Lenovo on Scaling AI Efficiently: Rethinking Infrastructure for a Power-Constrained Future

Enterprise AI infrastructure planning now starts with land, power, and water. As on-premises AI deployments scale, GPU availability is no longer the binding constraint. Site capacity, power draw, and water availability now decide what an enterprise can actually build and how fast it can bring new AI capacity online.

At the Six Five Summit: AI Unleashed 2026, Patrick Moorhead opened the Sustainability Track with Linda Yao, VP and General Manager of Hybrid Cloud and AI Solutions at Lenovo's Solutions and Services Group. Their conversation covers what happens as enterprises move AI workloads on-premises, and how decades of high-performance computing experience are shaping the way Lenovo designs for power-constrained deployments today.

Yao traces the shift directly to token economics. Enterprises that moved workloads on-premises for cost control found on-prem deployments up to eight times more cost-effective than token-maxing on a cloud subscription when facility planning, racks, and cooling get designed together from the start. Rack density has more than doubled since 2021, pushing liquid cooling from optional to standard for AI-scale deployments. Lenovo's own numbers back the approach: customers using Lenovo Neptune liquid cooling and services have cut energy costs 40% and sustained 10% higher performance, and DreamWorks Animation completed a full infrastructure deployment in 1.5 days with a 20% rendering speed increase.

Yao also pushes back on the idea that AI infrastructure requires wholesale replacement. Older systems remain valuable for lower-intensity or less time-sensitive workloads, and Lenovo's circularity practice, reusing, refurbishing, repairing, and recovering existing hardware, helps enterprises extend infrastructure life and lower technical debt without a full AI refresh. Her core argument: enterprises that plan for power from the beginning will be positioned to keep pace with AI compute demand five years from now. Those that treat power as an afterthought will find it limits how much AI capacity they can bring online and sustain.

Key Insights:

🔹Power and water availability now set the ceiling on AI capacity, ahead of GPU count in the planning process.

🔹On-premises AI deployment can be up to 8x more cost-effective than cloud token-maxing when facility, racks, and cooling are designed together from the start.

🔹Rack density has more than doubled since 2021, pushing liquid cooling into standard requirement territory for AI-scale deployments.

🔹Lenovo customers using Neptune liquid cooling have cut energy costs 40% and sustained 10% higher performance; DreamWorks Animation completed a full infrastructure deployment in 1.5 days and saw a 20% rendering speed increase.

🔹Circularity offers a practical alternative to a full-scale AI refresh. Reusing, refurbishing, repairing, and recovering existing hardware helps enterprises extend infrastructure life while reducing costs and technical debt.

Lenovo's approach treats power as a design constraint from day one, letting infrastructure scale directly with AI workload demand.

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

Catch the rest of the Sustainability Track and full Six Five Summit: AI Unleashed 2026 coverage at sixfivemedia.com/summit, along with Six Five's on-site coverage of Lenovo's Big Tent event in Las Vegas.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

‍

Linda Yao:
We reuse, we refurbish, repair, and recover to extend the life of a reliable infrastructure and make that AI adoption less expensive and less wasteful. Technology modernization, in our opinion, can still make use of past investments. And we've been able to help our cloud provider and our enterprise customers alike with circularity and true scale.

Patrick Moorhead: 

Hey everybody, welcome to the Six Five Summit 2026. The theme is AI Unleashed. And yes, we are seeing the enterprise benefits of agentic AI. It's amazing that three years ago, we were pretty much stuck on recipes and LLMs. We have come a long way. Security has come to the forefront, right? And it just makes sense that we have to secure All of this and different conversations are forming out there and power constraints are one of the key issues here. And it is wild how we've gone from, oh, AI is going to be all cloud to, oh, it's going to be a balance and it's going to be a hybrid. And by the way, when all those AI workloads come on-prem, decades of experience should come into play because the industry has had related challenges and solutions. And to discuss this topic here is Linda Yao, VP and GM, Hybrid Cloud and AI Solutions at the SSG division of Lenovo. Linda, welcome back to The Six Five.

Linda Yao: 

Thank you, Patrick. It's great to be here.

Patrick Moorhead: 

Yeah, it's always fun. It's good. You and I interact more than on video, which makes this fun. But I love this topic, and I love that you're doing this video, because quite frankly, you're right in the middle of everything at Lenovo, cutting across all the hardware with all the services, direct customer feedback. So thanks for coming on here. Hinted at this a little bit in the lead in but for years the assumption was that I was going to be you know a cloud only world and by the way, nothing bad about cloud only I mean Lenovo benefits a lot from that you have a lot of customers that you serve there. But boy, has the conversation changed or maybe it's the reality has finally caught up with what all of your customers have been saying forever. Why are we seeing this shift toward on-prem AI where putting power is at the center of infrastructure planning?

Linda Yao: 

Yeah, I mean, Patrick, you and I both spend a lot of time with IT decision makers, and CIOs, as well as line of business owners, are bought into the promise of enterprise AI, like you mentioned before. That being said, those AI workloads and the token consumption is driving up power demand, rack density, and cooling requirements. all at the same time. So what we're both seeing now is that the limiting factor on AI is no longer just how clean your data is, or how many GPUs you can access, or how many tokens your model can eat, but where the AI can be built, how fast it can be deployed, and how efficiently it can run. So I live in California. The scarcities that we deal with every day are land, power, and water. And that is exactly what you need to build AI. So because of that, the site, the power source, and the cooling strategy now matter just as much as the compute. The question that CIOs are asking me is no longer just like, can we get the GPUs? It's all about, can we power this data center environment efficiently enough to make the AI work safely, economically, and sustainably? So enterprise AI outcomes, model routing, performance optimization, all of that still matters for sure. But to use AI, you literally have to build AI first. So it all starts at the site, and those are the first fundamentals to solve. Land, power, water.

Patrick Moorhead: 

Yeah, that makes a lot of sense. And we talked about this conversation from all cloud to hybrid AI, and there's also a lot of interest in sovereign AI. The other conversation that it literally it took three months to change was we went from enterprise token maxing, having leaderboards who could consume the most tokens because we thought, hey, the more tokens that our employees are consuming, that must mean they're more productive, right? And then the bill started coming in and you had companies consuming 50% of their token budget in a month. And then we very quickly swung to token efficiency. So as more of these AI workloads move on-premises to be more efficient, how do enterprise need to be thinking about infrastructure related, you know, using as little power and as little water as possible.

Linda Yao: 

You and I have probably spent more tokens than we can afford in our experimentation, but we're not alone. So token economics or token ops, these are now the buzzwords of the day. We have seen organizations take models on-prem for more control over their data and their knowledge, but also their compute costs. So sovereign AI, hybrid AI, on-prem AI, it changes the equation when it comes to cost. we have shown it at Lenovo to be up to eight times more cost effective than token maxing on a cloud subscription. So to bring it on-prem, to get those cost savings, though, the site, facility, the racks, the water loops, they all have to be designed together from the start with your IT infrastructure. So we've found the smart move is companies who do their strategic planning up front rather than a bunch of tactical fixes later. That means they size the environment for the workload. They're not just filling racks and hoping that the infrastructure and the power sources will keep up. That means they plan for liquid cooling because rack density is rising fast. It has more than doubled since 2021. And the traditional cooling approaches are not going to be enough. They also know that they need to have the right experts on site with them to bring their data center up to peak performance and make sure that it stays there throughout its useful life. Our goal, both for cloud providers and for enterprises, is to use as little energy and water as possible while guaranteeing performance. And we've built our AI expertise around that, from Lenovo Neptune technology to our power and cooling services, and right-sizing through TruScale. It all comes into play.

Patrick Moorhead: 

Yeah, it's funny. I get in the room with somebody who is operationalizing AI inside of a a medium or a large enterprise. And they tell me what they're getting done by using their on-prem infrastructure and the type of compute that they have, leveraging small language models or very focused models. And it's incredible when they tell me what they're getting done in a single rack of air-cooled GPUs. It is an absolute mind blower to me. And these aren't little wimpy workloads. These are enterprise grade across half a million employees that they have. And I'm super, super impressed with that. You know, it's funny, Linda, I have people who say, don't tell anybody how long you've been in tech. They'll know how old you are, right? This is my 36th year in tech, but this relates to my next question I have for you. There's really nothing new that I haven't seen. It might have a different name, but the themes are the same over that time period. And related to Lenovo's experience, these power, cooling, and utilization challenges aren't new for Lenovo, right? In fact, you have decades of HPC experience that are shaping the way that you think about scaling this enterprise AI. Can you share how you're relating HPC to AI? A lot of people are like, oh, no, different market. HPC is about flops, and AI is about tops. How are you solving those problems?

Linda Yao: 

Yeah, I appreciate your recognizing that and making the connection. So Lenovo does have a long track record in high performance and supercomputing, whether air cooled or water cooled. And this makes it a natural conversation for us. We have been the number one provider of supercomputers in the world for several or so consecutive years now. And the fastest supercomputers they use Lenovo's Neptune liquid cooling and professional services. So we have decades of experience with these high-performance AI environments at scale. And with Lenovo Neptune, we can deliver warm water cooling of entire systems, component-level cooling for CPUs, GPUs, memory, what have you, liquid-assisted cooling, or even rack-level water cooling. And that is how, with that combination of skills, Lenovo has helped some customers achieve 40% lower energy costs, 4-0, while sustaining 10% higher performance. We can also do things like remove the heat efficiently from the full system, perform preventative maintenance on the data center environment beyond the GPUs, and use TrueScale to meter the power usage so customers can keep right sizing over time. DreamWorks Animation is one of my favorite customers. They're in California, where land, power, and water are scarce. They're a great example of what this looks like in practice. So Lenovo worked with DreamWorks to deliver a 20% performance increase. So that improves their animation rendering speeds and enables their artists to have much faster iteration cycles when they are creating those wonderful movies and films. And we were able to complete that whole infrastructure deployment for them in just 1.5 days. And we supported that with Lenovo services through custom hardware integration, white glove services, and proactive support. So they could get production ready really fast. That's just an example of how our liquid cooling, our professional services, it's not just working in theory. It's been proven over years at scale and at the highest levels through many tiers as well as tiers in the lab and on the field.

Patrick Moorhead:

I like that, cheers and jeers. A little side note, by the way, the HPC architectures and AI architectures are very similar. I mean, look at the networking. They used to be completely different, and now they're very, very much related in. And sometimes people forget that you were, I believe, the first water cooling vendor that was using x86 out there, and your lineage goes back really far. And it's funny, I had known you were number one in HPC, but had forgotten it. So thank you for reminding us all about that. One other thing that I've seen in technology is that Every time the next bright, shiny thing comes on, it says it's going to replace something. But the reality is there's still a lot of mainframes, a lot of minis, client-server computing, social, local, mobile. I mean, I remember the meme that tablets were going to replace PCs. Now we've got very robust tablet and PC markets. But the reality is, particularly in enterprise, is they want to keep hold of their investments as long as they can. And one of the memes out there for enterprise AI is that you have to replace everything, right? Take your data center, I don't know, gut it, move it to a colo, rebuild everything with DLC and higher wattage capabilities. But that's not an absolute at all. By the way, good for Lenovo, good for the industry if that were the case, but not necessarily optimal for every customer need. I've seen you talk about this, things you've written about this. What are ways that organizations can lower the barrier to AI adoption by getting more value out of the stuff they already own?

Linda Yao:

 Yeah, I mean, for sure. IT budgets are precious. And the goodwill that you can generate and the excitement around AI, that's also precious. So even though there is a lot of AI FOMO right now, not every organization is ready to rip and replace everything that they own. The smart path is to lower the barrier to entry by getting more value from the infrastructure that's already in place. Whether that means right sizing the environment, taking steps to reduce technical debt, or finding the next best use case for those existing assets. So older infrastructure doesn't always need to be retired. It might still be very valuable for lower intensity workloads or those that aren't time sensitive. And circularity also matters a ton. At Lenovo, we reuse, we refurbish, repair, and recover to extend the life of reliable infrastructure and make that AI adoption less expensive and less wasteful. Technology modernization, in our opinion, can still make use of past investments. And we've been able to help our cloud provider and our enterprise customers alike with circularity and true scale. It helps to monetize their tech debt, match capacity more closely to demand so that they can grow into their footprint and don't need to buy more than they need at the peak.

Patrick Moorhead: 

Yeah, it's funny. I think you might be the only tech vendor that is actually talking about this. So this is a very novel concept. I can actually take more advantage of what I have today. I love that. Hey, I want to go a little bit more long-term here, Linda. This has been a great conversation so far. So, I think this is true in every major inflection that we've had, that the companies that win the AI race won't just be the ones with the biggest balance sheet spending the most money. But five years from now, and by the way, I think you agree, maybe you don't, maybe you do, but if you do, five years from now, what will separate the enterprises, the organizations that plan for power from the beginning from those that will ultimately be limited by it?

Linda Yao: 

Oh my goodness. Five years from now in the world of AI is hard to predict.

Patrick Moorhead:

I know this is a tough one. Any, any time.

Linda Yao: 

Yes. No, but that being said, I expect that the winning organizations, I agree with you might not be the ones who spent the most money, but the ones who plan for power from the start. It is one of the main constraints now on how much AI capacity, a company can actually bring online and sustain for their users. So the organizations that plan ahead are going to be positioned better to keep up with this seemingly unlimited demand for AI compute and AI workloads, and keep up in a way that doesn't keep them up at night. That is where Lenovo comes in, right? We help these organizations design for performance and sustainability from day one, architecting the full data center for the AI mission, planning for the right power envelope, and providing the full lifecycle of professional services so the infrastructure can be built to scale sustainably and perform at peak from the start.

Patrick Moorhead: 

Yeah, that just makes sense. And what's interesting is I listen carefully to what you're saying, particularly around this strategic point of view on this. You literally have to, by the way, SSG does this for a living, right? You map out your prioritized workloads from top to bottom. I think we've done a lot of experimentation and a lot of POCs. And now it looks like the customers are really centralizing in on those that can be governed and those that are unique and of course, that deliver ROI. you have to have power as a power and water as a part of that conversation as well, as opposed to some mythical place that we can run this. And now that it's clear, and I call this years ago, just because history mostly repeats itself, we've gone from 80% training with 20% inference, and it's completely swapped. And I think for the enterprise, even more pronounced, because there wasn't a lot of training going on out there, maybe you're customizing some models for you. But yeah, if you don't have power as part of that conversation, you're not really having the correct conversation.

Linda Yao: 

Land, power, water, Pat, that's where it is.

Patrick Moorhead: 

No, I like that. And you're in California, I'm in Texas, where it seems like we mythically have unlimited amount of everything. But we're unique, we're special, I guess. Linda, I want to thank you for coming and kicking off the sustainability track. for the summit, really appreciate it. Always like your insights. And I do like the fact that we just don't meet on video like this. We're having some meaningful conversations outside of this forum. So thank you.

Linda Yao: 

No, it's great to be here and can't wait for the rest of the summit.

Patrick Moorhead: 

For sure, thank you. All the viewers out there, don't forget to hit subscribe, follow us on social media, and check out all of our Six Five Summit content and all the Lenovo content. Hey, check out, we were at their Big Tent event in Las Vegas, multiple days of coverage. You can see that on sixfivemedia.com. Stick around for more insightful conversations. Take care.

‍