When people picture a data center, they tend to picture compute servers, GPUs, watts and raw horsepower. But there is another side of the equation that decides whether that expensive compute power actually does anything useful: Storage. In this episode, we'll tackle what storage is, what decisions make the biggest impact, key challenges and how to handle the current storage shortage.
Welcome to The Five Nine miniseries Data Center 2.0, were we dig into what's changing in the infrastructure behind modern businesses.
Catch the video at top, listen to the audio edition and read our transcript below, or watch this and future episodes on YouTube. Missed the first episode in this series? Watch it here.
To learn more about the topics in this episode, check out:
- VAST Data raised another $1B. Here’s where its CEO says the money’s going
- Fragmented data drains the competitive 'moat,' says IBM
- Rakuten Mobile reveals its data alchemy secrets
- Nokia taps Pure Storage to build data pipeline for telco AI, CNFs
- Is object storage the secret to telco AI at scale?
- Here’s why Nvidia-backed startup VAST Data is all the rage
- The Five Nine: Telcos are drowning in data debt
This podcast is written and hosted by Diana Goovaerts. It is edited by Diana Goovaerts and Matt Rickman. Liz Coyne is our executive producer. Special thanks to guests Matt Kimble, Jay Subramanian, Rob Lee and Andy Persteiner.
Diana Goovaerts: When people picture a data center, they tend to picture compute servers, GPUs, watts and raw horsepower. But there is another side of the equation that decides whether that expensive compute power actually does anything useful: Storage.
Andy Persteiner, Field CTO, VAST Data: Processing power means nothing if there's no data to feed it. GPUs are now the predominant force in the data center. They need to be fed quickly and storage needs to be fast enough to keep up and make sure that they're always active and running.
Jay Subramanian, SVP and GM for Core Storage Platforms, Hitachi Vantara: If the storage underneath is not providing that level of performance and becomes a bottleneck, the entire chain above kind of gets distorted.
Matt Kimball, Principal Analyst, Moor Insights and Strategy: Storage has gone from being a repository for your data to an active participant in the AI equation.
Diana Goovaerts: Welcome to The Five Nine miniseries Data Center 2.0, the show where we break down what's changing in the infrastructure that powers modern business. I'm your host, Diana Goovaerts, and here's today's roadmap: First, we'll cover what storage means now, then the storage decisions that shape outcomes. After that, we'll tackle the biggest storage challenges – scale, complexity, distributed data – and finally, what's behind the storage shortage and how teams can adapt.
So, let's start with a basic question that turns out to be surprisingly hard. When we say storage, what exactly are we talking about?
Rob Lee, Chief Technology and Growth Officer, EverPure: So, the term storage, when we think about it from a data center context, encompasses historically many different solutions that are being brought to bear to solve one problem. And what I mean by that is if you think about the other legs of the data center, the compute, the networking, generally speaking, a data center will have one type of networking.
Historically, it'll have one type of compute. These days we're now adding GPUs to CPUs, so you have two types of compute.
When we look at storage, it might surprise you to learn that in a typical data center you might find four, five, six different types of storage all attempting to essentially do the same thing, which is store and retrieve data.
If we look at how storage infrastructure has historically been developed and deployed, it's historically been done so on an application or environment by environment basis. ‘Oh, I've got a database. Let me design a storage solution for that and deploy it.’ ‘Oh, I've got a file storage need. Let me go design a solution for that and deploy it.’ And over time, these different application needs have led to different solutions to the data storage problem.
Andy Persteiner: I think of storage as a way to store bytes. I don't necessarily think of it as a way to store files or objects or database records or event messages. I think of it as a way to store all of those things, and oftentimes people will, when you say storage, they'll get focused on a specific area or segments of storage.
But ultimately, if you have data that needs to be stored, it doesn't matter what format it's stored in, it needs to be stored somewhere.
Diana Goovaerts: So, it's less about the interface files versus objects versus databases, and more about how data can be stored, managed and accessed.
Kimball takes it one step further, saying storage now includes the controls wrapped around the data.
Matt Kimball: When I think of storage, I think of everything from holding those bits and bytes to managing the data that sits within those environments.
Oh, and by the way, providing an active layer of protection of that data.
Diana Goovaerts: That active participant idea matters because AI changes the shape of demand. It's not just that you need to store more data, you need to move it fast enough to keep extremely expensive compute busy.
Rob Lee: If you look at where AI technology sits and the promise of AI, which is to be able to bring new models, new ways of analyzing data to apply to data sets across applications, well now all of a sudden the challenge becomes very clear.
How do you do this in a consistent way when the underlying infrastructure supporting each of these environments was designed in isolation as wholly different? And that just brings to the surface a ton of inefficiencies and, operational challenges. And so, I think that's one of the largest forces that's changing storage infrastructure.
Diana Goovaerts: So, if storage is part data management, part data repository, and part protection layer, which storage decisions actually matter the most when it comes to getting results from your compute?
Well, Kimball frames the decision-making process in four buckets: two you'd expect and two that are becoming more important in the AI era.
Matt Kimball: There's a performance element, speed. How quickly can I serve up data to a workload, to an application, to a model, to an inferencing cluster? That's the first.
The second is obviously reliability. I think that's kind of table stakes for most storage systems today.
Resilience. How resilient is that storage environment that I'm deploying?
And then the fourth and probably obviously the most emerging I think in the AI era is how much data management is built into that storage platforms. We're past the point where we are deploying just a bunch of spinning rust or a bunch of flash media and storage, as I mentioned, has elevated itself from simply a repository to a very active participant in the AI equation.
Andy Persteiner: The factors that people need to consider when they're making storage decisions: I think that ensuring that people don't need to move data from place to place in order to satisfy the requirements. That's a key element.
And being able to seamlessly allow for both legacy applications and modern applications to access data sets without needing to make extraneous copies, that's also an important consideration.
Jay Subramanian: The most important thing that is driving storage today is that intelligence to be able to understand what data goes where, who gets access to the data, what is the level of performance required and should it be placed in a high-performance storage or is it in archive storage? In some cases, it's sort of interchangeable and you have to make sure that the decision to move things around happen very quickly. So, that's where the key word that I would tag onto is intelligence. The intelligence to understand all of the different attributes.
Rob Lee: I think that most people, when they think about the attributes of storage they start by thinking about performance, especially as we think about modern applications in AI. And performance is super important, but what people quickly realize is that just as important to performance is reliability, is security, is availability, is manageability, right?
We go back 10, 15 years and you stand up an application environment that thing's gonna run three, four or five years without changing. These days, with AI technology and how fast things are moving, you know, the environment needs could change from week to week to month to month.
And so that ability to be very agile, very elastic, very dynamic, and accomplish all that ability to navigate that change without creating downtime, without creating data loss, without creating a lot of the operational challenges, that's becoming front and center in terms of achieving outcomes versus just clocking a benchmark number but not actually being able to move an environment into production.
Diana Goovaerts: Put all of these things together and you get a simple test. Can you serve the data fast, reliably and safely without multiplying copies and operational overhead as demands change? But that doesn't mean storage comes without challenges.
Kimble outlined some of the biggest.
Matt Kimball: Data is being generated everywhere in the enterprise. It's on the edge, it's on the client devices, it's in the core data center, it's in the cloud, right? So it's the distributed nature of my data environment.
The second is the type of data that's being generated. Now we have unstructured data that is, you know, in the forms of photos and PDFs and recordings and all kinds of artifacts. So there's this hodgepodge of data that needs to be stored and used very easily, everywhere.
The third is the complexity of taking all of that data that's sitting in the storage environment and making it wholly usable by – obviously what is popular today – by AI, whether that's for training a model, or whether that's for feeding an inference pipeline.
Diana Goovaerts: Persteiner describes the same pressure as a scale problem: capacity scale, access scale and even geographic scale when a single data center isn't enough.
Andy Persteiner: People who are used to working in the world of terabytes [are] now migrating to petabytes, or people who are used to living in the world of petabytes now looking at the world of hundreds or even thousands of petabytes, or in another way, exabytes. That's the first point of scale.
The second point of scale is how much access is happening at the same time of this data. Is it handful of users who are running basic jobs or is it a large scale fleet of GPU servers that are performing train training and inference simultaneously on the same data sets?
What oftentimes people don't think about is what happens when you scale beyond what you can manage in a single data center, either for footprint or for power reasons, and then thinking about, ‘Well, how do I scale not only my storage, but also my applications and workflows to be able to be geographically distributed in the event that I need to grow beyond a single data center?’
That's the first place to start is when building something. Can it scale? Not only in terms of physical footprint or the amount of money spent, but also in terms of ensuring that the system can be supported by effectively automated processing as it gets larger and larger.
Another thing to consider is that the types of demands of workloads are changing. In fact, they're changing just as rapidly as many other technologies in the space are changing. And so what ends up happening is the workload that you had on day one may be different on day two or year two, and needing to sort of architect a storage solution and a data center solution around making sure that you don't need to bring in something new or migrate your data or applications to something new every year or every time a new workload is introduced.
Diana Goovaerts: And that's where a lot of teams get trapped: Treating each new requirement as a separate tool purchase. Kimball's advice, zoom out.
Matt Kimball: The first thing I tell 'em is step back from the equation and think bigger picture. When we go and we say, ‘I need to buy a product to handle this element of data management, to do this function,’ and we do that in isolation, we're really almost adding to the complexity that is today's data environment. Right?
I think you need to look at data management through the lens of how I support the entirety of my enterprise needs, not just one function, not just one workload, and not just one or two workloads. I think a lot of folks are getting caught up in this ‘I need to go buy, or I need to go look at these platforms that perform this very single function because Nvidia is telling me it's really, really important,’ and it probably is if NVIDIA's telling you that. But your needs are bigger than just what NVIDIA's telling you.
Don't miss the forest for the trees.
Diana Goovaerts: You've probably heard some version of this: storage demand is spiking. Supply is tight, and prices are moving. Persteiner explains why this isn't a quick fix.
Andy Persteiner: The manufacturers have a fixed amount of capacity they can produce. They are working tirelessly on innovating to be able to double the amount of data they can store in the same physical footprint.
But building fabs is expensive and time consuming and so the amount of capacity that's extracted from them may not scale as quickly as the demand. I think what we're finding is that the demand for so much storage has increased significantly in the last two years, and we don't see signs of it slowing down.
Our forecasting is looking, from our standpoint, at needing to triple the amount of capacity that's deployed over the course of this year as compared to last year. And so that's a very big challenge.
Diana Goovaerts: So, what do you do while you wait for supply chains and manufacturing to catch up? Kimball's take is practical. Stop leaving capacity unused. Be smarter about architecture and tighten up your data strategy.
Matt Kimball: Make use of what you have. As enterprises, we're deploying architectures that don't make best use of all that storage. They actually leave storage on the table.
And the second is really look at how you're making use of the cloud. And third is have a clean data strategy. This is not about, ‘I have duplicate copies of data everywhere in this AI world.’ It is ‘I have, when I say multiple, it's on the factor of many, many copies of data sitting everywhere across my data state.’
That's taking up a lot of precious storage.
Jay Subramanian: Effective capacity is becoming more important. Things like data compression, data duplication, all of that stuff becomes more important. The only way customers can get around this in the near term is to become more intelligent in terms of is this data set compressible, how do I make sure that I can use all of these data reduction technologies without losing performance at the same time, know where it stands, where to save what, where to store what, and how do I make sure that I don't keep moving data between these kinds of things?
Because a lot of infrastructure is just put in place because people are moving data, I don't wanna say arbitrarily, but a lot more than they should. If you have the proper planning techniques associated with it, we should be able to go in and both reduce the footprint as well as also ensure that you don't have to deploy infrastructure that is somewhat unused.
Diana Goovaerts: Before we go, there are two misconceptions worth clearing up. First, that modern storage automatically means all flash, and that modernization has to be a rip and replace project.
Matt Kimball: I think one storage misperception that people need to let go of is that spinning media is dead. AI is not served by flash only.
So when you think about your storage environment and you think about modernizing, you don't have to think about going directly to an all-flash environment all the time for everything. That's part one.
Part two is there's this notion that if I'm going to modernize my storage environment, I need to – if you listen to any vendor out there, because it's in their interest – it's a ‘you need to rip and replace. You need to take out everything you have and replace it with my solution only.’
Really what you need to be looking at is finding a universal control plane that allows you to manage your data estate in its entirety.
Diana Goovaerts: And second, that storage is boring plumbing. Persteiner’s closing line is the kind of perspective shift that tells you exactly why storage conversations are suddenly everywhere.
Rob Lee: I think there's probably two storage misconceptions that people need to work their way out of. I think the first is that storage is a dead technology or kind of a stagnant technology.
I think that there's been, and there continues to be, a tremendous amount of innovation in storage technology, both in media, in storage infrastructure and storage systems, but also in data management systems.
I think the second is that different types of storage and different storage solutions are needed for different applications. Again, I'll go back to if we can power the entire data center with largely one type of networking or one or two types of compute, why do I need six types of storage to go hold my data?
That just kind of doesn't really make sense. And I think that's a significant misconception that the industry has grown up on, but it's time to let go of.
Andy Persteiner: I think people should rethink how they look at storage. Don't just look at it as a necessary evil. It's not just the plumbing that sits underneath everything that you need to have working. It's actually becoming the lifeblood of how organizations are run and how they're going to achieve revenue success in the future.
Diana Goovaerts: Here's the takeaway: In the AI era, storage isn't a passive place where data sits. It is a performance system, a management layer and a resilience strategy.
If your CPUs and GPUs are the engine, storage is the fuel system. And as data volumes and access patterns scale, the teams that win will be the ones who reduce copies, simplify operations and design for growth without constant re-architecture.
To all of our listeners and viewers, make sure you like and subscribe to wherever you listen or watch, and we'll see you for the next episode.