How to Escape AI’s 'Open Bar' Era: Intelligent Token Routing

0:00Mike Trojecki: AMD's recent agreement with Core Scientific to buy more than 500 megawatts of AI-ready data center capacity is a signal of where the AI race is heading, beyond just chips and toward the ability to assemble compute, software, power, and facilities into capacity customers can actually use. Now that capacity is expected to help AMD accelerate AI deployments, but enterprise leaders still face challenges in how to best take advantage of the new infrastructure, namely how to route each workload to the right model, processor, and environment, and ultimately proving the return before consuming the next token.

0:35Mike Trojecki: In this episode, AMD's John Hampton and WWT's Mike Trajiecki join us to explain how organizations can build that kind of token economy, from infrastructure modernization and intelligent routing to the operational decisions that make AI sustainable at scale. John Hampton is AMD's Corporate Vice President of Global Enterprise Technical Sales and brings lessons from AMD's own routing efforts that yielded incredible results, of which you'll hear more about. And Mike is a veteran of the show as he leads WWT's AI practice with a focus on practical, scalable adoption and measurable business outcomes.

1:11Mike Trojecki: This is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

1:25Mike Trojecki: John, Mike, welcome to the AI Proving Ground podcast. How are the both of you today?

1:31John Hampton: Outstanding. Great to be here on this podcast. Big fan of WWT and this AI Proving Ground podcast. And to be here with Mike Trejecki. It's going to be a fun one today.

1:40Mike Trojecki: Yeah, I'm looking forward to it. Thanks for having me again, Brian. Absolutely. Well, organizations, you know, for the last few years have really been asking themselves, how can we move fast with AI? And certainly there's plenty of organizations out there that are still at that moment. But the conversation as a whole seems to be shifting a little bit to how can we afford AI? How can we fund AI? And those types of questions. So, John, I mean, just curious, start us out. What changed between those two phases that's caused this shift?

2:09John Hampton: Yeah, we are definitely seeing a shift in the market and what we're calling stage two of AI. This is actually the timing of this conversation is terrific because it comes off the heels of AMD's flagship event last week, AMD Advancing AI 2026. And this was the conversation. And, you know, Mike, my view is this is the number one topic right now in enterprise, because what's happening is we've moved beyond the experimentation for forever. it was, hey, we need everybody using AI, go in there and educate yourself on all the ways that you've been doing AI and improve your productivity. And so that's what's been happening.

2:54John Hampton: And it's great that we're seeing all the adoption increase, but now we're seeing this shift to where we're going beyond AI prompting and we're seeing self-prompting, autonomous workflows, and even going beyond just agents and agentic workflows to loops and 24-7, we are cranking on AI now. And now what's happened is just an absolute explosion in the token bills. It is becoming really unaffordable. Suddenly the ROI equation is broken. And so enterprise is trying to now solve for tokenomics. I believe that that is the conversation that we all should be having and And that's what I'd love to break down with Mike here on the podcast today.

3:41John Hampton: Mike, there's actually this hilarious analogy that we've been talking about where Enterprise accidentally created an open bar for tokens. And it was, hey, everybody, come on in here, grab a drink, top shelf for everybody. Let's go another round. And now all of a sudden, you check that bill, it looks like we just had a wedding reception. It's been kind of this explosion. And suddenly, we're struggling. We have to shut that bar down.

4:12Mike Trojecki: Yeah, well, we don't want to shut the bar down. I mean, I'm okay with Pappy for everybody. You want to kind of go down that route. I'm not going to be averse to a Woodford here and again. But it's definitely changed the dynamic. I mean, we're seeing that internally, even at WWT from a token standpoint. You know, we can look at the FinOps world and use that as kind of a guideline for what we're doing with kind of FinOps and AI. But people are realizing that this is not inexpensive. You know, just using a chatbot, you know, is much different than having 200 agents that are out there running around.

4:53Mike Trojecki: And like you said, on this kind of agentic loop, it's just churning out 24 hours a day. As far as even the coding side of things, are we seeing the return of productivity? In some cases, depending on the workflow, absolutely we are. But in a lot of cases, we have to really take a look at the financial aspects of this in addition to not just the regulatory, but the operational pieces. Hey, John, do you think that that's kind of a normal, natural course of action where those costs explode just as it relates to the cost of learning, the cost of innovation? And if so, what does a more responsible model look like moving forward?

5:34John Hampton: Yeah, we really need to evaluate what we're calling the full token economy. And I've been saying this, I believe the next generation of winners in AI will have the best token economy. So we can break down a lot of the different aspects of what it takes to set that up. We think about the innovation that we're bringing and both the financial impact and the technology impact that it brings. But really, if you take a step back, Brian, we really see such a diverse set of AI-based use cases. And, you know, Mike, I'd love to get your thoughts on this here, too. But it's important that we don't bring a sledgehammer for every job, right?

6:17John Hampton: We should be thinking about this having a full AI toolbox, this stage two, this new era of AI to get the right tools for the right use case. So we're really excited to start thinking about intelligent token routing and dynamically evaluating each of these use cases and routing them to the most optimal model. So to really answer your question, we have to think about, are we running this on-prem or on the frontier models in the cloud? What models are we using? What hardware are we routing to? Again, don't need a sledgehammer. We don't need big GPU clusters for everything. And again, when you have started running these AI use cases 24-7 and you're using a sledgehammer for a nail, it's really going to hurt your token economy.

7:10John Hampton: me. So Mike, talk to me about this. Are you, are you now seeing this kind of emerging intelligent token routing? We'd love to get your view on that.

7:18Mike Trojecki: Yeah. You, John, you look at it this way, right. You talk about, you know, nail on a sledgehammer and you talk about use cases. I think it goes even a step further into the actual workload itself. Right. We've, we've had this conversation in the past about hybrid now being a permanent state, right. It's not a transition anymore. So you look at the different paths that a workload can take. It could be in an on-prem environment and an on-prem AI factory could be in a co-location provider. It could be your equipment in a colo. It could be in a neoclap, could be in a hyperscaler. The reality is that I believe, and I think you and I both agree, that this is going to sit, these workloads are going to sit across the board.

8:01Mike Trojecki: And we're not going to build a giant AI factory on-prem for every single workload. Some workloads are going to do fine just in a hyperscaler. Some of them, if I'm doing training runs, maybe it's going to be in a neocloud. To me, that's where we get into this question of tokenomics is where is the best place for that workload to sit? And there's going to be hundreds of workloads potentially per use case. So we look at this and to me, that's a major contributor into the tokenomics kind of formula.

8:36John Hampton: Yeah, definitely. I like what you said there. I think everything for sure has to be hybrid. And you need to make sure that you empower your infrastructure with not only hybrid options, but an array of different models and hardware solutions so you can really tune those workloads and AI use cases to improve, number one, again, the financial impact in the tokenomics, but number two, again, improve your token efficiency, really improve the overall effectiveness and performance of what you're doing. Choosing the right solution there can also help on that piece. So we're definitely encouraging everyone to really hold AI use case workshops.

9:21John Hampton: This is the beauty of WWT. I'm really impressed, Mike, with your AI proving ground in your advanced technology center, where you can really discuss, run labs, run POCs and do some testing on your various AI use cases and leverage WWT and AMD to really do a lot of the heavy lifting here and help guide you through what the right solutions are for your environment.

9:48Mike Trojecki: Yeah, it gets you to scale quicker. Right. So for us, the thing that we're trying to do with our customers are trying to help them understand what that workload is going to perform like in their own environment. So when we bring them into the AI proven ground, we help them understand really how this is going to function. And when we do that, they get to scale their workload. They get to scale their use case much faster. And so for them, it gives them the business impact a lot quicker than trying to run an on-site POC where we're trying to find equipment. Oh, by the way, is there enough power? Does it have to be liquid cooled?

10:29Mike Trojecki: Do I have enough physical space to do this? And those are the things that we have to take into consideration when looking at any of these use cases and workflows. I don't need a GPU for everything like you said. a CPU for this workload, maybe it's a computer vision workload, CPU is going to do just fine in that type of environment. Am I going to use a CPU for large scale model training? Probably not. That's where I'm going to start looking at rack scale. That's where I'm going to start looking at some of the bigger kind of GPU environments. But it all depends. And, you know, Johnny, you mentioned this.

11:05Mike Trojecki: One of the things that we're doing is helping our customers going through this kind of AI capacity planning, right? What does it look like? So we talk about the technical piece, we're talking about the commercials behind it, but there's also regulatory pieces in there. There's a skillset conversation that comes along with it. There is the responsible AI side of things, space, power cooling. Like we said, there are just so many different variables that people need to consider and are going to have an impact on the overall cost, just in hardware, in what their monthly spend is, and from a tokenomics standpoint.

11:41Mike Trojecki: So if we, you know, jointly working with AMD, I know we have all of your stuff in our proving ground, and we use that to help our customers make the right decision on the workload, I think we'll continue to help AI advance. You know, funny little pun, you just talked about advancing AI. Thank you for that one. Nice work. It was a phenomenal event, by the way.

12:02John Hampton: No, thank you. And you're absolutely spot on, though. This is the kind of conversation that we all should be having, this type of collaboration. And I applaud those that are listening into the podcast to kind of gain some of these insights and just get the knowledge to work through this. But we want to help you work through that complexity and build some awareness for what the right tools may be. And I just want to amplify what Mike said there. In the WWT Advanced Technology Center and this AI Proving Ground effort, you can kind of put a lot of this to the test. We can certainly run TCO and financial modeling and give estimates of how we can improve your tokenomics.

12:45John Hampton: But it's about proving that out and testing it. So Mike's absolutely right. What we're seeing is a lot of these enterprise AI use cases could absolutely get routed to CPUs, routed off, again, frontier models in some cases to on-prem. Maybe it's a lower cost, lower power GPU. you. Mike, I have to tell you about AMD as an example, similar to WWT. Just think of AMD as one of the largest tech companies in the world, all the engineers, the software development. Really, that's become the killer app, right? That's one of the reasons we're seeing these token bills completely explode because of the benefits that we have in AI-based coding.

13:31John Hampton: So we we like I think we blew through our entire 2026 budget in the first month. And they had to kind of come and say, look, here's your your your drink tokens and and kind of really kind of reevaluate what is the most effective spend here. So what we did is we started routing some of our internal AI use cases from Frontier models onto our new Instinct MI350P GPUs. These are PCIe solutions that are more aggressively positioned. And by doing that, we improved our token bill by 43%. And here's the kicker. Here's the beauty of it. Again, the technology impact is important, too. So we saw a 2.9x improvement in faster responses by also doing this intelligent routing.

14:26John Hampton: So we start sharing that with enterprise and that's exactly what I want to be doing. Well, that's where, again, this partnership with WWT comes in. Let's go put this to the test and prove out what we could do for your token bill.

14:40Mike Trojecki: Now, again, going back to our lovely conversation around Top Shelf, if there was actually tokens that we were allotted per month for drinking, I think my wife would probably be pretty happy with that if she was able to give me an allotted set of tokens.

14:56John Hampton: Yeah, I hope she doesn't listen in. Yeah, that would be too bad.

14:59Mike Trojecki: Yeah, she's not going to listen. She's like, oh, you're on a podcast? Great. Well, John, let's stick with that internal pilot that you're talking about there. I mean, fantastic results, but I'm curious, what exactly were you all routing? Was it individual prompts? Was it entire applications? Was it models? Was it stages within an agent workflow? Walk me through a little bit of that process.

15:22John Hampton: Yeah, this is really the right question because really every enterprise is a snowflake. We're all unique and different. And a lot of AMD's internal engineering and AI use cases do different. So you really have to think, and again, you're dynamically analyzing what are the incoming AI use cases. and there's an on-the-fly evaluation of, okay, what would be the most cost-effective solution, but what is also going to be the most important routing to maximize performance as well. So this is really the next wave of intelligent dynamic routing. But for example, if you're thinking about a lot of software dev, it is more complex.

16:06John Hampton: We're getting into coding here. So there's absolutely a time and place where you need to be routing to the frontier models. But as you could imagine, through a lot of this engineering development, it could be a little bit less complex. We think about web searches, maybe it's small context where you have some reasoning, but you don't need to scale quite as much. So some of those types of AI use cases, Brian, we were able to route to these on-prem local model MI350P solutions and, again, offset that. And we're talking about millions and millions of dollars on the token bill. So, hey, there's a time and a place, Mike, to have a little sip of Pappy, and we can all appreciate that.

16:52John Hampton: But, again, that's not necessarily sustainable for large enterprise. So there's a time and a place maybe you would just crack open a beer or have a glass of wine and you're still able to enjoy yourself. I agree. We're not shutting the bar down, but we got to make sure that we're taking a reasonable approach to all of it. So, Brian, I think that's really the key here. And it's about examining and assessing those same set of workloads and AI use cases together as an industry. One last point on this. Together we advance is really AMD's motto. And we really have to be doing this as an open ecosystem.

17:35John Hampton: Last week, there were a ton of exciting announcements from our partners around the industry and many that want to come and support these types of efforts. So one example of that, Mike, was SpectroCloud. This is a new partner that WWT works with. And there's new emerging reference architectures that can absolutely help with this intelligent routing. They were talking about, this is actually a Goldman Sachs research reference that token consumption will increase 24-fold by 2030. So this is only going to continue. So SpectroCloud and this pallet AI solutions coming now alongside AMD and WWT to really think about bringing inferencing closer to where the work happens, intelligently routing requests across local frontier models.

18:27John Hampton: There's this concept of meter token usage. So we're actually now dynamically evaluating our token spend. And so I think really the point here is together, we can start to continue to scale and ramp these types of efforts to really dial in the token economy for enterprise. Yeah.

18:50Mike Trojecki: And, you know, absolutely agree with you on that. You look at like SpectreCloud and then you look at just some of the optimization stuff from, you know, companies like Raffae, Armada, formerly Arna, and you've got all these, you know, and you're starting to do software side. You're starting to see the optimization piece become more critical across the board. And it's funny because we talk about this kind of world of what we're seeing now, all these open-weight models, right? And push within our own government to, you know, with people saying, hey, please do not limit the open-weight models. And you have to weigh the risks of all of that.

19:30Mike Trojecki: But looking at these models, looking at small language models, that is going to absolutely impact the tokenomics piece of this. And, you know, it'll cut costs in a lot of cases, and it'll give people more control over what they're doing in their own environment, right? Especially as we see some of these workloads, you know, they may start in the cloud, might get repatriated to, you know, on-prem. They may stay in the cloud. But again, you have to look at what that model looks like and how you're planning for that workload in whatever environment you're looking at.

20:02John Hampton: Yeah, maybe just one point to just expand on here for a moment to come back to your point there on the software and how the software is evolving. The amazing thing about this is how AI is really helping improve AI based software. We actually recognize as a company at AMD that really the biggest area that we needed to focus on in the AI race was our software, was our Rockham-based solution. I'm very excited to now announce Rockham AI last week. And I'm telling you, Mike, it's incredible what we're seeing with AI is really helping us address that software challenge and scale faster than we've ever seen before.

20:47John Hampton: Our engineers, again, are so much more productive than we've seen across their careers. And we're just very excited to see how this will help really AMD become an excellent second platform solution for AI because of the accelerated software that we're now bringing. So that's an area I'm trying to call some attention to so that everyone's, again, aware of this just evolving software space. We were showing, Mike, AI having the ability to write kernels live in front of the audience just blew the minds of a lot of our partners and customers were there. So what are you seeing in that space?

21:28Mike Trojecki: Yeah, I will tell you, I've had similar feedback. I had a number of my team attending and seeing that just people are just stunned by what's happening in the software space. If you look at us, I will tell you the two biggest areas we're investing in right now are on the data side and on the software side. So, I mean, we've now built an entire team just on the software pieces for AI. And you look at where that's going and other teams throughout WWT where we built out organizations focused on the agentic piece, on the optimization piece, the automation side. Software is king. So what you guys are doing with ROKM and kind of stepping into that world, it's not like you haven't been there before, but really upping the game there is a differentiator for you guys.

22:15Mike Trojecki: And I will tell you, it's making a difference in the conversations or will make a difference in the conversations we have with our customers. This episode is supported by Lenovo. Lenovo provides enterprise PCs, servers, and edge computing solutions engineered to power the modern hybrid workforce and accelerate intelligent edge and AI deployments. Equip your organization with Lenovo's reliable, innovative technology built for the future of work.

22:42John Hampton: Yeah, excellent. I really have to say, too, I've been very impressed by WWT really shifting to becoming an AI native company. So a lot of respect for Jim Cavanaugh and the WWT leadership team and certainly have a fantastic relationship with Dr. Lisa Su and our executives. But these are the things that as leaders, Lisa and Jim are driving our companies to become more AI native, where everything we do starts and ends with AI, helps us scale to this next era that we're talking about. But again, it all does come back to the data, the software and the token economy in order to scale in the right way.

23:23John Hampton: This is a conversation that we want to be having with other large enterprise.

23:28Mike Trojecki: Yeah, we don't want to give all the credit to Jim. I mean, he's not a little bit to help the organization out, but I'm sorry, Jim. I love you.

23:36John Hampton: If you're listening, Jim, I'm a big fan. And yeah.

23:39Mike Trojecki: No, but here's the thing there is you talk about an AI native organization. When I got here two years ago, the conversation was we are going to be an AI first company. And Jim drove that vision into the organization. And he has been a, I mean, he's been an advocate and a participant. I mean, he hosts a call on Monday mornings, an executive call, and all we're talking about is how do we use AI to improve WWT? And that should be a lesson for every other company is don't just do this in your tech organization. You need executive level leadership, and you need somebody like Jim who is a visionary when it comes to saying just those simple words.

24:25Mike Trojecki: We are going to be an AI first company. that changes everything for everybody in the organization. Well, Mike, before we go any further, I do want to understand as it relates to that internal pilot with AMD, I mean, fantastic results, 43%, token bill reduction, 2.9 times response speed, amplification. What should listeners learn from what John just detailed? And what should they also be a little bit cautious about in terms of not making an assumption on how it might work in their environments? Yeah, I would tell them not to make an assumption. I typically leave that to you to make the assumptions.

25:05Mike Trojecki: So just to work earlier. So here's what people should be looking at with what John said. You do not have to, like you said, hit a nail with a sledgehammer. There are a ton of options out there and you need to look at every different path for the workload. you need to really look at the cost control piece of this. To me, seeing that type of change in an organization, especially an organization that is as big as AMD, should serve as notice to every company out there that AI can absolutely impact your business and you can see the financial return out of it. Obviously, costs are increasing, tokens.

25:52Mike Trojecki: We've talked about that in the podcast so far. But it's going to shift. That token cost may increase, but other areas of the business, the costs are going to decrease. So you need to look at this, again, from a business standpoint, look across your organization. And that's what AMD did is they looked across the business and figured out what are the appropriate use cases and where they can see the most benefit.

26:19John Hampton: Yeah, just one other thing I want to bring some attention to because, Mike, I know you hear this too. And I've actually even heard this discussed on your podcast. But what we're also seeing is a lot of enterprise really struggle with the budget and the power. Most of them are out of physical space to deploy on-prem solutions. And so one of the things that AMD and WWT have partnered on is to help them modernize their infrastructure for AI, right? So I think if you back up and, again, kind of assess your existing infrastructure, you may be surprised with how much you can really modernize, consolidate your servers down.

27:03John Hampton: We're seeing, Mike, I mean, as much as seven or eight to one consolidation, in some cases, take a thousand servers down to 150 maybe. And imagine the budget that you free up from hardware reduction, software licensing cost reduction, all the power, electricity costs that have been freed up, administrative warranty costs. I could go on and on. And imagine the physical space that you're going to free up. So again, the AMD and WWT have really a proven formula here that can be tested and proven out in the WWT Advanced Technology Center. We have POCs set up. You can run labs. Maybe you need to test on it in a virtualized environment and put this to the test.

27:50John Hampton: And that's really a great call to action there is to really make sure that you have modernized for AI and you now come to the table with budget power and space that can then be reinvested in the most optimal solution here for best overall tokenomics. That's what I think is different now, Mike. I've been around a long time. I've got the gray-haired. I noticed you had a little. This is different. This is different than it used to be where it's just like a simple refresh, you know, lift and shift kind of thing. We have to be modernizing, preparing the infrastructure now for the most optimal AI and most optimal token economics.

28:33John Hampton: Like it's a business decision in addition to just a pure tech refresh. Right. I don't know if you want to expand on that.

28:42Mike Trojecki: Yeah. And I just I make my hair like this just so I seem a little bit wiser. That's all. I'm actually younger than I look. But that's the thing is when you look at this, this is why we've created, we have a whole facilities team that is dedicated to looking at space, power, the cabling costs. Cabling costs alone are extraordinary. So if you're able to consolidate down and you can do and get these workloads supported where I don't have to have such a power hungry environment, that saves a ton of money from a cost standpoint. And just the power consumption, I mean, we're looking at some of these racks, you know, are what, 600, getting up to 600 kilowatts per rack, getting towards a meg per rack.

29:35Mike Trojecki: I mean, you've got to look at that in what you can actually support. I mean, even in the AI proven ground a couple of years ago, we realized like, all right, we need to go and upgrade all of our power, all the cooling pieces in here to support all of these different environments. So we can look at this and say, okay, customer A, customer B, whatever industry you're in, going and helping them determine the workload, part of that is the power of cooling. And that goes into that commercial side of things from a cost standpoint. Because a lot of enterprises, unless you're a huge enterprise, you're in the space and power piece.

30:12Mike Trojecki: You're going to need to work with a colo partner, one of the partners like Equinix or a digital realty that we work with. I mean, those are the things that we look at and say, okay, it's not just token costs. It's not just hardware and software costs. It's the actual operational cost of your environment. So I couldn't agree more. Yeah.

30:33John Hampton: You know what that was just now, Brian? Wisdom that you just heard there.

30:39Mike Trojecki: All in the hair. That's right. He was right. I mean, John, so the first half of this conversation was mostly about routing efficiency, routing workloads intelligently. Now you guys are starting to talk about systems efficiency. AMD has said it's next-gen Helios architecture producing more tokens per dollar at the rack level. I think it was up to 30 more tokens per dollar or somewhere in that area. Let's dive in there because that's solving a different part of the equation.

31:09John Hampton: Yeah, thanks. It's such a fun topic. AMD has really now evolved so much. Again, I've been around a long time. I like to think I have some wisdom, but AMD used to be just kind of this low-cost alternative in the CPU space. And now you're seeing announcements like Helios, which is full rack-scale solutions with GPUs, CPUs, AI NICs, networking, DPUs. A lot of that's come through major acquisitions and just bringing more and more innovation. just Lisa Su's maniacal focus on engineering execution. But that's exactly what we're doing is bringing really a second alternative solution to the market that desperately needs it.

31:53John Hampton: Mike, I don't go a day without hearing, look, I need another player in here to help me manage my both financial and technology impact within my organization. So I think it's so important that We have choice. We have silicon diversity. But we've really seen this been reinforced with a lot of major announcements and partnerships around Helio. So, Brian, even last week, we heard from Anthropic. We heard from Microsoft on the Azure side, OpenAI, Meta, Oracle. I could go on. Really, the point here is the largest mega data centers have all adopted multiple gigawatts of power in compute. I believe that could power the DeLorean, multiple gigawatts of compute.

32:42John Hampton: But the real exciting thing is how all this comes together in a full rack scale solution. Again, this is an area where WWT can help AMD bring this to market in the ecosystem. Brian, if I could, I have a little show and tell. I want to showcase really next gen technology that was really announced and launched last week at AMD Advancing AI. I'm holding up here the new MI455X GPUs. In a Helios rack, imagine 72 of these GPUs clustered together. I got to tell you, the industry, I think almost pervasively, is very, very excited about that MI455X solution, which is really going to bring a level of competitiveness that we all really need.

33:34John Hampton: In addition to that, you will see in Helios our 6th Gen Venice Epic CPUs that really help drive a more efficient host node environment and really take the CPU aspects of this solution to the next level. And we even have AI Nix that will be in these solutions. So, again, really the point here is we can scale up into that compute AI inferencing spectrum all the way up to these Helios solutions that will be as competitive as any solution in the market. But the point is we can also right size the AI solution to the specific needs of every enterprise. And the beauty here is AMD and WWT can serve up that whole array of inferencing solutions for you to test and fine tune right size for your environment.

34:30John Hampton: Maybe just get a normal hammer versus a sledgehammer for some of those use cases. And that's exactly what we're together trying to bring. I may just add one thing here to Brian that often does get forgotten about, but that to really complete that full AI inferencing spectrum. we really ought to be thinking about desk side solutions. So I'm referring to AI PCs in this new category that we're talking about, agent PCs. And for those of you listening in, we'd love to talk to you more about this space. Again, if we are offloading from the cloud in some cases, and we're now doing various AI use cases and local inferencing on-prem, we're going closer to where that inferencing happens.

35:17It makes so much sense to have a high-powered agent PC to do a lot of that development.

35:24John Hampton: Agentic workflows can be developed, desk side. And again, a lot of these tasks that we've been talking about throughout here together, you could run on AI PCs on a laptop. And really, what we've noticed is that is really becoming an important extension of the tokenomics equation. and it really is that taking it a little further to improve that token bill because again there's a time and a place to to run various use cases on these frontier models you could scale all the way up to helios if you need to on the hardware side but again so many of these new emerging vertical ai use cases small language models and local inferencing low latency type environments can absolutely be run on AI PCs and agent PCs.

36:15John Hampton: So again, we really just encourage you to look at that full end-to-end spectrum for your infrastructure and look at every way you can to totally dial in your tokenomics.

36:26Mike Trojecki: Yeah, and what, John, the cool thing about this, right? You look at this and you go from, you know, Helios, which, you know, I think it's obviously just a beast, right? It's what you guys are doing there, from just a scale up and scale out standpoint is incredible. But you look at where you guys are going and starting just from you look at Epic and then you look at what you guys are doing with Pensando and on the Nick side, look at Instinct, like you said, with the 455. Now we're talking about doing this at the desktop level. That is the next wave, right? We've talked about all this training stuff, but really where AI comes to life is at the edge, at inference.

37:14Mike Trojecki: And you're starting to see a lot of these, you know, we've got a lot of our partners that are, you know, these GPU as a service providers. But what we're starting to see now is inference as a service provider. Again, don't need massive GPUs to run from an inference standpoint. Now I've got the ability to keep cost control down at the desktop. I'm able to do that in my AI factory, in the cloud. And you guys have a solution that kind of spans across the board. And obviously throw the Rockham piece on there with software. And you guys are in a great, great position from an AI standpoint. We're excited about it.

37:56John Hampton: Yeah, thank you. I just maybe want to throw one more thing in there. Again, we want to do this as a full partner ecosystem. We know really your top partner is Cisco. And I just wanted to give an example of where some of these solutions are emerging. We got to hear from G2 Patel on stage with Jack Kuhn last week at AMD Advancing AI. I also had the privilege to hear G2 speak with Jim Cavanaugh at your flagship event last year at WWT Championship. And so a very exciting announcement that we heard from Cisco. They're bringing now this type of agent PC, this desk side solution to the environment with AMD and talking about really the importance of networking, the importance of security, governance, some of the things that you've mentioned here.

38:46John Hampton: The ability to track your tokens, your tokenomics. A lot of that really should and can be done desk side on-prem. But again, we also now are seeing Cisco GPU solutions like this MI350X based solution, both of which are going to be available for testing in the WWT AI proving ground. So it's really exciting to see how AMD has evolved to have a really broad spectrum of various AI-based solutions. But then to see Cisco and WWT kind of wrap around that, all of your strongest capabilities, it really starts to come to life in a very compelling way.

39:27Mike Trojecki: Yeah. And we live, I've said this a number of times, and Brian's heard me say this on other podcasts, but AI is an ecosystem world. There is no one out there who is doing this on their own. Not WWT, not AMD, Cisco, anybody. So we've got to look at this as how do we build that ecosystem out because there's just so many different pieces and parts from software to component level to management and orchestration. It's all about the ecosystem. And that's why I look at this as this world being a partner first world. it certainly takes a village and um based on what the uh two of you have provided so far today certainly we're in good shape to do that and uh i don't know what time everybody out there is listening to this episode right now it's almost 3 p.m central time and uh given the uh the top shelf uh nature maybe happy hour starts a little bit earlier uh today but mike john thank you so much for joining the podcast uh this afternoon um we'll definitely have you on again in the near future to talk about what I'm sure will be some big updates.

40:31All right.

40:32Mike Trojecki: Thanks to Mike and John for joining us today. The takeaway from this episode, winning with AI will not come from throwing the biggest model at every problem. It'll come from matching each workload to the right model, processor, and location, and proving the economics before scaling. This episode of the AI Proving Ground podcast was co-produced by Nas Baker and Kara Kuhn. Our audio and video engineer is John Knobloch. My name is Brian Felt. Thanks for listening. We'll see you next time. Bye-bye.