Good afternoon. Welcome to the HPE Discover Investor Relations Summit. All participants will be in a listen-only mode. Should you need assistance, please signal a conference specialist by pressing the star key, followed by zero. Please note this event is being recorded. In just a moment, I will turn the conference over to Shannon Cross, Chief Strategy Officer.
Certainties materialize or if the estimates or assumptions prove incorrect, our results may differ, perhaps materially, from those expressed or implied by such forward-looking statements. HPE assumes no obligation to update such statements. Please find more information regarding forward-looking statements on our website at investors.hpe.com. With that, let me welcome Antonio Neri, HPE President and CEO.
Good afternoon.
Antonio, we all enjoyed your keynote today and Rami's talk during the networking general session. We've also had an opportunity to tour the show floor. It's clear HPE is leading the adoption of agentic AI in the enterprise. We're leveraging our innovative networking cloud and AI portfolio, helping customers move from AI experimentation to fully autonomous operations at scale. During your keynote, you shared that we are deepening our work with NVIDIA with the next phase of AI factories and HPE Private Cloud AI. We are excited about the opportunity to work with AMD on Helios. Rami also talked about HPE extending our leadership in self-driving networks as critical foundation for agentic AI from campus and branch through data center.
I'm excited by what we've announced so far at Discover, and I hope everyone listening on the webcast, as well as everyone here in the room, can tune into Fidelma Russo's general session tomorrow, where she'll discuss how HPE's innovation will continue to support customers on their AI journey. I look forward to this Q&A session with the investors and analysts, and we'll take questions from the audience shortly. First, let me kick off with a couple of my own. First, networking has become such an exciting part of our portfolio and our story since we combined with Juniper. How do today's announcements show the extension of our networking leadership in the AI era?
First of all, good afternoon, and those who are tuned in on the webcast, thank you for joining us today. For those here in person, I hope you enjoy so far the event and the day. I understand you did the tour of the show floor, which it takes two days to see everything. I hope you got a glimpse of the amazing portfolio that we have curated and built. At this event, you can see the continuous innovation we continue to bring to the market to address the needs of both cloud and AI. Obviously, the biggest topic is AI. In that context, we think about AI as a productivity tool that will change forever how we live and how we work, but fundamentally is to power it. At the core of that powering is the foundation which sits on the network.
We talked about the need to improve the cost per token or the first time to token, and fundamentally every aspect of that stack needs to be productive. Today the network is a bottleneck, no question about it, because we saw the tremendous advancements with the compute and accelerated computing. What we have done with networking and the portfolio that we built with the acquisition of Juniper is address the demands of AI and cloud from the edge of the network, which obviously is the on-ramp for many things, including going forward, the inferencing component of this, all the way to the training side. I thought what I covered this morning and what Rami covered whatever, just an hour ago, is a testimonial about how well this integration has gone for us.
It is not just integrating two great companies and two great assets, which were very complementary with each other, but really scaling that integration with innovation perfectly timed to the inflection point. Whether it's in AI, scale up, scale out, or scale across. We have an amazing portfolio, and that's why we see the results we saw in the previous quarter in terms of orders, in terms of backlog, obviously because of the supply availability, and in terms of durability, because the network for AI demand is insatiable. For us, I think we are perfectly timed for that moment. Look, I believe in the agentic enterprise is also using agentic AI to make these solutions more autonomous and intelligent. This concept of self-driving network is something that we started with Rami a while back, but now it's live. It's available.
If you see some of the demos we put on the floor, even on the routing side, it's a little bit scary to see how far you can self-optimize traffic across data center interconnect. To give a sense, that 12,000 router, you can put together the entire population of New York and London together, and concurrently 60 million people can't broadcast a movie or watch a movie on that single rack. The performance and the ability to do it in an autonomous way, it's just remarkable. That's just an example, that's why we say, well, we'll extend it to everything. We already had it anyway in the campus and branch, and now we brought it in through the Aruba side as we brought the switching to the Juniper side.
I think it's fantastic what we've done. It's a testament to the hard work I think the team did in terms of the integration.
Absolutely.
The IT that sat in Juniper and in Aruba and now is being supported by the total of HPE.
If you think about it for a moment, right? We closed the transaction on July 2nd. By January 2nd, just exactly five months, we brought in 10,000 Juniper employees inside the company. We announced our strategy for networking. We announced the roadmap, across the four key networking segments, campus and branch, data center, switch and security, and routing. We integrated the sales force into one unified sales organization.
We announced all these products along the way. Of course, we are doing really well from a synergies point of view. What comes next now is that vision to build the best networking business on the planet, and that includes also the back end of how we do business. Fundamentally, also the next chapter is also the synergy with the rest of the portfolio, particularly with the cloud portfolio. Which we are integrating products, whether it's software in the virtualization stack or whether it's in the private cloud stack or whether it's with storage. Which are sources of revenue and profit as we think about 2027, 2028- 2029.
Well, that's a perfect segue to my next question, and then we'll open it up to the floor. We've come out of a great quarter. How are you thinking about the durability of these results? What gave you the confidence to provide the fiscal 2027 framework that had called for double-digit growth for basically both revenue and EPS at the midpoint?
I think at the core is structurally, the portfolio of our company has changed forever with the addition of Juniper, right? The mix has changed dramatically. I argue we're still undervalued in many ways from a P/E multiple perspective against not just today 2027 guidance, but against the long-term potential. Especially because the networking demand is very, very high. My view is that, first is the mix of the business. Second is the demand that we see in the market, right? We grew in security mid-teens all the way to the routing and 30%, and in between you have upper 20% in the rest of the portfolio. We have an enormous backlog, obviously, that we need to clear. What give Marie and I confidence to give six quarters of guidance, because that's exactly what we did, second half and 2027.
The pipeline is multiples of our backlog. Number two, our portfolio is being seeked by customers. Networking clearly is the driving force that's doing that. Obviously, the supply constraints in many ways are driving demand because everybody wants to get in the line, in the queue to make sure they don't have to wait too long to get the supply. Let's be clear. That supply problem is not going to be solved anytime soon. It's like I make the analogy, one time happened to me, I made a mistake. I went to the DMV. I took the ticket. I have to wait, whatever, 40 people ahead of me. I left, and then I come back the next day. Shoot, now I have to wait 60 people. Right?
Look all the programmatic things we have done with Juniper synergies and Catalyst, particularly Catalyst in the way we work inside the company and the ability to improve our gross margin profile and operating margin profile to the OpEx. It was a no-brainer for Marie and I to go out and give all that. Which for us was the concept of durability because that's the key here, right? It's not one-time. As I said in the earnings call, Q2 was not just a one-time event. It was a combination of many things we have done for many, many quarters. In the ISL, we have been very disciplined about what capital to deploy for what return. That's it.
Great. Well, thank you, Antonio. With that, let's open it up to the questions from the audience. I see hands.
These are 115-
We will have mic runners, so please wait until you have a microphone before you begin. We are webcasting this session, so please state your name and company as you're asking your question. Finally, can you just stick to one question? We will come back around as time permits.
We are under the earnings rules here.
Yeah. Wamsi, you have the mic.
I'll stick to it. Wamsi Mohan, Bank of America. Thank you for doing this. Nice to see all the integration progress you made. We heard a lot of exciting things about the networking portfolio. When you look at your Q2 results, 10% growth, there were differences within the subsegments in there. You've guided next year 8%-12%, I'm thinking, like, why should that not be viewed as a very conservative guide, just given the fact that we heard so many things here in the pipeline that are coming through in the next several quarters?
Yeah, because you guys always look at revenue. We look at orders and ability to convert to revenue. That's the thing. Look, in many ways, it was a prudent guide up from a pure revenue perspective, which is what drives profit and eventually free cash flow. As I said, the supply availability will continue to be severely constrained into 2027. To give a perspective, we already have the capacity allocated for 2026. What we do, the way it works, every 90 days, we tell our suppliers how we want that capacity to be divvied between, I don't know, server storage, between this 64 GB to this 128 GB to 256 GB, this speed, that speed.
I met one supplier here, which is a great partner, I said, "It's great what we're doing together, but I need more supply." Then he goes on and says, "Yeah, we'll see what we can do." Right. It's really that the issue, Wamsi. We expect to exit 2026 with a higher backlog in many ways than we are today. We felt it was prudent to give that double digits based on what we believe we're going to be allocated in 2027. That process is a process that's still going today because we have no firm final numbers, because we have negotiated now the LTAs, those LTAs are not just one year. Now they are multi-year commitments. We need them to come back with the final. That's the reason why.
We expect in many segments of our business to continue to grow faster than that number on an orders perspective. Yeah, we're going to go back.
Tim?
Tim was first. Yeah.
Thank you. Tim Long at Barclays. Sorry about your ticket, but given what the stock's done, you didn't have to go. You could have just paid the full price, I think. Anyway.
The ticket for what?
The DMV.
Oh. Oh.
I wanted to get back on networking if I could, a two-parter. You showed a lot of really good technology today, in that AI piece, you raised the numbers a little bit last quarter. Two parts. One, could you talk a little bit about leverage of the strength you guys have in server and storage, and have you started to see that at all impacting that line? Second, some of these newer layers like scale up, where you have the nice AMD deal, and much more importance on scale across. Can you talk about how that might impact positively that AI data center line? Thank you.
Yeah, Tim. Fair to say that one of the areas that are growing the fastest is actually the scale across, for sure. I think in above the 10,000, 12,000 and the MX product are becoming key references for DCI and on-ramp to the edge. Scale out, which is the QFX products, continue to grow, and we have a number of marquee customers which are adopting that. That, think about it in the case of NVIDIA, right? NVL72. That's the NVIDIA GRID announcements we made at GTC. There is a number of hyperscalers and I'll call it large service providers in neocloud that actually have adopted the QFX on top of the Spectrum-X to do the scale out, right? There is a number of reasons for that. They love the management control plane. They love the AI operations that we built into that.
They love the performance of the actual switch. As you saw, we announced the 1.6 Tb first industry time to market with Broadcom on the Tomahawk 6. Even there, we offer the two distributions between Junos OS and SONiC-based OS, and both have AI embedded into it. With the Junos OS, you get more telemetry by definition. The reality is that, depending on the type of customer may pick one versus the other one, how their environment work. In the case of scale up, in the NVIDIA case, it's NVIDIA. That's a given because of the Spectrum-X, ConnectX and BlueField. In the case of Helios, which I hope you saw there, is I call it a double fridge decker. It's an OCP design.
That's going to be our SONiC OS, with our 52/52 QFX switch, which we are excited because as customers adopt alternatives to the NVIDIA for certain training, large training environment, this is a large environment for training, obviously that's our networking in it. Everything else will scale out and scale across, will continue to be the same with Juniper. That obviously gets deeply connected with our compute. Inside that Helios or NVL, there is an HPE server that comes with it. Right? In the case of the AMD, there is a tightly integrated work between AMD, our server team, and our networking team. It's not just GPUs from AMD with the network fabric. It's actually all three together. Obviously, you have to see it at the fall when that comes out, right?
I expect a number of large marquee hyperscalers and/or service providers to adopt that. That will be clear tailwind for us as we go forward. Right? Yeah, Asiya. Oh, behind you, and then you.
Great. Asiya from Citi. There's investor perception that in this cycle where we have so many component constraints, HPE has done a really great job, and specifically as it relates to networking chips, access to having silicon. Maybe you can just help us understand, why do you think HPE is better positioned in this cycle to get access to components? Where could there be some upside to driving better component access as you go into FY 2027?
Well, thank you for the question. I think it's important to remind ourselves, our portfolio and what IT we own across the portfolio. First of all, on the routing side, we have two dedicated silicon roadmaps that we run. One is for the router MX, that's the Trio silicon. Our design, our silicon. The other one is the PTX, which is our Express 5 silicon. Basically, we do not use merchant silicon for any of our routers. That's number one. Number two, when you go to the campus and branch equation, our entire HPE Aruba CX portfolio, maybe just a couple of products on the fringes, is our silicon. We designed that silicon for many, many generations now, and it was part of the original portfolio we owned, which I reversed it, integrated into Aruba in 2015, which is our ProCurve business.
Now that silicon is across the entire campus access layer and aggregation layer for the campus and branch. What we announced today is that those switches now are also managed and available to the Juniper Mist. Okay? In those two aspects of the portfolio, we have our silicon. We don't buy anyone else's silicon. When it comes down to security, you will see very quickly that that campus switching silicon, I just talked about it, is converging with security. Unlike others that converge security at the software layer, we are converging security at the silicon layer. The next generation of CX switches will be a converged silicon between networking and security. That's a unique value proposition that is going to give us a huge advantage because that silicon is truly programmable.
All the algorithms are built in the silicon, so we can program that from our cloud control plane, whether it's Mist or whether it's Aruba Central. Then in data center switches, we use, of course, Broadcom. Broadcom, we are now the largest OEM partner for Broadcom. That's another reason why we have an advantage when it comes down to that. Look, there is constraints there in networking too, and the networking constraints are mostly aligned to the same constraints you see in the market, which is memory. Even though the memory footprint in a networking switch is much smaller, it's actually the older technologies, the DDR4, not even DDR5, which people have, of course, de-emphasized. We are moving to the latest design in some of these switches. In fact, some of them, we may skip all the way to HBM.
Lou?
Lou?
Okay.
Yeah, we go here.
Thank you, Antonio. Lou Miscioscia, Daiwa Capital Markets. Looking beyond networking, I guess, for a moment, you had some very good questions on that. Can you differentiate the demand and the highest supply chain constraints, for, let's say, GPU servers, AI GPU servers, CPUs, Vera, normal x86? Maybe, leading the answer here a little bit. Few years ago, thought that maybe CPUs would see 10 times the capacity needs of GPUs once you get to inference. Seems like you're starting to see that, maybe you could help us with differentiation.
Yeah. On the GPU side, I would say there is not severe constraints, the model works slightly different. Unless you're willing to invest way up front and take a bet, you normally place orders or POs based on orders. You don't place and build a huge inventory. We learned our lesson early on in the cycle, right? Especially with the life cycle of these GPUs moving so quickly. Maybe at the beginning, it was great to have some, look, if I knew what I knew today, I would have bought power and GPUs. That was not the case. It's less about the constraints, it's more lead times. If you build, it's going to need to build a large AI training system, that's based on lead time, based on where you place the POs.
Obviously, generally, it tends to be a three-way conversation between the customer who has sometimes significant relevance with NVIDIA and then just us, right? We work together on that prioritization. However, there are other constraints around that. Power loops, cooling loops. You think hoses will be a problem? That's a problem. In networking transceivers, that's a challenge. There is a number of things, peripheral things that goes around that. On the CPU side, there have been constraints. I think we have done a very good job navigating that. Look, you can have the CPU, but if you don't have the memory, it's a waste of time. It's like, I have the car, I have no wheels. Okay, that's not very helpful, right? You have to move as a system.
Look, we have done a very good job in partnership with both AMD and Intel, because of a long-term relationship. Vera CPU is just early, right? We just introduced the product, I think that product is going to do great in the inferencing space. When I think about between now and end of the decade, the vast majority of demand will be in the inferencing, not in the training side. That's our view. Now the question is, number 1, where the inferencing will be done, what type of architecture you're going to deploy for that inferencing, it's not going to be one kind of unique architecture. Use cases and verticals will variate. Ultimately, will be more centralized or decentralized. These are all things we're going to learn as we go forward.
The ratios, CPUs to GPUs, I think is going to variate based on the type of inferencing, whether it's four or eight, I don't know. Look, having a server business that has that scale is important in that context. What I'm working with the team is that, okay, how the architecture really collapses. You don't want to build more layers and overhead in that architecture. I do believe there will be new emerging architectures between KV cache, cores, and networking fabrics coming together in a more efficient way because you have to solve for scale, cost, right, and eventually energy. A lot of things will happen. That's why in my head, I always think about it, I have a server, I have a storage, I have a networking. No. I have network fabrics, I have cores, I have software, I have memory.
Storage is an extension of memory. Then how you bring it all together in a way that creates some differentiation. Yeah, he was asking for a while.
Thank you, Antonio and Shannon. Erik Woodring, Morgan Stanley. Antonio, can you maybe help us just to build on that question by contextualizing how your enterprise customers' compute needs are really changing in this environment with agentic AI, with inferencing, more getting done on-premise? The question maybe is this customer's materially growing their server installed base? Is this just refreshes to modern architecture, the sustainability behind that? Would just love to understand, because we do seem like we're at an inflection point, just the context behind that that you're seeing would be really helpful. Thank you.
Yeah, sure. Look, first of all, I spend more than 50% of my time with customers because I always say the truth is in the coal face when you talk to customers, right? They will tell you exactly what's going on. Fair to say, maybe a year ago, things were a little bit slower. People were sitting on the fences understanding how this is going to evolve. It's fair to say, at least in the last 6 months, there is a significant acceleration. We're still early. That's the interesting part, right? If you look at the ratio of compute for training versus inferencing, it's still maybe 70/30 right now. At some point, that has to go the other way around. Then again, where it's done and what level of scale you need. Look, I give you the example of us as a company.
We as a company are aggressively using AI everywhere we can. We have 1,200 + use cases in the life cycle. 250 are in production, meaning already deployed in production. Marie is not here, but Marie in finance have been super aggressive deploying AI everywhere she can. This is part of the modernization with Catalyst. Interesting, we are striking some unique partnerships where they use our AI factory for enterprise, AKA Private Cloud AI, to develop the agentic models that eventually get packaged, and we can take it together to market. That's an example of Deloitte with Zora AI suite for CFOs. When I talk to customers, and I was a couple of weeks ago in Chicago and in Europe, they are aggressively moving forward.
What is interesting, people started with large language models, we call it frontier models, that's fine because it's very isolated, very contained, and you can see the benefit of it. I think the way you become an agentic enterprise is by actually bridging together, stitching together agents in a workflow by first digitizing, automating, then deploying AI on top it. There is a little bit of process engineering going on, understanding value streams, then doing the hard work. One of the barriers has been governance, data, preparing the data, and all the regulatory goes around. That's why this morning on stage, I talked about everything we built in that AI factory. Unlike some of the competitors we have, they just resell just the hardware.
We went on and built an entire software ecosystem inside our GreenLake cloud that ultimately the infrastructure that sits underneath is tightly coupled for that, whether it's RAG, whether it's small language model training or giving context to multi-modality, and so forth. When you say, "Okay, how many GPUs are you going to use?" I can tell you inside our company, 300. Don't need more than that, right? What I always said is that if you picture on a whiteboard two axis You have the axis of training and the axis of inferencing, right? The question to ask. Clearly, training is all GPUs for the most part, and the inferencing will be a mix of things. If you take a different view of that, which is service providers, model builders, hyperscalers, these are very low number of customers, right?
Although it has been growing now because everybody wants to participate in the build-out through some sort of financial engineering that's going on. Let's say 50, that they are big enough to make a difference. Those 50 will consume millions of GPUs. It's like here to here. You have the opposite, which is hundreds of thousands of customers who are going to consume the opposite, very low amount of GPUs. This transaction value is significantly lower than that transaction value. That's why you have to find the right balance. Ultimately, I argue it comes down to honestly working capital, and that return on that working capital, and obviously the margins you can generate. More software, more services embedded into it, the better it is because if you believe end of the decade, that's what it's going to be.
You want to be ahead of that and now build an enormous revenue with huge amount of capital that eventually that's going to be a tough compare. In either case, we lead with networking. Doesn't matter.
Katherine.
Oops.
Thank you. Kat Murphy from Goldman Sachs. To maybe extend on the conversation around the AI server opportunity, talking about sovereign in particular. Realizing that sovereign use cases are not monolithic. Can you share anything to think about the size of the opportunity and the maturity of the market, and then where that falls on the training versus inference spectrum, and what in HPE's portfolio and your existing relationships is unique in kind of addressing that opportunity? Thank you.
Yeah. Sovereign is a unique customer segment because it's a combination of many things. You have what I call the traditional labs and government entities that they are making investments in order to deliver an AI cloud under the principle of sovereignty, a lot driven by the geopolitical environment we live in, also as acting some sort of service provider for the communities that obviously their country or their regions are. Look, their ability to raise capital, it depends on the geo. In the United States, it's very easy. How much you need and how much you're willing to pay. You go to Europe, is an ongoing fight, I would say. They have a lot of ideas, but the ability to raise money is complicated.
In fact, I was with one he said, "We are very excited." I said, "Great." "We're going to build a gigawatt factory." I said, "Fantastic. How much money you have?" That was the simple question. "We have $4 billion." I said, "Yeah, I can't get out of bed for $4 billion." Right? It's just literally, the understanding of scale is so off, right? It goes through that process where basically they're trying to attract the private sector to bring the capital. Through regulation, through geopolitical insight, their own country is long. It's a long sales cycle. The bottom line is a very long sales cycle. Look, we built some AI clouds already. If you think about the U.K., the AI in Bristol cloud is a sovereign cloud.
If you think about the European Union, the LUMI system is an AI system in Norway, it's in service of the European Union. These are examples, we are working with a number of them. I would say it's maybe 12 to 15 in total, right? What's going to happen, the sovereign AI clouds are going to be these neoclouds that have the capital to become the driving force. They're going to act as a sovereign AI cloud, although they are a neocloud, just happen, I'm going to build it in this country. I'm going to open with the same regulations so that you feel that we can serve you. Obviously, Middle East was going strong until the conflict started, that put a significant slowdown to the process, although, UAE is still going, I will say.
This is the challenge, right? That we see. It's not just AI, because all of them also need supercomputing. Supercomputing is very important because it's an incredible adjacency to AI. If you look at the United States, all the national laboratories are supercomputing entities. Oak Ridge, Argonne, Los Alamos, Lawrence Livermore Lab, all are HPE supercomputers. Now, we built those three years ago, in some cases. They're all exascale systems. Now they are adding to it an AI system. In the case of the Oak Ridge National Laboratory, that it was Frontier, the first exascale. Now we're going to what you saw a little bit of a cat two or two cabinets, mission. Next to that, you have Lux, which is the AI Pure system.
That for us is an opportunity to grow in sovereign because we have the expertise and the trust, which is super important when it comes down to sovereignty.
The other part is the networking.
The networking, obviously, look, we're trying to be simplistic in the way we tell stories so that it becomes super technical. That's why I said, "Rami, go do that later." It's like when you build your house, you're not going to start putting trimming around the doors and all that until you lay the plumbing and electricity. To me, the network is exactly that. Without the plumbing and electricity, and the plumbing being the network in this case, you're not going to finish the data center. Then obviously you have the cooling that goes with it. Networking is going to be that core foundation, because without a robust foundation, you can't deliver the rest.
Victor, did you have a question? No.
Thank you. Victor Santiago with Evercore ISI. Antonio, could you help us better appreciate how customers are navigating this higher-priced hardware environment we're in based on the conversations you're having with customers? Just given the recent strength in traditional servers you guys saw, where prices are now multiples of where they were from prior year, how are customers managing their IT budget allocations? Are they extending life of existing assets or reallocating spend from other areas?
Look, of course, they're going to try to extend where they can. However, the need to modernize their infrastructure to be able to adopt AI is stronger than ever. Look, if you're going to host this on-premise, you need to find the space and the power. The demands for power are pretty significant. When you look at the Helios Rack, there is a chance, depending on how you deploy that thing, you may need 700 kW power. That's whatever, 5 ft wide, whatever. The reality is that they all need to modernize and save space and save cooling. We can show customers that we can take seven Generation 10 servers of any vendor for that matter, and reduce it to one. That's a seven to one reduction just on space.
We can save up to 65% energy, then increase the performance by a factor of X in terms of core and memory density and so forth. That's why we see the momentum we saw in then what we call traditional servers, which was the orders were up triple digits year-over-year. Of course, a lot was also driven by the cost. Cost is clearly cause of concern, but it's not the reason why stop it. They are not stopping. Then we have a portfolio with HPFS that actually helps them. With HPFS, we can come in, accelerate depreciation of those assets, remove legacy infrastructure, then free up the capital for them to reinvest where it makes sense. In many cases, they actually pivot to OpEx versus CapEx.
It's not just, I don't know, free up $3 million, let me buy $3 million of CapEx. They actually shift to OpEx because they believe, particularly with AI, it is probably a proven way to start small and then start growing from there on. That's why our portfolio is not just technology, it's also the financing capability. It's all delivered also to one unified control plane, which is our GreenLake cloud. Everything I have done now for 5 years-plus, doesn't matter how you pay, you will be using our GreenLake cloud to manage it. That includes a subscription model to the software associated with that infrastructure. There's a number of ways, look, we have not seen a slowdown because of the cost. If anything, we have seen an acceleration.
We believe that's going to continue to be the case because even in 2027, that cost curve will be more stable, but it's going to stay very elevated. Do I wait 18 months, two years? Who knows, right? Yep. Michael.
Thanks so much. This is Michael Tsvetanov with Wells Fargo. With your new networking product announcements, which you talked a lot about today, obviously the integration of Juniper going quite well, it seems you're very well positioned to attach your networking, alongside the rapid server growth you're obviously seeing. I'm curious if you can just speak to the attach rates that you sort of see today between your compute and networking, maybe if you would expect that to trend moving forward. Just generally, I assume as customers upgrade their compute portfolio, you kind of need to bring networking and storage with it. What's the timing lag between those two things? If you can just expand on that'd be really helpful.
Well, look, we are very early in the process in term of integrating that data center switching with the rest of the portfolio. On the AI front, that's happening at the customer side. Because again, in the case of NVIDIA, you sell the NVL72, we win the footprint above on our own architecture for scale out and eventually scale across. There is no really attach of anything. You need to win the footprint inside the data center in the AI space. When you go to the traditional enterprise or cloud, that's where the data center switching attached to the rest of the portfolio is very important. The way you do it is not just here is the switch. You have to integrate the entire life cycle.
What we are doing- we already kind of done is the integration of the lifecycle provisioning management for networking, which we call Apstra, with Morpheus Enterprise, eventually with the full rack for private cloud. At that point, you basically slide the switch into the top of rack and you're kind of done. The hardware part is the easy part. It's not different than any other vendor top of rack. The hardest part is the software integration, and the team have done a great job. That's one part. The second part of this is the integration of the software-defined layer of networking with Morpheus and VM Essentials, which is our virtualization and container environment. That's done. We did that in record time.
Now in our Morpheus Enterprise software, not only we provide orchestration and brokering with the public cloud and on-premises, but also we provide the full software-defined layer from network to compute, obviously virtualization, all the way to the storage layer. We integrated there also the OpsRamp for multi-cloud and multi-vendor observability. That will drive an attach rate, of course, of networking with compute. On the storage side, as we now have the 1.6 Tb, there is a lot of workloads that are perfectly tuned to use Ethernet as the fabric inside the storage. That's going to be a Juniper QFX switch inside our Alletra MP portfolio, where it makes sense. Once we integrated the Apstra software with our switches and with our control plane, everything else just flows.
This is why we expect that the revenue synergy of this will start 2027 and continue. The private cloud footprint is the perfect instantiation because a server storage network and an infrastructure with a software-defined architecture with GreenLake all in one tightly coupled package, whether it's virtualization or AI, because it's the same thing. The only thing that changes is really what type of GPUs and storage you use underneath. In that case will be a server with GPUs, and in other cases will be the storage file and object for unstructured data. I think Wamsi, you had a question? Yeah.
Yeah, thank you. Wamsi Mohan, Bank of America. Antonio, in this world of agentic AI, where there is more attach of traditional server storage, why should HPE not consider maybe selling to Tier 2 CSPs in this world where the margin structure of the entire portfolio can be higher relative to maybe the opportunity of selling just AI-based servers?
Look, the traditional servers in the context of hyperscalers and large service provider, that already moved in the cloud space long time ago, right? If you recall, in 2017, I decided to stop selling that because the margin rates were inexistent. They, in many cases, have their own design to begin with. In many cases, they have their own silicon. If you think about AWS Graviton and so forth. There are elements of the hyperscalers where we participate and will continue to participate. It's more the edge of the hyperscaler, where they use products like us as their on-ramp into the large data center, right? That's more the distributor element of the hyperscaler as a point of entry to the large cloud. Once you are inside the large cloud, it's not like I'm selling ProLiant DL products or any other products.
In fact, none of us, including our competitors, sell into that. They go straight to the CMs or ODMs who build that for them because they have a unique design based on the footprint, how they laid that cloud with CPUs. I think for them is another CPU recipe that now just serve inferencing for those who centralize. If there is an opportunity, of course, we will do it, but I don't think that's going to be the biggest opportunity in my mind. We're going back there, and then we'll come back. Moving back.
Austin Baker with Loop Capital. Just to kind of double-click on some of the things you've already talked about. In the last six months, you've talked about the significant acceleration in demand. Could you elaborate a little bit more on how much of that you're seeing is due to inference on-prem, and how you think HPE will kind of continue to outperform its peers if it is kind of in an on-prem world with inference?
What we're seeing is that, and it's a pattern that we see, right? You have to see multiple data points over time, when enterprise customers in particular verticals decide to go all in on the AI, and they pick a mix of expert's models. Once they attach that data to that model, they tend to want to have it under their control. Yes, we are making things like private cloud with the security around punching an API out more secure. There are industries that even though you make that secure, they said, "No, I'm going to have it on-prem, and I don't want anybody to touch it." There are number of verticals and use cases that will be definitely on-prem, and as they grow in deployment, the inference is going to grow.
That's why I said earlier, we believe that at least 80% of the demand by 2030 will be inference. What we need to see is what's the mix between on-prem, off-prem, meaning large centralized environments. Interestingly enough, how many will be at the edge? We showcase, I think, on the floor, you have four products that they are designed specific for inferencing at the edge in large distributed enterprises. You can see, warehouses of sorts, manufacturing floors of sorts, healthcare where they need to process. There's a lot of regulation. They are HIPAA and everything else, that putting around all this control is more expensive than just putting an inference server in their environment and be done with it. It's just math and physics. We saw it.
We believe it will continue to grow, but we need to see more about eventually how these architectures evolve and ultimately what decisions to make. Great. He has a question too, if you can bring the mic with him. I'll get it to him. We go back to Tim.
Cool. Just very quick one for you, Antonio. Erik Woodring, Morgan Stanley. Quantum. We saw a quantum computer out on the show floor. Some of your competitors that have talked about quantum benefit significantly when we think about valuation multiples. How relevant is quantum to HPE?
I will not comment on valuations of sorts because sometimes, look, you can go back to other valuations of other companies. I'm speaking from a practical point of view, and obviously as an engineer, I used to be at least one, I'm bringing a pragmatic view. Look, quantum is awesome for many things, but the scalability of quantum is not there. To put in perspective, that chandelier there probably has, I don't know, 75 qubits. Obviously, it's not on because you can't do it here. It needs to be in a vacuum location. If you really stretch it, we may be, as an industry, 1,000 qubits. We have multiple technologies being tested for creating the qubits. That's a fight we don't participate, and that's why we are qubit agnostic. Now to do something very useful with quantum, you need at least 10,000 qubits.
We are far away from 10,000 qubits. However, our approach was, look, we can accelerate quantum by focusing on three things. Number one is the ecosystem. Number two is the network. Networking, again, comes to play, and I will explain why. Number three is the environment to develop quantum applications. On the ecosystem, you saw a number of vendors with us. Number two is the network. What if you use the same principle of traditional computing, applying scale-out architectures to quantum? Normally, if you want to get to 10,000 qubits, you have to scale up. Remember, HPE had, and still has today, Superdomes. Superdome is a scale-up system that today can scale to 64 TB of memory. You can put an entire database there. It was great at the time, and still is great for many applications for that.
To get 10,000 qubits on that, it's going to take time. What about if we take 1,000 qubits, scale it out? That requires networking. Now with Juniper and other assets we have inside the company, including silicon photonics and other things, we can go enable that. Number three, probably the most important one, which is how we develop quantum applications today, not when we have 10,000 qubits. That's where we use traditional computing environments in this scale-out model to allow them to develop applications for those who wants to build it. That's why the connectivity between the quantum system to traditional system like HPC or supercomputers are very important, so you can start simulating this thing. The question is. Will people develop scale-out applications in quantum or wait for a scale-up? That's the question, and that's a technology decision people have to make.
That's my view is the pragmatic real. The reality is that, look, I am old enough to have seen every inflection point in IT, mainframes to PC clients, to the internet, to mobile, cloud, to now AI. All the other stuff still alive. Many of you who works in banks here, you have a mainframe guaranteed. You still have applications that were developed 20 years ago. Quantum will be another one. I think quantum will be great for cryptography and other things, but ultimately, I think as a form of a accelerator to traditional computing.
We already connected a supercomputer to a quantum computer. What happens is that the supercomputer is doing all this work, and then, say, for this specific task, let me give it to the quantum. They do it faster for that thing and give the answer back so that the outcome of the whole thing goes out faster. That's how we think about it. Tim.
Yeah. Sorry. Tim Long at Barclays. Antonio, I wanted to go back to the core enterprise business. Obviously, servers, you went through some of the reasons why they're so strong. Just curious your take on why not just HPE, but the industry seems to be seeing less storage growth in the enterprise than compute server growth. Why do you think that is, and do you think at some point there will be a catch-up or any perspectives there? That would be great. Thank you.
I think there are two reasons for that, Tim. First of all, on the training side, you can't think of storage the way you normally think about storage, right? I think about a large amount of boxes with traditional compute and a lot of SSDs that the fast object gets laid on top of it as a software-defined layer. That's in the training side of the equation. We go to enterprise, look, you already have data all over the place, and what we did with the team is given an intelligent context to that data through the MCP aspect of this. I expect data to grow once the inferencing continues to grow. My view is that that data will grow as the inferencing grows, not because of what we're doing today. If that inferencing kicks off, then it will go faster.
The question is how it shows up. I don't think it's a traditional storage only. I think this concept of memory in KV cache, attached straight to the fabric. Let me put it this way, I don't think it will be a storage array only. Okay?
Okay.
Lou?
Quick question, Lou. We're going to wrap it up.
Okay.
I think it's pretty quick. Thank you. Lou Miscioscia at Daiwa. Obviously, you talked about the biggest complaint is really the supply chain. What about power and data center space in the U.S.? Will that be a constraint?
Power and da-
And da-
Data center space.
Yeah, both individually, in 2026 and into 2027.
Yeah. I think, Lou, it's fair to say we are behind. Look, I'm sure you do the analysis and all of you will probably have different numbers to begin with. That's a given. How many gigawatts have been announced? I don't know, 150? We think by 2030, there will be 250 GW of data center somewhat announced. I don't know how much of that 100% will be in production. The biggest limitation today is power and cooling. In the U.S., we don't have space problem. You may have a community that does not like to have a data center in the back of their house. Look, I think, as I reflect the back history, this is also an opportunity to innovate around the entire ecosystem. There will be new sources of power. I think it will be independent grids to power this data center.
I don't think the vast majority of the energy will be connected to the main grid. That also forces new innovation in gas turbines. You saw Siemens Energy has used an AI as a way to create next generation of gas turbines, obviously nuclear power and small reactors of sorts. We talk about putting AI data center in space. Look, funny enough, HP already has a very small AI data center in space. It's going around the Earth 256 mi above our head every day of the week. That's called Spaceborne Computer-2, and the astronauts are using that to do research on the International Space Station. This is small scale. Then here, probably by the end of the year, October, November, Artemis III is going to take the first rover on the Moon. That compute module and the network to connect back to Earth is Hewlett Packard Enterprise.
That's Astrolab. The first flip rover on the Moon will be powered by Hewlett Packard Enterprise. Because before you show up there as a human, you need to have some infrastructure, right? Astrolab is going to control that rover through that connection. Yeah, I expect, Lou, that This is why the U.S. has a huge advantage, put aside the politics. Getting into this faster, including regulations, compared to other geographies of the world, I don't think anybody can really match us.
We had a suggestion that the next investor event we have is going up to the International Space Station.
Yeah. Good luck with that. I want zero liability with that. I can take you to one of the space mission controls where actually we are going to power that as well. The mission control where the rover is going to be controlled is also Hewlett Packard Enterprise. It's a fun thing to do, but also you learn a lot.
Yeah. That's great. Well, thank you everyone for joining us here in the room and on the webcast, and thank you to Antonio.
Yeah. Thank you for everyone that logged into the webcast. I appreciate it.
Yeah.
Thank you for spending the time with us in the next couple of days. All right?
Great.
Thank you.