The Deep View: Conversations

AI's appetite for compute keeps growing, but so does the pressure to deliver more intelligence per watt and per dollar. Can AMD's first rack-scale AI system open up an ecosystem dominated by Nvidia?

In this episode of The Deep View Conversations, we sit down with Andrew Dieckmann, AMD's general manager of its data center GPU business, to unpack the company's Helios platform and the rapidly changing economics of AI infrastructure. 

Dieckmann explains why frontier AI requires more than just GPUs. It demands tightly engineered racks that combine GPUs, CPUs, networking, software, cooling and serviceability. The conversation examines the tension around AI data centers: hyperscalers still cannot get enough compute, while communities worry about power, water and whether the benefits justify the buildout. Andrew argues that responsible deployment and open ecosystems are essential as these systems become intelligence factories.

The conversation then turns to Helios: AMD's performance claims against Nvidia Vera Rubin, pricing and value, the first likely customers, and the Cerebras partnership for high-throughput, low-latency inference. Andrew closes with his advice for leaders navigating AI velocity: reassess priorities more often and use coding agents as force multipliers for scarce engineering talent.

Topics covered:

• Why AMD is moving from chips to full rack-scale systems
• AI demand, data center constraints, and community impact
• Open hardware, open software and customer choice
• How agentic AI changed infrastructure planning
• Helios performance, efficiency, pricing and customers
• AMD Helios versus Nvidia Vera Rubin
• How AMD and Cerebras split inference workloads

This conversation offers a clear look at the technology and economics shaping the infrastructure that will power everyday AI and the breakthroughs to come.

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm 

And don't forget to sign up for The Deep View daily newsletter. We don’t just cover AI, we decode it. In a world flooded with hype, we deliver sharp, no-nonsense insights to keep you ahead of the curve and help you put AI to work every day: subscribe.thedeepview.com

Creators and Guests

Host
Jason Hiner
Editor-in-Chief of The Deep View

What is The Deep View: Conversations?

From frontier labs and enterprise platforms to emerging startups reshaping entire industries, The Deep View: Conversations podcast interviews the brightest minds and the most influential leaders in AI.

Jason Hiner: All right, Andrew, thank you for being on the Deep View Conversations. For those who aren't familiar, tell us a little bit about what you do here at AMD.

Andrew Dieckmann: So I'm the general manager of the data center GPU group at AMD, so that involves our Instinct products, including the Helios Rack that we launched into production today. My job is to manage the overall business, customer relationships, product management teams that just manage the overall P&L and business moving forward.

Jason Hiner: We're going to talk a lot about Helios for sure. But before we, you talked to me a little bit about, you know, what's the stuff that gets you most excited to come into work every day right now?

Andrew Dieckmann: Oh gosh. I mean, I started out my career as an engineer, so I mean, I get excited about the technology, the pace of change is very exciting, also challenging. So I think just what gets me excited is every day, every week you learn so much. There's always new challenges to solve. The breadth of what we're solving for, you know, it's not just silicon. I mean, silicon is the foundation of what we do, but, you know, doing data center planning and everything from picks and shovels to building the chips and the systems and everything in between and the right go-to-market motions and customer relationships. I think it's a lot of diversity and interesting problems to solve every week.

Jason Hiner: Yeah. Very good. You know, you worked in your career across a lot of different areas of infrastructure. How has that prepared you to work in the work that you're doing now, especially with Helios? But even just more broadly, the way that AI, you know, has changed, has meant sort of the demand or maybe the shape of the demand and of the products that AMD is building, you know, look a lot different, right?

Andrew Dieckmann: Yeah. I think, so I joined AMD in this role about four and a half years ago. And so it's been, fortunately, I had a lot of background to draw on from working in semiconductor for data centers, infrastructure products for many years. And so I was very familiar with data centers and the ecosystem around data centers. And I think that, you know, luckily has prepared, you know, me well for the challenges that we're embracing now. And I think AI has, like I said earlier, expanded the scope of what we think about when we bring a product to market because it's not just, hey, we're releasing a chip, you know, have fun designing it into a system or here's a basic reference design. It's moved to, you know, what we announced or disclosed more today, the Helios system. It's a full rack scale architecture. And although AMD is not selling that platform directly to the customers, we're enabling an ecosystem to build that rack scale architecture. It requires AMD to be very prescriptive in terms of the amount of design work that we do, the operational aspect of plumbing all the supply chain to make sure that all of these piece parts exist in the appropriate quantities for our partners to be able to obtain, make sure that the rack is, you know, highly reliable, serviceable and all of those things. So it's really expanded the scope of what we think about, you know, yes, we're a chip company, but we're much more than a chip company. There's an enormous amount of software and system engineering that goes into delivering these solutions.

Jason Hiner: You know, compute is the, you know, the fuel that drives the AI industry forward. And there's a lot of mixed messages around compute these days, you know, it's that there's so much that's being spent. There's enterprises that are worried about their costs. There's you know, hyperscalers that are and AI model providers that are concerned they don't have enough like that the demand is, you know, going through the roof and they're not going to have enough. And at some point they're going to, you know, run out of tracks, right? The train's going to run out of tracks. Those were a lot of different signals. Like how do you all at AMD sort out all of those different signals to understand the industry right now?

Andrew Dieckmann: I mean, it's been a tremendous, there's been a lot of change in a short period of time and there will continue to be a lot of change. The thing that has proven to be true is every time we think, oh, we might have solved for enough compute, there's like, you know, we can, we've incrementally every six months or 12 months, sometimes on a monthly basis had to raise our view of the market opportunity because the signals that we're getting from our customers is they see really real value in the technology. It's providing real benefit in terms of business outcomes and therefore they are investing in it. And it is, I think the scale of that investment and how quickly that investment has come upon us that if people debate, hey, is it too fast? Is it too much? But it's one of the probably the largest infrastructure project on planet earth that we've had to date. Right? And I think there's, these are very well educated decisions that are being made by well educated stakeholders that are spending enormous amounts of money to build up this infrastructure based upon the promise of the technology and really the economic value it provides.

Jason Hiner: Yeah. Okay. So there's this story, you know, even the wider public sees data centers being built, right? And so there's this very challenging narrative around data centers right now. When you all think about that, how do you tell the stories about like, oh, well, the benefits that this is going to bring are X. So people don't just think about, well, why are these big buildings being built? And what is that going to have impacts for power or water or other things? You know, do you all think about that? Do you think about how can we help tell the story of the benefits of these things?

Andrew Dieckmann: Yeah. I mean, I think there's, we're building intelligence factories, right? Which have a lot of promise in terms of what they can provide in terms of value to society, right? And I think it's, you know, society has to figure out, you know, we have to work as really it's a humankind thing to make sure that the value of these systems accrue in a beneficial way. And like any powerful tool, it can be misused. It can be put to very beneficial use, right? And we see a lot of the beneficial uses in medicine and education. And I mean, the list goes on. It's like, of course, you used to, you know, a very small portion of society could afford a personalized tutor for their child or whatever. Now you get a personalized, everyone that has $20 in their pocket can get a personalized tutor to really accelerate, you know, the educational experience for their child, for their children. That's just one example. But yeah, I think it's, we have to build data centers in a responsible way in a way that sensitive to the communities that they're in. We have to make sure that we are, you know, using all of the technologies at our disposal to make sure that they're done in a way that is ultimately accretive and beneficial to the communities that they're in.

Jason Hiner: You all talk about, when you talk about Helios, when you talk about a lot of your technologies, you talk about the benefits of a more open ecosystem. Do you see that as one of the parts of the good that you all are bringing to both the sort of the dialogue, but broader sort of good to the ecosystem itself? Like there's less lock in. There was more choice. There's more opportunity for your customers and partners to sort of shape what the product looks like and how it gets deployed. Is that one of the things when you think about, and how do you think about that as one of the sort of important aspects of what you're building?

Andrew Dieckmann: I would say it's a core principle at AMD to have that open ecosystem. I mean, we saw, we announced a wide breadth of products that we're doing everything from client to edge to data center with Helios. AMD has a broad range of a broad portfolio for many different applications, but we're not about locking customers in. We feel like there is, customers should have choice and there's not a one-size-fits-all. As Lisa constantly says, there's not one compute solution for everything. There's optimizations and it's such a broad market and with a variety of applications that no one company is going to provide all of those things. Now we think AMD can provide many of them and we're happy to provide very competitive solutions in a number of areas. But whether it's specific platforms like Helios, we have an open rack architecture, which allows customers to have choice on different components. Probably more importantly, we have an open source software stack.

Jason Hiner: Yeah.

Andrew Dieckmann: And that has been, I would say, I don't want to overuse the word game changing, but I mean, with the advent of agentic AI and the ability for that to supercharge the software development, the fact that our key partners like you heard on stage, like whether it's OpenAI or Anthropic or Meta, they contribute, you know, them and other, you know, very valuable partners contribute to that open software stack and it's really supercharged our ability to provide more value to our broad customer base through that open source endeavor.

Jason Hiner: Yeah. So you mentioned agentic AI, you know, this 2026 trajectory changed a lot in January in the first quarter in January, February, when sort of the agentic moment happened with OpenClaw. OpenClaw was like the very, very tip of the iceberg. Like there's been so many things that have happened, you know, since then. How has that had caused you all to recalibrate even what you do, the way you think of demand and sort of the way you build things? Because you plan years ahead of time to build the roadmaps that you build.

Andrew Dieckmann: Yes.

Jason Hiner: So how does that work, the change in the world in the past six months?

Andrew Dieckmann: Yeah. Our planning cycle, I mean, the development cycle of a chip is like close to two years start to finish, right? And so I think as we learn a great deal during every planning cycle, through every development cycle really, and we try to retain optionality and modularity wherever possible. They constantly reassess and, you know, sometimes, you know, what's not publicly reported is we have plan A, plan B, plan C, we're constantly looking at in adjusting what's the right ratio of this to that, you know, what's the right solution in this particular system area or, you know, our chiplet architecture for our, our silicon products has proven to be extremely valuable and that we've been able to incrementally spin off, you know, either product variants or shift a little bit what we're doing partway through the development cycle. And we constantly have a feedback loop with our customers as they're learning too. Like we're all learning like, sure, we, we sit down and say, okay, what did you learn this week? What did I learn this week? And we collaborate really closely. There's a lot of co-design. We get feature requests from our customers well into the development cycle. We do not ignore those. We assess how do we, you know, it drives some of our engineers crazy sometimes because we're coming with, okay, I need this one more thing. But I think due to our silicon architecture being modular, our software stack being open, it's given us a lot of flexibility in terms of incorporating that new learning into our product deliverables.

Jason Hiner: So the world with agents, the world has kind of moved in AMD's direction a bit, right? Is that fair to say?

Andrew Dieckmann: We feel that way.

Jason Hiner: Okay. Because before that, it was very GPU AI accelerator focused. And with agents, it's much more orchestration focused. The CPU is needed, you know, where you have a leader, much more sort of a leadership sort of position in server CPU. And so what is, what has that meant for the roadmap as well as sort of the way you're rolling out this technology, the stuff that was announced even today?

Andrew Dieckmann: Yeah, in a few different areas, I mean, one, it's always nice to be able to provide more value to your customers, right? And so, you know, now our major customers, they need oodles of GPU compute, they also need oodles of CPU compute. To your point, we have a very established leadership position in CPUs, which we've been working on for many years. So I think it's allowed us to bring more ingredients to the table that our customers really care about. And so that's been nice. And then also, even within like the GPU systems, you know, the benefit of our high performance CPUs actually comes forward in some of the total system performance. Okay. So it's both, you know, the silicon and architecture on the GPU side, but also it's the whole system structure. Because again, our customers are not buying components, they're buying a system. And the system performance, you know, both per watt, you know, per per dollar competitiveness is really the metric of is it good, you know, how does it compare to our competition, all that stuff?

Jason Hiner: There's this incredible ROI story that's also happened because of this, because of agents doing so much, you know, costs for enterprises have soared, you know, and they're feeling the pain of that. How much pressure does that put on you all? Do you get from customers of like, I need this to be more efficient? I need it to do more, I need the compute, but it's got to be more efficient. The my cost, my, you know, cost per watt intelligence per watt needs to be better, all of these kinds of things. How much does that pressure come to you all?

Andrew Dieckmann: I mean, I think it's, it's one of the foundational design goals for every generation, like we need to improve performance, performance per watt, performance per dollar, we need to, you know, pass a good amount of incremental improvement to our customers. They need to pass that on, whether it's cloud customers, they need to be able to, every generation do so much more for less because the demand is like exponential. Yeah, I don't, I mean, it's, that's a core principle. Does it make it us try harder? I don't know that it makes us try hard. We were, like,

Jason Hiner: Everything was, like, you're always working that to the max.

Andrew Dieckmann: Exactly, like, all those knobs are dialed up and we debate that vigorously. And I think that's like just core operating behavior, at AMD, but it is the importance of it, I think, just because the, the, the spend is growing. And it's just, it's a large, a larger problem, right? Because, I mean, we're deploying, the world is deploying more compute, you know, per year than in previous decades, right? So, and that's accelerating. So, I think just the scale of the problem is, is to your point, I mean, the scale of the problem is, or the, the deployments is daunting. And so, every percentage point of efficiency is extremely valuable, extremely valuable.

Jason Hiner: Okay. We heard a lot this year, from the beginning, starting the beginning of the year around, you know, competitive products, Nvidia's, you know, Vera Rubin, as like, this is the extreme example of, you know, what's going to build the next models, what's going to solve the biggest problems and that Helios is, is sort of the first thing that's the first competitive product, my understanding, that's on that, you know, same level. When you think about what's unique about Helios, when you talk to customers, when you sort of tell the story of Helios, you know, how do you tell the story?

Andrew Dieckmann: I think the story of Helios is, it is extremely competitive rack scale architecture. We've, we think we have produced in the spec show, it's the highest performing GPU, best performance per watt, best performance, just straight out absolute performance. Those integrated into the rack, we believe, provides the best single rack AI performance in the market and we showed a bunch of specifications today and we think we're, you know, 10 to 15 percent better workload performance than a, than a Vera Rubin rack. You know, we will continue to evaluate their products as they, as they come to market, as I'm sure they will evaluate ours. But yeah, I think there's, it's a huge market. There is a opportunity for many companies to be engaged.

Jason Hiner: It's not a zero-sum.

Andrew Dieckmann: It's not a zero-sum thing. And I think, you know, we're doing our best to provide strong value to our customers through the, those deep co-engineering partnerships. And I think it's a healthy market to have multiple companies engaged and competing hard.

Jason Hiner: So, you all haven't said how much Helios costs, but there are some industry reports about how much it costs and that it is more expensive than, than Vera Rubin. Do you have any metrics in terms of the cost kind of, you know, per performance per watt? Do you, do you feel like it's competitive, that it's better when you get to the, down to the like, what you're paying sort of per, you know, unit cost?

Andrew Dieckmann: Yeah. We, so Helios is not quote more expensive than a, than a Vera Rubin rack.

Jason Hiner: Okay.

Andrew Dieckmann: I think we, we will, we engage our customers in a collaborative way. And I think all rational customers are willing to pay for the value that the product provides.

Jason Hiner: Sure.

Andrew Dieckmann: So, you know, if their assessment is similar to ours, which we think it is that it's the highest performing rack in the world, they are willing to pay a reasonable amount of money for that rack. And, and we are also cognizant of the fact that we are the less established competitor in this particular market. So we're being aggressive in the market, but also very rational and our customers are very rational. They're willing to pay for the value that we're providing. So.

Jason Hiner: Okay.

Andrew Dieckmann: We think we're, we're going to be well positioned to gain share and do so in a way that is beneficial to our customers as well as to our shareholders.

Jason Hiner: When I think of Helios, you know, you think about the first places that's going to be deployed, you've already sold it to your first customers. I'm sure you've already sold it to like your second wave of customers, you know, to before those are even made because there's so much compute demand. But those first customers, when I think of like, who is going to use these sort of systems, the level of the Helios or a Vera Rubin, I think of it and you tell me if I'm wrong, that it's the, it's the hyperscalers, maybe the Neo, the larger Neo scalers, the, you know, cutting edge frontier labs, that they're the ones who need, you know, that level of performance. Not everybody needs the performance of Helios. And you have a multiple levels of products that can serve multiple levels of customers.

Andrew Dieckmann: Yeah.

Jason Hiner: Who are the first ones that are going to be deploying it that have already bought it?

Andrew Dieckmann: Yeah. So, some large committed customers, I mean, we have five very large committed customers that we've talked about deployments with. The latest ones, Anthropic. So, we've discussed that the press release with them was, you know, up to two gigawatts of Helios deployments. They will obviously be a large customer. OpenAI was the first large customer we announced last year. Meta, we announced this spring as a multi-generational, multi-gigawatt customer. Those three customers will deploy a lot as foundational model builders. Microsoft, we had an announcement on Monday where they're going to be deploying a Helios at scale. Oracle, we have a partnership through Oracle Cloud that they're deploying a bunch of capacity. And then we have partnerships with a number of Neocloud partners who are also deploying Helios at scale. And I think, you know, to your question of who needs the rack scale architecture. Yeah. I mean, anybody who's working at the frontier, you know, basically, you know, is a primary candidate for the system. Anyone who's doing any large-scale pre-training. But then, like these guys that are serving, you know, frontier models, large models to, you know, hundreds of millions or billions of users, it is the most efficient way of providing that scale of inferencing also.

Jason Hiner: Okay.

Andrew Dieckmann: So I think those are kind of the primary customers. But there's like, you know, we have sovereign initiatives that we're working on where, you know, both in Europe as well as Asia, countries looking to deploy. So I think it's not a specialized piece of equipment. It's for people who are willing, who are looking to do training or inference at large scale. It's a great solution. And like to your point, we have other solutions for on-prem where people maybe don't have liquid-cooled data centers or we have air-cooled solutions, everything from PCIe cards to eight-way servers to, you know, inferencing on the edge and laptops. So there's a spectrum of solutions. But, you know, the Helios solution is really targeted to the large cloud deployments for at-scale training and inferencing.

Jason Hiner: Okay. You also announced a partnership with Cerebras.

Andrew Dieckmann: Yes.

Jason Hiner: Tell me a little bit about that. What's the, what's that partnership focused on?

Andrew Dieckmann: Yeah. So that is an emerging segment of inference. So inferencing is such a large market, there's segmentation now that's happening. And the low latency serving is something that Cerebras is best in class at. And the combination of their wafer scale engine with Helios allows us to scale that offering to a larger set of customers. So it dramatically improves the economics of their offering on a standalone basis. So.

Jason Hiner: Okay.

Andrew Dieckmann: In fact, they can disaggregate the inferencing workload. You have a prefill and a decode. You sort of put, you use the right tool for the right job. Like we have incredible throughput, they have very, very strong low latency responses. You kind of combine those things and it allows the service to provide low latency attributes at a lower cost per token than just scaling out the wafer scale engines independently. So that's what the partnership is all about.

Jason Hiner: Okay. Last two questions I ask everybody, you know, that comes on the podcast. One, in the age of AI, leaders use this term leverage a lot. Like how do I get maximum leverage for my time? And so what's your best tip for leaders who are really trying to maximize their time in the age of AI?

Andrew Dieckmann: I think the amount of change in the market, so quickly we discussed that a little bit. It causes you to need to reassess. It's not new things, but you've got to ruthlessly prioritize and you have to do that more frequently now, I find, because it's just like every month is like a year went by in terms of like, where am I spending my time?

Jason Hiner: Sure.

Andrew Dieckmann: And how to have things organized appropriately because there's new tools coming to market, coming to bear all the time that you can use to make your time more effective. So I think it's really, you know, evaluating those. I mean, as human beings, we have a throughput limitation of our own, right? But you have to take the time to actually pause and be like, and assess and then recalibrate. And I find that for myself, you have to do that. I have to do that more frequently now than in the pre-AI era. Just because of velocity.

Jason Hiner: Yeah. Okay. How about what's the AI tool that you're using right now that maybe not everybody knows about that you think people should be aware of because it could really change the game?

Andrew Dieckmann: I mean, as a company, we talked about today, the coding tools from our partners have been game changing in terms of the effectiveness of our engineers. And it has, as Vamsi talked about in the keynote, it's unlocked things that were just not possible before.

Jason Hiner: Okay.

Andrew Dieckmann: Like you could never, it doesn't matter how big your company is, I mean, we have 30,000 engineers, we're not a small company, but you could never hire enough of certain skill sets to actually like get done what you wanted to get done.

Jason Hiner: Sure.

Andrew Dieckmann: And so now I think the combination like, you know, kernel optimization is a perfect example. Like the, the agentic force multiplier on something like that has been a game changer for us. We've been able to do, just do things that weren't possible before have more optimization points for more different models. And it's really been an unlock in terms of ability to get, you know, our leadership hardware into the hands of more people doing like leadership things. So I think that has been, and that's been an unlock really in the last six months that took another step function forward. So I think that has been very exciting to see and it's been extremely beneficial to our customers and to our business.

Jason Hiner: That new box that you announced today, which is essentially like a team of agents on a desk, right?

Andrew Dieckmann: Exactly.

Jason Hiner: Is your team using those already?

Andrew Dieckmann: We're using those already as part of our workflow. And yeah, I think we're experimenting with all types of different, you know, tools and pieces of hardware. And the story is the same in our company. There's never enough compute. It's like you give the engineers more compute, they want more compute. So it's a good problem to have, but you know, good problems are still problems and we've got to solve them.

Jason Hiner: Yeah, very good. Andrew, thank you for the time.

Andrew Dieckmann: Yeah, thank you.

Jason Hiner: Great stuff.

Andrew Dieckmann: Yeah, great talking to you, Jason.