The reported $10 billion deal between Anthropic and AI cloud startup Volta is more than just another big-number headline in the AI infrastructure race. It’s a sign of where cloud computing is headed. When a frontier AI company needs massive amounts of specialized compute, and a relatively young cloud provider claims it can deliver that capacity, billions of dollars shift hands before most enterprises even figure out their actual AI infrastructure requirements.
The Anthropic-Volta deal is notable for its size, duration, and specialized AI hardware involved. But it’s also a signal, and that’s the real story. We are watching the rise of neoclouds, a new class of cloud providers built less for generalized enterprise hosting and more for focused, high-performance infrastructure for AI training, inference, model serving, GPU clusters, and related workloads. The hyperscalers are still growing, and recent revenue announcements from Amazon, Microsoft, and Google show that the Big Three remain very much in control of the broader cloud market. However, AI has created enough demand, specialization, and scarcity to open a lane for a new group of infrastructure providers.
Large commitments have long shaped the cloud market. Enterprises signed multiyear spending agreements with hyperscalers. SaaS companies prepaid for capacity. Startups took cloud credits and grew into large consumption accounts. What’s different now is the scale and urgency of AI infrastructure deals. These are not ordinary enterprise hosting contracts. They are multibillion-dollar bets between companies that both need the deal to validate their futures.
For the AI model company, the contract is about access to compute. Without processors, power, networking, cooling, and memory, there is no competitive frontier model strategy. For the neocloud provider, the contract is about credibility. A $10 billion commitment from a major AI player signals to investors, chip suppliers, data center partners, and future customers that the provider is not just another GPU reseller with a website. It becomes part of the AI supply chain. This is one reason neoclouds are emerging so quickly. They are being financed, validated, and scaled by demand that did not exist in this form even a few years ago. Many of these providers will serve small and midsize businesses that cannot get meaningful GPU access from the hyperscalers, or they do not want to navigate the complexity of massive cloud platforms. At the same time, they will also serve the largest AI companies in the world. That range is unusual. The same class of provider may offer on-demand inference capacity to a startup on Monday and negotiate a multibillion-dollar training infrastructure agreement with a model company on Tuesday.
The rise of neoclouds is not only about clever positioning. It is also driven by scarcity. Enterprises and AI companies cannot always buy the processors they need, even when they’re within budget. Advanced GPUs and AI accelerators are constrained by supply chains, manufacturing capacity, allocation decisions, and the simple fact that everyone wants the same hardware at the same time.
Memory is another bottleneck. Shortages of DRAM and high-bandwidth memory are central constraints. A model training cluster is not just a pile of GPUs. It requires memory, networking, storage, power infrastructure, cooling systems, and operational expertise. If any of those components become constrained, the entire infrastructure plan slows down.
Neoclouds can step into this gap by aggregating access to scarce resources and turning them into consumable infrastructure. In many cases, customers are not buying compute because it’s cheaper than owning it. They buy compute because they cannot get the hardware any other way or because they cannot build the operating model fast enough. This is on-demand infrastructure, but with a narrower, more urgent purpose than traditional public cloud consumption.
That demand creates a specific niche. Neoclouds will focus on serving the training and inference needs of companies that require specialized AI infrastructure but do not want (or cannot manage) the capital expenditure and operational complexity of owning it. Some will specialize in training clusters. Others will focus on low-latency inference. Some will build around specific chip architectures or data center geographies. The winners will not just have GPUs. They will have reliable capacity, predictable economics, strong networking, useful software layers, and a clear understanding of AI workload operations.
It would be a mistake to think neoclouds will replace Amazon Web Services, Microsoft Azure, or Google Cloud. That’s not happening. The hyperscalers also benefit from the AI boom, and they have enormous advantages in enterprise relationships, global infrastructure, platform services, security, compliance, and ecosystems. Their revenue growth shows that enterprises are still buying cloud services from them at massive scale.
However, demand for AI infrastructure is expanding faster than any single class of provider can absorb. Hyperscalers will capture a large share of this market, but not all of it. Neoclouds will continue emerging because the market needs more capacity, specialization, and options. Over the next two to three years, I expect many of these neocloud providers to inflect sharply. Revenue will rise quickly for those that secure hardware, power, and customers. Some will become acquisition targets. Some will fail because they overcommit, underdeliver, or discover that running AI infrastructure at scale is harder than raising the money to build it.
The most interesting outcome is not a world where neoclouds beat hyperscalers. It is a world where enterprises use both. General-purpose workloads, data platforms, application modernization, and enterprise integration may remain with the big cloud providers. Specialized training and inference workloads may move to neoclouds when economics, availability, or performance make sense.
This is where the gold rush becomes dangerous. Many enterprises are making strategic decisions about where AI training and inference will run before they understand their own requirements. This is a recipe for disaster.
Ten to 15 years ago, enterprises rushed into public cloud because they believed speed mattered more than architecture. They lifted and shifted applications as quickly as possible, signed commitments, and celebrated migration numbers. Then they spent the next decade dealing with cost overruns, poor workload placement, security gaps, operational confusion, and technical debt. In 2026, many organizations are still repairing the damage from cloud decisions made in a hurry.
The AI infrastructure push is even more extreme. Mistakes will not merely cost two or three times more than expected. In some cases, they could cost 10 to 20 times more than a properly designed equivalent solution. For a Global 2000 company, it can become a board-level issue. In the worst cases, it can become a bankruptcy-level decision, especially when companies lock themselves into capacity levels, architecture, or operating models that do not align with real business demand.
Enterprises need to slow down. Don’t press the “Buy” button until you understand what the infrastructure is for. Are you training foundation models, fine-tuning existing models, running retrieval-augmented generation, hosting inference for internal applications, or experimenting with AI features that may never reach production? These are different problems. They require different infrastructure, economics, and risk models.
The most important question is whether the workload needs AI at all. Many applications are being pushed into AI because executives want AI attached to everything. That is not strategy. That’s branding. Traditional analytics, rules engines, search systems, automation platforms, and better application design may solve many of these problems without expensive AI infrastructure.
Neoclouds will be an important part of the next cloud era. They will provide the capacity the market desperately needs, and some will become durable infrastructure companies. But enterprises should not confuse availability with suitability. Just because someone can sell you GPUs does not mean you should buy them. The smart move is to define requirements first, model the economics second, and select providers third. Anything else is just another land grab, and we already know how that ends.