Inside Pure Storage’s FlashBlade//EXA: Scaling AI Without Bottlenecks - Six Five In The Booth

Can enterprise storage evolve fast enough to support the explosive growth of AI and HPC workloads?

From Supercomputing 2025, host Daniel Newman is joined by Pure Storage’s VP/GM Artificial Intelligence, Matt Taylor, to discuss strategies for scaling AI workloads without creating bottlenecks – a critical goal for leaders building AI infrastructure. They explore why organizations are doubling down on high-performance, AI-ready storage, what the latest MLPerf 2.0 results mean, and how new S3/object storage capabilities are driving changes in data pipeline design.

‍

Key Takeaways Include:

🔹Market Shifts in AI and HPC: AI workloads are accelerating, creating a tipping point where infrastructure needs rapid adaptation, prompting Pure Storage to invest heavily in AI-ready storage solutions.

‍

🔹Challenges at Scale: Traditional storage faces significant challenges with large-scale AI and HPC, leading organizations to reconsider architectures and prioritize seamless scalability.

‍

🔹MLPerf 2.0 in Practice: Pure Storage’s MLPerf 2.0 results offer unbiased insights into real-world AI training and inference performance, cutting through market hype for enterprise decision-makers.

‍

🔹Impact of S3/Object Support: Newly added S3/object support in FlashBlade//EXA is transforming how teams design and scale data pipelines, especially for sophisticated AI and ML workloads.

‍

🔹AI Ecosystem Momentum: Increased partnerships with software vendors and ecosystem players are accelerating deployment and operationalizing AI and HPC architectures for the future.

‍

Learn more at Pure Storage.

‍

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍Disclaimer:Six Five In The Booth is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

Daniel Newman:

Hey everyone, welcome back to another episode of The Six Five. We are here in the booth, St. Louis, Missouri, Supercomputing 2025. In the booth at Pure Storage. Excited to be joined today by Matt Taylor. Matt is the Vice President and General Manager of AI and HPC at Pure Storage. Big job right now. Matt, thanks for joining. Yeah, thanks for having me. Really enjoying it, really excited to be here. Yeah, it's good to have the chance to sit down. I kind of laughingly have said this used to be kind of a show for geeks and nerds, like researchers, universities, national labs, you know, maybe a couple of enterprises on the far end of the spectrum that have really big compute needs.

‍

Matt Taylor:

And now it's like an AI enterprise show. Oh yeah, I mean it's almost like GTC level AI focus. Which is a really interesting trend actually is, one of the big things we see is this traditional world of HPC and AI are really merging together. And even like traditional HPC applications, you're now seeing AI loops in them. And so the worlds are no longer distinct like they used to be. It's very much one world. And many times the teams that were the traditional HPC teams and enterprises are now being asked to do AI. One of the big trends we're seeing is these worlds are definitely merging and there's no longer the HPC team and the AI team.

‍

Daniel Newman:

It's one now. Well, there's a lot of overlap in terms of the infrastructure, the amount of compute performance. Now, HPC likes to talk flops, AI likes to talk tops, but really we go flops, tops, back to flops. But there is a massive explosion of AI workloads. In your view, data is such an important part, and where the data is stored, how quickly it's accessible, latency, cloud, all these things are important. What are you seeing driving this explosion beyond the obvious?

‍

Matt Taylor:

Yeah, I think there's two big trends that I see, and they're kind of two different sides of the market. So on the big foundation model builders, the open AIs, et cetera, just the amount of data that they are using to train these models is exponentially growing every day. Because as they train the models and deploy them, then they get more data from the usage and feed that back into the model. And so for these large models, the data is just exploding. And then on the other side of things, with real enterprises, what they're recognizing is they have these just treasure troves of data, but they don't know how to actually bring the data to the AI models. And so what they're really seeing is, I need to modernize my overall data infrastructure, move data out of these silos, and make it accessible to these great LLMs and great JetAI models. But one of the challenges they're having is actually doing that. And so at Pure, we're trying to do a little bit of both. So we've launched this new platform called Exa, which is really focused on high-performance storage for the large language models and the Neo clouds. And the other side, we're spending a lot of time on data pipelines and helping customers break out of their silos and get to use that data to generate more value.

‍

Daniel Newman:

Yeah, so the storage side's evolving quite a bit too, right? So, you know, as AI continues to proliferate really quickly, exponentially, more data needs to be more readily accessible in real time. I mean, the enterprise statistics that I'm hearing is like 90 plus, maybe 95% of enterprise data is behind the firewall. And of that 90, 95%, maybe 1 or 2% has actually really touched AI yet. hugely opportunistic, but it definitely changes how you got to think about building your storage, right? Yeah, completely.

‍

Matt Taylor:

I mean, you know, the traditional sort of way you thought about is, I deploy an array for an application, and if I need to do more of those, I just kind of replicate that, creating these like major silos. And now it's like, great, I have all this data, but how do I get my most precious data, like customer data, to these models becomes a really hard challenge. So you have to think about not just the infrastructure, but the layer above that. Think about modernization of Kubernetes and containers and then the tools that you're going to plug those into. It becomes a really different challenge versus just needing to have an application and keep it stable and be able to support X number of users. I now need to figure out how to take that data and actually get more value out of that. And that's where we're spending a lot of time helping customers figure out that strategy as well as figure out the infrastructure that needs to go along with that.

‍

Daniel Newman:

So as part of MLPerf, they're starting to look at storage technologies that are able to support ML and AI workloads. Recent MLPerf data came out. What was your takeaway in terms of how you scored, how Pure Storage landed in those new results?

‍

Matt Taylor:

Yeah, so full transparency, you know, we at Pure have not done a lot of this historical benchmarking work. And so when we launched the FlashBlade EXO product, which is really focused at high performance storage for the Neo clouds, the foundation model builders, people that really care about the performance of the overall system. We made a decision we were going to go and invest heavily in actually building out a performance lab and be able to do these results. The first one that we chose to do was MLPerf. The reason why we like MLPerf is it's a great storage test that actually tests GPU utilization based on the storage infrastructure. So what it basically tests is, okay, I need to be able to utilize my GPU at a certain rate, and then how many of those can I do in parallel? And so we chose that as the first benchmark we wanted to publish. What we were able to show is that we were up to 2x better than our closest competition on ML Perf, the three training benchmarks they have. And it's really a testament to the system we built that really focuses on massive throughput and scalability. So we're super excited about that. We've got a bunch more coming, but that was really the first one we focused on because it really does show the value of storage and being able to not bottleneck your GPUs and get the maximum amount of performance and usage you can out of a pretty expensive asset.

‍

Daniel Newman:

Yeah. Rapidly evolving and changing assets too. It's exciting times in that area. You also have evolved in your types of storage, right? Object. Talk a little bit about why you went that route and what does that mean from an AI standpoint?

‍

Matt Taylor:

Yeah. So, we've had object support natively at Pure pretty much from the beginning of our FlashBlade product. So, 10 plus years we've had object support. But it has been candidly mostly used in the cloud, is really where a lot of S3 has been used. And enterprises have started to use objects more and more. What we saw with AI was this just explosion of massive object tiers of data. And a lot of this is coming from the foundation model builders and the large Neo clouds. that are looking at their data available to them. And what they realize is, they've got these massive data sets, but they're pretty much in an object pool in a cloud. And what they want to do is figure out, how do I get that data into my models? And so, the process of taking that object data, basically moving it into a file format and into your training set, has become challenging when you get to these really, really large data sets. And so what we decided to do was actually take the EXA product, which is really focused on this really high performance training and checkpointing use case, and then actually make object support, bring the entire object stack that we had from our FlashBlade product into the EXA product. So the benefit we get is a really, really robust S3 compatible feature set with the massive performance we have in this new platform. And the reason why we've always said we were going to do this, but we got a lot of requests from the big foundation model companies. And actually, the reason why we actually pulled this in and announced it is we're actually starting to actually do this with customers in a limited fashion for these really, really large foundation model building use cases.

‍

Daniel Newman:

So everything's moving really fast. Part of the challenge of being a vendor leading in technology innovation is the entire ecosystem. You've got your partners, you've got the customers. Enterprises have historically not moved quite as fast as, say, hyperscalers when it comes to adopting. You actually just talked about that a little bit with S3. But at the same time, your partners really need these new technologies, you need to pull them along, they need to pull the enterprises along, because this change is very real. Talk a little bit about what you're kind of seeing in the partner and customer ecosystem and how that's moving along.

‍

Matt Taylor:

Yeah, I'd say, from a pure perspective, we do everything through channel partners, so we're very channel partner-centric. We have a very good, I think, understanding of what's going on in the channel. What we've seen is there's a handful of our partners who have really leaned in heavily and have invested in building practices and specialized sales forces to go after this big AI and HPC space. And so we've picked a handful of them and I'm starting to invest pretty heavily. Some of the examples of this is we actually just went into the WWT AI Proving Ground here in St. Louis, actually just got the EXA platform up and running there. And so, you know, that's really meant to help customers test new technologies, actually build a complete stack of their application they want to go build and deploy, but do that in an environment where they have access to the latest and greatest technologies from a variety of vendors, not just us, a variety of vendors. And what we're really focused on is trying to help our vendors be an accelerant for customers. So work with our partners, work with the WWTs, the Penguin Computings of the world, to really help customers say, I want to adopt a new technology. I can get all the pieces of this really rapidly changing ecosystem together and bring a solution to a level of maturity before I actually go and deploy it. And so we're going to continue to do a lot of work with all of our partners on the software side, on the channel partner side of things, and increasingly obviously with NVIDIA as well. We'll have our bunch of certifications coming out here over the next few months as we've got to keep doing the certifications and keep making it easier for customers to go and deploy these challenging and rapidly changing technologies.

‍

Daniel Newman:

You've got to pull them along, give them the education, the knowledge, the access, build those POCs, help them develop them, get them to scale. It takes a village. Right now with AI. For sure. Matt Taylor, thanks so much for joining me. Great to meet you. Thanks so much. All right. Thank you. And thank you, everybody, for being part of this Six Five. We're in the booth here at SC25 St. Louis. We're with Pure Storage here today. Hit subscribe. Join us for all the coverage here at Supercomputing25, and also, of course, all the great coverage on the Six Five. But for this episode, I got to say goodbye. See you all later. 

‍