FinOps X 2026 | Mike Thompson & Jim Greene, AMD

Clips
More from FinOps X 2026

Mike Thompson

Director of Cloud Product, FinOps Foundation Governing Board

AMD

James Greene

Director of EPYC Technology

AMD

How data center modernization creates room for new AI workloads

As enterprise AI budgets hit their limits earlier each year, the pressure to fund new agentic and inference workloads without expanding total spend is forcing a fundamental rethink of infrastructure and data center modernization.The State of FinOps 2026 Report found that 98% of practitioners now manage AI spend, even as most organizations still overspend on AI workloads by four to five times their original budget. The core tension — more AI demand, constrained budgets and aging hardware — is exactly the problem Advanced Micro Devices Inc

play_circle_outline AMD EPYC Cuts OpEx ~30% via Consolidation; Rationalizing Agentic AI Spend in the Token-Maxing Era

play_circle_outline Replacing Idle Servers: Modernizing Data Centers for AI Workloads and Energy Savings

play_circle_outline Hybrid cloud and common architecture benefits for workload portability and sovereignty

play_circle_outline Optimizing TCO and Revenue with FinOps: Right-Sizing AI Instances and AMD End-to-End GPU-to-Rack Solutions

play_circle_outline Supply constraints (memory, components) push enterprises to cloud and force engineering optimizations

Info
Transcript

Mike Thompson & Jim Greene, AMD

Mike Thompson

Director of Cloud Product, FinOps Foundation Governing Board AMD

James Greene

Director of EPYC Technology AMD

In this interview from FinOps X 2026, Mike Thompson, director of cloud product at AMD and governing board member of the FinOps Foundation, joins Jim Greene, director of EPYC technology at AMD, to talk with theCUBE's John Furrier and theCUBE Research's Paul Nashawaty about how the end of the "token maxing" era is forcing enterprises to treat FinOps as a discipline for managing AI value, not just cutting cloud costs. Thompson frames the core problem: AI budgets are being exhausted one to two quarters into the year, leaving no runway for new agentic workloads. H... Read more

explore Keep Exploring

How are companies responding to rising AI compute costs and mid-year budget squeezes, and how can AMD's EPYC processors help free up OpEx to fund new AI initiatives? add

How can enterprises modernize their data center infrastructure to meet the capacity, power, and performance demands of AI workloads? add

How can organizations optimize value for AI and other use cases across on‑premises, cloud, and hybrid infrastructures while addressing data sovereignty, performance, cost, and availability requirements? add

How should organizations choose the right hardware (CPU vs GPU and what size of instance) for different AI workloads such as training, inference, and offline or small/medium models? add

During a period of constrained component supply, who will obtain the available memory and servers, and how are enterprises responding (e.g., moving to the cloud and right‑sizing memory)? add

bolt Powered by CUBE AI

Mike Thompson & Jim Greene, AMD

search

John Furrier

>> Welcome back to theCUBE's live stream coverage here at FinOps X in San Diego 2026. I'm John Furrier, host of theCUBE with Paul Nashawaty, with theCUBE Research, my cohost. Saving cost is a big thing, but the value creation is what everyone's talking about, because revenue's being generated, teams are being transformed, value's being created, but you got to do it with the cost process driving TCO. New ways to value projects, workloads, and of course, productivity. Got two great guests from AMD. Mike Thompson back, CUBE alumni, director of cloud products at AMD and FinOps Foundation governing board member. Great to see you. Thanks for coming back on.

Mike Thompson

>> Thanks. It's good to see you again.

John Furrier

>> Jim Greene, director of EPYC technology at AMD. Good to see you. Thanks for coming on.

>> My pleasure. Good to be here.

John Furrier

>> Well, first of all, congratulations on AMD on the business performance. Obviously everyone's talking about the results we're seeing at the AI infrastructure. So thank you very much for doing that. We want more horsepower, faster, cheaper, better. We keep doing that, but this is turbocharging the era we're living in, which is the agentic AI, tokens. We heard on the keynote. The token subsidy is over. I love that line, because we've been kind of letting people run with the AI, but now people are like, "Wow, I see the value. How do I scope it?" You guys have been contributing that with some architecture, some product features. But here, this group and this community is going to be setting the agenda for how to financially understand what the hell's going on. Mike, this is kind of what we talked about at Google Next, Cloud Next. What's the update? Have we solved the problem yet?

Mike Thompson

>> Well, similar discussion. I think with AI, token maxing, I think is what JR was referencing this morning in the keynote. The era of token maxing is kind of over, because the spends are going through the roof. And so there's a period of rationalization that I think the industry is going through now, particularly on AI applications and agentic, because those ones are really practical. So I think aiming to actually get value out of AI investments is one of the major focuses. But then one of the problems that they quickly stumble on, and this is something we were talking about before going live, was budgets tend to get spent one to two quarters into the year. There's so much dynamic development in the AI space that two to three quarters in, budgets are already tapped out. There's new AI applications that simply must happen. How do you fund that? And that's one of the things that AMD really likes to bring to the table with our EPYC processors, because you can free up easily ... I'll speak across clouds, maybe Jim can speak to a little bit of the on premise applications, but easily 30%, at three clicks of a button, 30% OpEx return, OpEx savings that can then be reinvested in these new novel innovative applications like agentic.

John Furrier

>> Jim, how about the enterprise? Because again, this is the number one focus, because the people doing AI are builders, operators, and investors. They're investors because they need to squeeze the dollars. They got to think like an investor.

>> Yes. And that's a big challenge for a lot of them, because it's not only the rate of change, it's the scale of the change, right? It's literally almost Pandora's box. So one of the things that we find on the enterprise side is saying, "How do I get the capacity to deal with all this new demand?" Because a lot of their infrastructure is still really kind of old. I mean, some of it's pre-COVID, really old, six, seven-year-old servers, because they kept extending the life of those servers. Now, there's no way you can run AI workloads on these. So they have to find, how do I get the, A, the budget, but even just physically, the space and the power to accommodate this crazy growth in demand? And that's really where the data center modernization theme comes in. So we can have customers coming in saying, "Hey, I can replace eight of those old Ice Lake servers," or some old prior generation Intel, "and I can replace it with one EPYC server, that's going to save me the power, that's going to save me the space." And oh yes, over time, because it's so much more energy-efficient, so much more license-efficient, you save the money as well.

Mike Thompson

>> Yeah. So that consolidation gives you more rack capacity, you draw less power, and that's how you make room for innovation.

>> For the new stuff.

Mike Thompson

>> I've got to soapbox about this. I've talked to dozens and dozens of companies. I look at their infrastructure, what they're running on their servers, maybe 70, 80% of the market is running at 10% utilization, meaning the CPU is only being used up to 10% of the time, 90% idling, just wasting power. To me, that's gross. And I think the leadership performance that EPYC brings, that high-performance processors bring, enables you to take those older generations, n-1, n-2, n-3, and consolidate many of those into one server, and actually boost performance and save money at the same time. And so I think there's an awful lot of opportunity out there, with the egregious waste with low-utilization servers that are out there now.

Paul Nashawaty

>> So John opened up with optimizing for value, right? And that's really what you're kind of alluding to, is you're really kind of driving towards that value, but there's the use cases that really come into play, right? You have AI for automation, you have this crawl, walk, run stage, where maturity, where organizations are at, but you also have things like the tier two and the neoclouds that are kind of starting to use this. But a lot has to do with AI and data sovereignty, in some issues, in some cases. So talk a little bit about that. How do you optimize for value when you're talking about these different use cases across on prem and in the cloud, hybrid infrastructures and such?

Mike Thompson

>> So to me, having the commonality of that infrastructure across an on prem versus a cloud-based infrastructure, and across multiple cloud-based infrastructures, really gives you the ability to address and pivot to almost any one of, "Oh, I have a sovereignty concern," or, "I have a performance concern. I have a cost concern. I have a availability" ... You really have a lot of tools in the toolbox if you have a consistent infrastructure across all of those, and it allows you to put the workloads where it's most cost-effective, or where it's legally required, or any of those other kind of drivers for the business discussion, having that capability is super important.

John Furrier

>> I love having you guys on, because you got the cloud and on prem. And one of the things we've been saying on theCUBE, you know our AI factory narrative, on the app dev side with cloud native, we've been talking about the same thing. Hybrid cloud one. So you have a distributed computing, true distributed computing environment. Okay, great. There's no debate on that. But now the data center has exploded, because of the diversity to bring in the kind of optimization with AMD and others. You can run stuff on prem and do the cloud. So the cloud's not going away.

Mike Thompson

>> As long as you're running on x86, that's true. On ARM, it is not.

John Furrier

>> Okay, explain-

Mike Thompson

>> There's no on prem ARM solution, realistically. So if you want to have a reasonable hybrid environment, you're not going to run x86 one place and ARM the other. That's just too difficult to manage. It doubles the resources.

John Furrier

>> It defeats-

Mike Thompson

>> Operationally defeats the purpose.

John Furrier

>> Yeah, explain the consequences and scope, the impact of that. Explain. This is a very important point.

Mike Thompson

>> Well, x86 was defined to be basically a compatibility standard, so that there can be multiple suppliers supplying these compute platforms. You can migrate between Intel and AMD and any other x86 supplier pretty much at the push of a button. When you bring an alternate architecture in there like ARM, and especially because a lot of the CPUs that are available now on ARM are highly proprietary to each one of the hyperscalers. While I love working with the hyperscalers and all, those are proprietary platforms where your applications get stuck on an island. You don't have any other choice to go somewhere else to flex your workloads and find the right place to land them. This is kind of one of Jim's talking points. You like to call it flexing your workloads.

>> Yes.

Mike Thompson

>> If you're on x86 on prem, you have a single pane of glass to manage it all, and literally spilling over into cloud or bursting to cloud is the push of a button. If you try to do that between ARM and x86, you got to recompile everything, have a whole different environment. So in my opinion, hybrid environments beg for a single architecture. I mean, there can be multiple architectures, but you want to be running the same architecture on premise and cloud.

John Furrier

>> All right, so explain the impact to the ops team, because that push of a button and the EPYC example are great examples of how you can optimize. If I have ARM, what's the cost to me? Time, time, people.

John Furrier

>> A lot of people now will say, "Oh, a recompile isn't that big a deal. AI can actually do it for me and simplify my life." Yes, but most IT practitioners will tell you it's not the port that takes the time, it's the maintenance, and supporting it, and keeping track. And there's no world where keeping track and building two to three to four instances of the same thing is more efficient than one.

John Furrier

>> It's a silo and more staff. So more staff. It's more staff-

Paul Nashawaty

>> issue.

Mike Thompson

>> Well, it's also a flexibility issue, but even outside of that, so I think kind of what you're saying, Jim, is the overhead and maintenance and management is roughly double what it would be if you had a single architecture. And then I would add, this is one of the things AMD and FinOps Foundation, what we try to bring to the FinOps Foundation is the notion of shifting left should involve wise platform choices. There's a 30 to 40% OpEx difference, annual OpEx difference between a couple of compute platforms that look the same. If you take an M8a in AWS, and you compare it to an M8i and then you consider the performance, the delta is 30 to 40%. A lot of folks don't consider that nowadays, and particularly when you're landing the applications, making those wise choices upfront is better. Otherwise, a year or two years later, the FinOps team is going to find this, see the egregious waste, and then force you through a more painful change, rather than making a smart choice upfront. That's how we shift left.

John Furrier

>> And then there's cost on licensing. There's all kinds of hidden costs.

>> Yes. The cost has a very, very long tail. Again, whether it's the administrative part, the operational part, if you look at things like on prem, the power bills, all of these things really ... And quite frankly, they are a much bigger number than the price of buying server A or server B, or instance A or instant B on a isolated basis.

Mike Thompson

>> And I think the tricky bit there is it's easy to walk into a train at the end of the tunnel, when you thought it was sunlight. If you look at a lot of the marketing that's done on ARM platforms, the claims look great, looks cheap, looks fast, looks energy-efficient, and on the surface that's actually true. But if you double click down and look at how many more servers or how much more runtime it takes to complete a job on an ARM architecture, that offsets everything. That's the part they're not telling you.

Paul Nashawaty

>> So I would like to up-level this a little bit and say, "Okay, there's a big impact to what you're talking about." It makes a lot of sense. The TCO, the flexibility, all those pieces matter. The other factor that take into consideration is, and this is what I'd like the audience to understand, is what does this mean for workloads? What does this mean for optimization? I know we talked about all the impacts underlying pieces, but if you're running an AI workload, and you have different architectures underneath, what does it all mean? Is it going to run better? Is it going to run different? Is it going to run indifferent? Like what does it mean?

>> So architecturally, they're going to run different, because they're running on different silicon. The key thing here is how do you operationalize that difference? So can you build the orchestration that's going to place the right workload in the right environment, right? You have an SLA, you've got some other type of requirement. How do you make sure that that environment is going to give you what you need for that workload? If you have a predictable, common hardware architecture underneath that, that's a little bit easier than to say, "Oh, well, it's supposed to be 30% faster with this other thing, but it isn't for my workload." So you kind of want to minimize some of the variables there. And the fact is, you don't have to choose one of those other things. You could get the same thing here, the same presence in not only that, not only did hybrid cloud win, multi-cloud one. So again, that just multiplies that.

Paul Nashawaty

>> But it's important for the audience to know, because a lot of times, maybe the DevOps teams or the platform teams, they're like, "I don't know, there's something under the covers, whatever. As long as it works, it works," but the impacts are real. And those data points you just mentioned, that's something that maybe people will oversee.

>> Yes.

Paul Nashawaty

>> So I think that's something to kind of highlight.

Paul Nashawaty

>> And let's face it, the whole premise of the cloud, no one wanted to care about the hardware. Ideally you don't have to. But if you do, there's tons of optimization benefits to paying attention to that, because that's where you get your scale, that's where you get your flexibility for deployment, that's where you get the flexibility to choose the most cost-effective platform. It's where you can more easily instrument automation, and that's where things like AI are really going to come in.

John Furrier

>> Okay. So I want to get into TCO, because that was on your guys' agenda, and I want to get into it. Because the old school TCO calculator was, you type in the purchases and you look at over a span of time, over quarters. You mentioned they're blowing their budgets. How should TCO be thought through? Because things have changed a little bit. You got real time, you got growth, you got architectural growth on the AI side. So you have now hybrid, architectural choices have impact. How should people think about TCO?

Mike Thompson

>> I think a lot of it really boils down to right-sizing, or making the wise decisions about where to land your applications. To start with, not every AI solution wears a leather jacket. There's plenty of applications where CPUs, which are way cheaper, way more ubiquitously available, do just fine for small and medium models, and offline models especially.

John Furrier

>> Agentic.

Mike Thompson

>> Yeah.

John Furrier

>> Inference, too. And for inference, many use cases.

Mike Thompson

>> But also, we went through a token maxing era, where you just tried to get the biggest GPU to do the AI that you had to do, because you had to do AI. And now it's a question of, it's not even that difficult of a decision, in my opinion. You look at what are you trying to accomplish, and then put it in a right-sized instance. And for AMD, we're not a one-trick pony. We have products that cover everything end to end for AI solutions. So we can start from one GPU that slips into a PCIe slot in a server, that's under Jim's team now, up to the eight by GPUs, up to the rack scale. You got to think about landing it in the right performance level for a GPU or a CPU, or you're wasting tons of money.

John Furrier

>> And now you have developers and builders that aren't like coders. That's evolved and that's on the scene now. So you have a lot more action happening on premises, but then you got this push button burst to the cloud option. So it's not a one-environment pony to use a word.

Mike Thompson

>> Yeah. Not a one-trick pony.

John Furrier

>> I mean, it's multifaceted. Not a one-trick pony, you're distributing it. Talk about that dynamic and what that means, because people are going to enable teams to start coding.

>> Oh yeah.

John Furrier

>> "Here's your tokens on prem, go play with it." And then, okay, we got the scope, like a sandbox and then I got to go to the cloud to scale it.

>> Everyone's a developer now, and not only that, everyone's developing something that could become a developer and release a new ... I mean, you see this, just almost kind of the bowling pin effect of, you build an agent, the agent builds sub-agents, and it really does scale really, really quickly. Again, having a common infrastructure across wherever, again, "Hey, I'm out of capacity here. Can I burst into some other part of the infrastructure?" Having that be totally compatible when you need to bring it back down because you have additional capacity is a huge benefit. It runs natively where you-

John Furrier

>> Guys, talk about this development and innovation. Let's talk about the innovation side of the equation, because we're seeing constrained environments.

>> Yes.

John Furrier

>> In my lifetime, my observation is anytime something's been constrained, entrepreneurs or engineers will engineer around it. So there's a balance between managing supply chain, like memory. If memory's going through the roof, and storage, and now you have networking is kind of a bottleneck. So all these things, it's the classic hyper-converged variables. What is the innovation? Because you guys are in that supply chain game, you got memory, we see the prices, that's to be innovative. Whoa, forget right-sizing, it's called stay in business. So there's a whole nother level of right-sizing around the constraints. What are you seeing? What are people doing to optimize the engineering side of it? Because a lot of people are penciling out the architecture now to make those calls.

Mike Thompson

>> Can I lead in with this, and then you can say something about the testing we did-

>> Yep, yes....

Mike Thompson

>> to look at memory downsizing? All right. The first thing I see is the hyperscalers have some of the most buying power on the planet. So during a period of constrained component supply, who's going to get them? So I see a lot of enterprises who otherwise would have had loads of buying power unable to source either the memory or the servers at all, or at the price point that they can manage. So I see a lot of them going to the cloud to take advantage of the sourcing there. And then there's another right-sizing thing, right-sizing for memory and you guys did a bunch of testing to look at downsizing memory.

John Furrier

>> One of the things that we're finding is you got to kind of break muscle memory. People bought what they always bought from an IT perspective, because that's what they do, right? They just keep ordering it. We're really trying to drive more awareness in the market about saying, "You know, A, there's probably a vast amount of overprovisioning," because at times memory's cheap and cheerful and you just put-

John Furrier

>> it all....

>> the same model in there-

John Furrier

>> Take it all in....

>> put it all everywhere. That way I could put any workload anywhere. But the real thing is, there's a lot of applications that aren't incredibly memory-sensitive. So we're trying to drive awareness about right-sizing memory, even on on prem configs, right? Saying, "Hey, you always bought two terabytes, you're only using a small fraction of that. Let's look at what you really need." So we're putting tools into the market today to allow people to do some level of analysis to say, "Gosh, I am overprovisioning. I can now make some better sense." It's really helping them understand what they really need versus what they've always bought. And again, no one's going to run into their data center and rip DIMs out, right? But next batch of purchases you make for that class of workload say, "Hey, you know what? I can buy -"

John Furrier

>> So mindset and muscle memory, basically don't think the old way.

>> Exactly.

John Furrier

>> Think the new way.

John Furrier

>> And this environment is just totally challenging everybody on that.

John Furrier

>> And what's the new way of thinking? How would you describe the new mindset, if you had to break-

>> To me-...

John Furrier

>> everyone with one swing of the bat?

John Furrier

>> To me it's how do we not only forget our muscle memory, but how do we reacclimate to the world of 10x, right? Because our needs are changing not incrementally now, it's going exponentially. So that's going to be the challenge for people to overcome.

John Furrier

>> And I think this community's going to be instrumenting it all. Guys, great to see you here. Great to see you come back on. I love this conversation. We'll see you guys at the Accelerating AI event you guys are having in San Francisco in July. theCUBE will be there.

Mike Thompson

>> Advancing AI. I'll be there.

John Furrier

>> Advancing AI. Sorry.

Mike Thompson

>> Just up the road, yep.

John Furrier

>> Advancing AI. I got it wrong there. And we need more advancement. Right now, more than ever, the costs are front and center. CFOs are leaning in, very operational, not just, "Here's your IT budget." They're like, "Give me the data." Thanks for coming on. Appreciate it.

Mike Thompson

>> Thanks for having us.

>> get run over.

John Furrier

>> AMD, having the products, having the migration, the portability, workloads running in an environment that's cost-effective, still drives it and advances the AI. I'm John Furrier with Paul Nashawaty. Thanks for watching.