VAST Data and the evolving neoclouds

VAST Data has emerged as the dominant storage supplier to neoclouds, with a 600 PB capacity win at Australia's Sharon AI. Neoclouds, which specialize in GPU-compute-heavy AI services, saw collective revenues of $23-25 billion in 2025, more than triple 2024 levels, as they compete with hyperscalers in the AI GPU-as-a-service market.

VAST Data and the evolving neoclouds VAST Data https://www.blocksandfiles.com/ai-ml/2026/05/29/mistral-ai-chooses-vast-data-for-its-ai-data-foundation/5248594 is the dominant supplier of storage to neoclouds with its latest 600 PB capacity win at Australia’s Sharon AI cementing its position. The neoclouds are evolving, establishing sovereign datacenters and expanding into the managed service provider MSP market as they compete with hyperscalers. Six or seven years ago, there were the hyperscalers providing their processing and storage services services based on x86 or Arm compute hardware, small scale MSPs providing lower-cost niche equivalents, and crypto-currency mining bit barns, using GPUs. Then there was a crypto market crash in 2018 ,and the crypto-miners switched to using their GPUs to offer 3D rendering, scientific simulations and machine learning cloud compute services. These evolved into large language models LLMs and the neoclouds “new”clouds competed with the old clouds, the hyperscalers, who did not treat GPU-based cloud compute services as a priority. Then a gigantic and fortuitous event took place in late 2022; ChatGPT was launched. It generated a sustained frenzy of AI foundation model training by AI developers - and the neoclouds had the GPUs and associated infrastructure, not the hyperscalers. The neocloud suppliers specialized in GPU-compute-heavy AI services and their revenues rocketed. Collectively, they saw revenues of $23 to $25 billion in 2025, more than triple their 2024 revenues. Market leader CoreWeave https://www.blocksandfiles.com/ai-ml/2025/11/06/vast-data-lands-117-billion-coreweave-deal/1608868 reported $5.13 billion revenues in 2025. The overall neocloud revenue number was significantly more than AWS, which had a $15 billion-plus AI revenue run rate earlier this year. Of course, AWS overall revenues were far greater, at $128.7 billion on 2025, but the public cloud hyperscalers were wrongly-positioned for the AI GPU-as-a-service boom and are now in a catch up position, using their own silicon as well as Nvidia’s. The neoclouds are seeing their market evolving, away from domination by a few large AI model training customers towards a broader one as AI inferencing becomes enterprise-wide. Against this background all-flash storage company VAST Data has emerged as the dominant neocloud storage supplier. Co-founder and CTO Alon Horev told us: “We have very good efficiency of how we use Flash across our neoclouds. We're getting two to one data reduction. And SSD-HDD tiering and some of the methodologies that were implemented for many, many years; they risk GPU sitting idle. So if you're going to store a dataset or a checkpoint on hard drives, you end up easily leaving the GPU idle reading that data. … we see customers going all Flash, not thinking twice, and it's definitely a challenge and an opportunity for us to also utilise Flash in the best way possible.” We asked if now, with neocloud customers storing many petabytes of data on VAST flash, there was a cost pressure to move old data into archives? Not at all, Horev said. The neoclouds; ”at the end of the day; they're looking at that target of keeping it extremely simple and giving the customers the confidence that their GPUs are not going to be idle because that's how they get compared often to other providers.” The cost of the GPUs and their infrastructure, and the need for them to be constantly bus, helped by VAST’s software, is far more important than the cost of VAST Data’s storage. The neoclouds need their underlying storage infrastructure to just work, as “the cost of making a mistake is just extremely big.” If the infrastructure doesn’t work, the the entire GPU compute farm doesn’t work. Horev said: ”When customers train video models, they can bring in half an exabyte in a single dataset. In that scenario ,where you're taking a bunch of thousands of GPUs and training them against 500 petabytes of VAST, VAST is becoming a bigger part of the cake.” “I think the challenge they ask themselves is how much would it cost me to do it with alternative technologies? How much would I pay for hardware, for software? How would my uptime look like? Would the customer see something as simple as this, a single file system or a single bucket that gives me all this capacity?” He said that, originally, neoclouds built some dedicated customer environments, but this is changing as the neoclouds want to become more like hyperscalers: “There's a very big difference in building infrastructure for a specific customer doing one-offs or dedicated environments. And these have been very profitable for some of the infrastructure and neoCloud builders because some of the end customers are big, and the end customer may say, "Hey, I'm going to give you a big contract. I want my own dedicated environment." But the biggest opportunity for the neoClouds to graduate into hyperscaler mode is that they can decouple the infrastructure from the end user. So it becomes fungible. The customers can grow, they can pay as they go, creating an alignment in the consumption model.” He explained it this way: “The consumption model that the end customer wants is pay per gig. They don't want to prepay for a massive cluster in advance. This is where we actually create the highest margin services, that's true for both GPU and CPU and storage and databases and everything. “If you look at what the hyperscalers built, they build infrastructure where they can partition and segment it in logical terms and not send someone to the data centre to recable something, not force the sellers to close three-year contracts, otherwise we're not going to open the door and talk to you. So I think this is where the margin grows, and this is where neoclouds have some insight into how the future looks like. So I could actually onboard a customer that wants to do only three months. I need to have full automation. I need to reduce friction. It means everything needs to be multi-tenant.” This is closer to the MSP business model. “They're going to build one data centre that has 100,000 GPUs and they're going to partition them to different customers, which is not easy, by the way, because of security and isolation. It's not an easy feat, but doing it for the data as well. So you're going to have a VAST cluster with lots of flash capacity, lots of compute power, enough to satisfy all the requirements, but then you need to be able to partition that. … the business outcome for them is that they can now charge very differently versus showing the customer an ROI that, if they sign up for a three-year or five-year, it's going to be worth it.” Are you seeing signs that traditional MSPs are looking to become neoclouds? Horev said yes: “You see telecom providers upgrading into neoclouds. You see other managed service providers upgrading into neocloud. So the opportunity is there. Each and every one of those organisations today is doing the math to see if they can build a successful model based on that.” “We see customers testing the waters, talking to us, trying to understand what it takes, they have a challenge and an opportunity because again, if they just look at the margins of renting out GPUs or even buying VAST and reselling it, it's one certain cost profile or margin profile that they can have and they're trying to understand where it can go. What can they do more in terms of providing more services, providing some differentiation in the market? But it does feel right now, because of the demand for GPUs, that anyone that puts GPUs on the floor can rent them out.“ But it is not easy for MSPS to become neoclouds and offer, for example, AI inferencing services. “ I think the world of inference requires expertise. It's extremely dynamic. If you just take the best in class open-source today, it's not going to be the best in class open-source in a month from now. … If you want to enable your customers to bring their own models, that's a type of service. How do you enable them to upload the models? How do you scale them up and down dynamically?” So this is nowhere near as simple as offering virtual machines as a service? “Exactly. I think there is a recipe for how to do infrastructure. There's IP experts, networking experts, admins that are upgrading their skills from doing classic computing into AI computing where they need to learn InfiniBand and Spectrum X https://www.blocksandfiles.com/ai-ml/2025/02/04/nvidia-says-spectrum-x-adaptive-routing-can-boost-storage-fabric-bandwidth/1606433 and liquid cooling. They need to learn how to do those things. They often also need to learn how to drop what they know on virtualization and adopt environmental hosting and Kubernetes, but it's within their expertise.” It can be done. VAST can help the neoclouds offerAI services to enterprises because VAST already has enterprise customers: “I feel that we're helping the NeoClouds understand what enterprises need because when NeoClouds had to support an AI lab during training, they needed a fast file system, they needed a fast object store. When they need to support an enterprise that now wants to do inference on the NeoCloud, they need to have FIPS compliance, they need to have auditing, they need to have security, they need to have support from multiple different protocols. They're storing user information.” “You need platforms that have been stress tested and been proven to be able to actually support multi-tenancy ,where you have multiple different companies, customers sitting on the same physical infrastructure without seeing their data, without having noise neighbours, and having all of those security features alongside it; encryption, various types of encryption, and external key managers. This came from the enterprise. So today we're actually helping the neoclouds going through security checklists and onboarding their customers.” We think it is unrealistic to expect the neoclouds to become hyperscalers, but they can expand into offering MSP-type services, while MSPs grow upwards, as it were, and start offering AI services. VAST will see its neocloud market expanding in both cases.