughly double its 2025 spending. The company has raised approximately $28 billion in combined equity and debt over the 12 months through March 2026 to finance that expansion. The debt load is substantial, and GAAP profitability remains a future goal rather than a present reality despite strong revenue growth.

Customer concentration is an equally significant risk. Microsoft accounted for approximately 67 percent of CoreWeave's full-year 2025 revenue. While the backlog is diversifying toward OpenAI, Meta, Anthropic, Jane Street, and others, the dependence on a small number of very large customers means that a renegotiation, cancellation, or decision by one of those customers to expand in-house compute capacity could materially reshape CoreWeave's revenue picture.

Hardware obsolescence adds another layer of complexity. AI chip generations are shortening. The transition from NVIDIA's Hopper architecture to Blackwell accelerators illustrates the speed of these cycles. A neocloud that took on debt to deploy a large H100 cluster may find the residual value of that hardware under pressure sooner than originally underwritten.

The Pricing Advantage (And Its Limits)

Neoclouds have competed primarily on price and speed of access. Without the overhead of hundreds of enterprise services and the organizational complexity of hyperscale cloud providers, specialized operators can undercut the majors on raw GPU hourly rates. Companies like CoreWeave and Lambda leverage volume commitments and direct supplier relationships to negotiate discounts that they can pass through to customers.

That pricing advantage, however, is not guaranteed to persist. Hyperscalers are not standing still. AWS cut prices on certain H100 instance types in mid-2025, applying direct margin pressure to neocloud offerings. Google, Microsoft, and Amazon have all been investing in custom AI accelerators, including Google's TPUs, AWS's Trainium chips, and Microsoft's Maia architecture, that could reduce their dependence on NVIDIA silicon and give them more pricing flexibility over time. If hyperscalers bundle AI compute cheaply or migrate their best customers to proprietary hardware, the neocloud pricing advantage narrows.

The shift from AI model training toward inference workloads may also reshape the competitive dynamics. Training runs require enormous, sustained GPU clusters and favor the kind of dense, bare-metal infrastructure that neoclouds excel at delivering. Inference workloads are more distributed, more latency-sensitive, and more amenable to the managed-service layers that hyperscalers already operate. ABI Research analysts have forecast that inference workloads will account for roughly 80 percent of the neocloud market by 2030, which means neoclouds will need to evolve their platforms to serve a substantially different kind of customer demand.

What Happens Next: The Agent5 View

Reasonable people following the neocloud space make widely varying predictions about where this ends up, and that uncertainty is itself informative. The range of plausible futures is wide.

In one scenario, AI compute demand remains structurally supply-constrained for the rest of the decade. Hyperscalers cannot build fast enough on their own, neoclouds lock in long-term contracts and use them to service their debt, and the GPU-first cloud model becomes a durable fourth pillar of global cloud infrastructure alongside AWS, Azure, and GCP. CoreWeave's contracted backlog of $99.4 billion as of early 2026, backed by some of the largest AI spenders in the world, is the primary evidence for this view.

In a second scenario, the AI infrastructure build-out is running ahead of actual end-user demand. Hyperscalers eventually bring enough capacity online, custom silicon reduces NVIDIA's dominance, and the neocloud segment undergoes significant consolidation as smaller operators cannot service their debt at lower utilization rates. Providers competing on generic GPU rental rates without meaningful software or vertical differentiation face existential margin pressure. The risk profile of these businesses, as some analysts have noted, resembles infrastructure rather than software.

A third and perhaps most likely scenario is something in between: the market bifurcates. A small number of well-capitalized neoclouds with strong supply relationships, deep customer diversification, and genuine software capabilities survive and scale. The longer tail of smaller GPU rental businesses consolidates or exits. The winners become embedded partners in the global AI supply chain, functioning as a kind of permanent overflow and specialty layer beneath the hyperscalers.

Getting smart about AI means holding all three scenarios simultaneously and updating your probability weights as new signals arrive. When CoreWeave announces a new multi-billion-dollar contract, that is a signal toward the durable model. When a major customer signals it may build in-house capacity, that shifts weight toward consolidation. When NVIDIA reports that neocloud builders represent a growing slice of its indirect customer base, that tells you the chipmaker is structurally committed to this channel.

The neocloud story is, at its core, a story about where the AI economy is placing its highest-confidence bets on infrastructure. The companies and capital flows building this layer are making predictions, in the tens of billions of dollars, that AI workloads will remain compute-hungry and that purpose-built GPU clouds will remain the most efficient way to serve that hunger. Watching which of those predictions are vindicated over the next two to three years is one of the clearest indicators of how the broader AI industry is actually maturing.

DEV Community