For VAST Data, the promise to build the AI Operating System is more than just clever marketing. It is steadily expanding its own offering to deliver on that promise. During the very first edition of VAST Forward, it is emphasizing this ambition even more.
VAST Data occupies a somewhat special position within the IT stack. It is difficult to pigeonhole the company. Of course, it is a company that deals with storage, but it no longer sells hardware (it stopped doing so a long time ago). Nevertheless, it competes to a certain extent with traditional storage suppliers. Furthermore, it has no particular preference for the infrastructure on which the software runs. That software is in fact a continuous development on top of the DataStore with which it all started.
In previous articles, we have explained in detail which components VAST wants to use to build the operating system for the AI era. Feel free to reread those articles if you are looking for more background information. We will not go through this whole story again in this article. We will focus mainly on what VAST Data’s announcements say about where the company is headed.
VAST wants to take on the VMware role
According to Renen Hallak, founder and CEO of VAST Data, the company wants to be the Switzerland of the world of (AI) data infrastructure. In principle, it wants to work with everyone and on any infrastructure. That is not usually the easiest proposition, because it offers so many possibilities that it can also scare off potential customers. On the other hand, it also means that other choices customers make in terms of IT infrastructure don’t really matter. Whether they choose servers from HPE, Cisco, Lenovo, or Supermicro, or are deeply invested in Azure, AWS, GCP, or Oracle, it doesn’t matter.
If we think carefully about how we would characterize VAST Data today, the analogy with VMware comes to mind. The role that company took on more than twenty years ago in the field of virtualization on top of whatever companies already had in place can be seen as the role that VAST Data also wants to take on. Again, that is both good news and bad news. The good news is that this approach has the potential to lead to enormous growth. The bad news is that it could also lead to enormous uncontrolled growth. VMware suffered greatly from the latter, particularly in the last five to ten years before it was acquired by Broadcom. VAST is still a long way from that level. In our opinion, it should see this as a warning.
New engines for more control in runtime
Looking at the announcements VAST is making this week at VAST Forward, we start with a couple of new engines: VAST PolicyEngine and VAST TuningEngine. These are not the first engines VAST Data has. There is already the DataEngine (the runtime of VAST Data), but also InsightEngine and AgentEngine. In short, InsightEngine provides context when retrieving and ingesting data from a variety of sources. AgentEngine ensures that multi-agent workflows run smoothly over a longer period of time.
With the PolicyEngine and TuningEngine, VAST adds two more important elements to what is needed to keep AI workloads running for longer periods of time. The TuningEngine can be seen as an extension of the InsightEngine and AgentEngine. In other words, this engine works with the results of the other two. The TuningEngine receives the telemetry and uses it to continuously improve the performance of the AI models and workloads.
In addition to good and reliable performance of AI models and workloads that organizations run, it is also important that they behave properly. In other words, there must be governance. That is where the VAST PolicyEngine comes in. This can be seen as a control plane through which all interactions between AI agents and between AI agents and data must pass. The PolicyEngine evaluates things such as identity, context, and specific policies of an organization before the agents can take action.
If you are now enthusiastic about the PolicyEngine and the TuningEngine, you will unfortunately have to be patient. They will not be available until the fourth quarter of this year. VAST is therefore very early in announcing them. Apparently, it was keen to announce something in this area during the conference.
Polaris as an overarching control plane
The VAST PolicyEngine we discussed above functions as a control plane at a relatively low, fine-grained level in the data pipeline. However, VAST also has ambitions on a more overarching level. It also wants to offer customers a control plane for the entire AI infrastructure. According to the company, this is necessary because AI infrastructure is no longer limited to a single rollout in a single location. As soon as you start working with AI in multiple locations (on-premises, in the public cloud, at a neocloud), it becomes important from an operational perspective to have a good overview of this. That is why VAST is announcing Polaris today. This should be able to offer this to customers.
Polaris is a Kubernetes-based control plane that is suitable for multi-tenant use (and therefore also interesting for service providers and the like). It is an agent-based environment. This means that an agent is installed on all VAST nodes that organizations have spread across multiple locations. Polaris then automates tasks such as provisioning and integrates with marketplaces in the cloud to keep track of how much space is left to scale up, for example. It also ensures centralised rollout of upgrades, expansion and replacement of nodes.
Polaris does not simply start working on its own in customer environments. It is intent-based. This means that customers themselves indicate how they want their infrastructure to look. Based on this, Polaris sets to work to achieve the desired state and maintain it. To a certain extent, Polaris is an extension of DataSpace, VAST’s global namespace. That provides a data fabric across everything. Polaris delivers the other side of that coin. While DataSpace focuses on the data, Polaris deals with the infrastructure.
Polaris is now available, although it is not yet fully developed. It is currently only available for cloud environments. The extensions to neocloud and on-prem will come later.
CNode-X brings VAST to the GPU (or vice versa)
In its quest to further optimize AI infrastructure, VAST is continuing to expand the capabilities of its self-proclaimed OS for AI. Until now, however, there was a hard limit at the GPU. That did not fall within the VAST stack. VAST focused on making advanced storage services available to GPUs, but could not do much else with those GPUs itself. With today’s announcement of CNode-X by VAST Data, that is set to change.
VAST developed CNode-X in collaboration with Nvidia. You could see this as a further development of the certifications that Nvidia has long awarded to products from storage suppliers. As far as we know, that program does not exist anymore. VAST CNode-X, however, is a fully Nvidia-certified system. Vendors such as HPE, Cisco, Lenovo, and Supermicro will sell it once it is generally available. Nvidia and VAST have, as it were, reversed the process.
At first glance, VAST CNode-X is a smart way for Nvidia and VAST to provide more guidance on what a good AI stack should look like. And, of course, to give the market an extra boost in terms of sales. Under the hood, the new systems also mean that GPUs are now part of the VAST ecosystem. The GPU-systems can now be managed, read, and deployed at the same level as the rest of what VAST has to offer.
Specifically, this means that VAST can now integrate all kinds of Nvidia libraries and APIs directly into its services. Think of cuVS for vector search, Nemotron models, and NIM, but also Nvidia’s Context Memory Storage (CMS) platform. This should increase performance and reduce latency. In other words, the whole thing should become a lot more efficient.
Tighter stacks, but more possibilities
We see today’s announcements as the next step in the maturation of VAST Data. Especially now that it is increasingly focusing on rolling out AI in enterprise environments, the issues described in this article are important developments.
In collaboration with major AI players such as public clouds, Nvidia, or neoclouds such as CoreWeave, the main focus is on being able to deliver the right components. However, to be successful in the enterprise world and with system integrators and other partners, there needs to be more consistency in the offering.
That is why VAST Data is increasingly incorporating more elements into its own platform, focusing heavily on the development of multi-level control planes and developing fully integrated stacks together with Nvidia and hardware suppliers. Developing an OS for AI (whatever that may mean exactly) is certainly a worthy goal, but customers generally get the best experience when the OS and hardware complement each other perfectly. That is what VAST is working on.
VAST is by no means finished. There is still plenty of room for further development. Take security, for example. It has now announced a partnership with Crowdstrike for this, but it is undoubtedly something it wants to take further itself. But that’s something for next time.
Read also: Who is developing the OS for AI? VAST Data is going all in