Enterprises have finished poking at artificial intelligence. Now they’re putting it to use and paying for it. Way more than many anticipated.
Research firm Deloitte calls this the “pivot to tokenomics,” a shift from experimentation to deployment that’s forcing organizations must learn to manage rising use and cost of using LLMs. This is driving IT teams to ask a blunt question: move all AI capability to the cloud or own the infrastructure on-prem or in a colocation facility.
"The one great way to address that increasing cost is to own your own token machine," Sumit Puri, founder and chief technology officer of Liqid, told The Forecast in a video interview captured at the July AMD Advancing AI event in San Francisco.
CIOs and IT teams are figuring out how to get the most out of scarce resources. They don’t want to get caught on a runaway token train, Puri explained.
“The cost of these tokens is front in mind for them,” he explained. “Large organizations are now saying, ‘Hey, we're going to limit the amount of tokens that you as an engineer inside of our organization can consume because the cost of that is becoming very high.’”
One way to address these rising costs is to “own a token machine,” he said. “Bring the infrastructure, bring the GPUs on-prem or in a colo (colocated data center) so it's a one-time expense. You can consume all the tokens that you want out of that investment that you make.”
How Are Enterprises Managing IT Costs to Meet Rising Cost of AI?
Many organizations are shifting to on-premises infrastructure to transform unpredictable, ongoing cloud token costs into a predictable, one-time capital investment. In 2026, Tokenomics and the technologies involved are evolving rapidly. IT teams must now track and govern token use with stronger control over their infrastructure. Then they need to manage what new and future AI agents actually do on that infrastructure. As a result, attention is zooming to agent gateways that route traffic, enforce access policies, track token consumption and record every interaction for audit and compliance purposes. For example, the agent gateway in Nutanix Enterprise AI software governs the behavior of autonomous agents that can initiate actions, chain decisions, and interact with multiple LLMs databases and business tools in real time.
As agentic workflows juice token consumption, enterprises face usage caps, rate limits and unpredictable repricing from model providers. This is well documented in the recent report by EY, Unlocking agentic value: a new investment discipline for the agentic era. Puri helps Liqid customers leverage composable infrastructure software that pools GPUs and memory into shared resources dynamically assigned to any server, the data-center equivalent of a shared tool crib where every wrench reaches every worker the moment it is needed.
“We can have this matching of infrastructure to model to not have any wasted resources and improve the tokenomics,” Puri said. “In a resource-constrained environment, utilization is critical. You cannot afford to have your GPUs at 30, 40% utilization. So if we can implement technologies that drive utilization to 100%, we can do more with the same amount of infrastructure.”
How IT Teams Maximize Hardware Efficiency and Support Massive AI Models?
In his first interview with The Forecast in 2025, Puri described how composable infrastructure allows IT teams to dynamically pool GPUs and scale resources independently, enabling them to run massive frontier models on a single server rather than expensive clusters. Today, IT teams now know well that AI model size dictates infrastructure needs. A small model may need a single GPU. A medium model may need eight. A frontier model may need dozens. Puri said customers cannot predict which size they will need, so Liqid removes the guesswork by pooling resources and spinning up GPU quantity to match the model.
At the AMD show, Liqid announced it can scale a single server to 30 AMD GPUs. The benefit, Puri said, is that large frontier models can run on a single server rather than a cluster.
"Now we come in and say we can take those models that were previously clustered and consolidate them down to a single server, a single instance, back to that tokenomic story," Puri said. "That's how we reduce the cost of deploying these models."
The orchestration layer that makes this work is Kubernetes, explained Puri, who pointed to the Nutanix Kubernetes Platform (NKP) as the mechanism enterprises use to deploy and manage those models, and said Liqid integrates closely with it.
"If we can implement technologies that drive utilization to 100%, we can do more with the same amount of infrastructure," Puri said.
Liqid does this by setting policies that move GPUs to where they are needed. When a workload uses its GPUs only 30% of the time, the software shifts them elsewhere to raise utilization. Puri said this works with brownfield environments, so IT teams can use existing servers and avoid challenges of purchasing new hardware.
"We're giving new life, new capability to the infrastructure that customers already own," Puri said.
Containerized Inference in Two Minutes
How Can Enterprises Simplify and Speed Deployment of AI Models?
By leveraging Kubernetes orchestration, enterprises can automate the deployment of containerized inference microservices in just two minutes, ensuring hardware resources are instantly allocated and returned to the pool once tasks are complete, explained Puri.
Inference runs in containerized environments, Puri continued. NVIDIA created NVIDIA Inference Microservices (NIMS) and AMD created AMD Inference Microservices (AIMS) because the complexity of kernel, operating system and library revisions is too much for average enterprises to debug. This can become a big, continuous problem for IT teams to wrangle.
Liqid built a Kubernetes plugin that automates the entire deployment. A user asks for a model like Llama 7B. The software pulls the container, determines it needs two GPUs, places them on a server and launches the instance in two minutes. When the workload finishes, the GPUs return to a free pool for the next model.
How Does CXL Memory Pooling Optimize Data Centers?
DRAM pooling enables the dynamic allocation of memory by workload, effectively ending the practice of over-provisioning servers and allowing data centers to reduce costs by precisely matching memory to application needs. While Liqid started with GPU pooling, it now extends the model to memory. The company says it is the first to offer an end-to-end DRAM pooling solution built around Compute Express Link (CXL), letting the two most expensive resources in the data center be dynamically allocated by workload.
"The days of over-provisioning your server with excess memory, those days are done now," Puri said.
Liqid is in a growth phase, Puri said, with new CEO Rick Hagberg and fresh capital. When he tells customers the vision, the reaction is consistent.
"This is obviously the right answer of how this should be done," Puri said. "Why isn't everybody doing it this way? That's my mission as the founder here, to make sure eventually everybody is doing it this way."
Related:
Building Solid Enterprise AI Infrastructure Strategy
AI Shifts from LLMs to Inference
From Cloud Native to AI-Native: The Evolution of Enterprise Infrastructure
AI Trends in 2026: Finding the Right Compute Platform for Each Workload
Video transcript:
Sumit Puri, Founder & CTO, Liqid: Since we spoke to you guys last, one of the major trends that we're seeing is the enterprises are finally past the initial exploration phase. Now they're in the adoption and deployment phase of their journey of AI. And I think a lot of them are trying to figure out how exactly they are going to do this journey. And the first crossroads that they're up against is, am I going to take all of my AI capability and move it into the cloud? Or am I going to own the infrastructure on-prem or in a colo that's required to do AI? And what it all comes down to is tokenomics. At the end of the day, these companies have a limited amount of dollars and a limited amount of power in many cases that they can deploy. And what they're trying to figure out is how do I get the most amount of tokens out of that very, very scarce resource?And so those are the discussions that customers are having now. And the cost of these tokens is front in mind for them. If we notice what's happening in the industry, large organizations are now saying, "Hey, we're going to limit the amount of tokens that you as an engineer inside of our organization can consume because the cost of that is becoming very high." And so the one great way to address that increasing cost is to own your own token machine. And so that's one of the conversations that we are having with our customers is one way to reduce the cost of those tokens is bring the infrastructure, bring the GPUs on-prem or in a colo so it's a one-time expense and you can consume all the tokens that you want out of that investment that you make. It is about the model that you are looking to deploy and you're going to, in any environment, you're going to have a variety of models.
You're going to have big models, you're going to have small models, you're going to have medium sized models. The model actually, the size of the model will dictate the type of infrastructure that you need. If I'm running a very, very large model, I'm going to need a large quantity of GPUs in order to run that model. And so you have to figure out if I'm going to build a very big system, how do I get those large quantity of GPUs into play? One way to do it is independent scaling of resources. We come in and let customers say, "I don't want to scale my compute as I'm scaling my GPUs. Allow me to just scale GPUs." As an example, today we're at the AMD show. The announcement that we're making today is our ability to take a single server and scale up to 30 AMD GPUs to that single machine.
And the benefit of that is we can now start to run these very large frontier models on a single server infrastructure. These models are so big. Previously you had to run them in a cluster. Now we come in and say we can take those models that were previously clustered and consolidate them down to a single server, a single instance, back to that tokenomic story. That's how we reduce the cost of deploying these models. Models will change over time. And so some cases you might want a very small model that will require, let's say, a single GPU. Sometimes you will want a medium sized model that will require eight. And sometimes you will require a massively large frontier model, which will require dozens. You as a customer, it's impossible for you to know ahead of time what size model I will need when. And so our vision is let's take the guesswork out of that.
Let's have a pool of servers, a pool of GPUs, and dynamically spin up GPU quantity to server based upon models. And the reality is the way that'll be done is something called Kubernetes orchestration will be the way that they do it. Nutanix has a phenomenal way of going off and doing that through their NKP platform. We're a big believer in NKP. And so we integrate closely with NKP. And so we say, let's allow people to deploy these models, the variety of models that are supported with the perfect set of infrastructure so we can have this matching of infrastructure to model to not have any wasted resources, improve the tokenomics. In a resource constrained environment, utilization is critical. You cannot afford to have your GPUs at 30, 40% utilization because you can't get enough. And so if we can implement technologies that drive utilization to 100%, we can do more with the same amount of infrastructure.
So that becomes critically important and we do a lot of that today. We look at an environment and we say, "Hey, listen, this workload is only using those GPUs 30% of the time. Let's set up policies that allow you to move those GPUs to different parts of the infrastructure to raise the utilization." The second thing is Brownfield. We support brownfield environments. So imagine an environment where I can't buy servers because memory's limited, CPUs are limited, servers are very expensive. We come in with our solution, which is disaggregated pools of GPUs, and we attach them to existing infrastructure. And so now we're giving new life, new capability to the infrastructure that customers already own. As inference becomes a thing in places like the enterprise, it will only be done in containerized environments. NVIDIA has created something called NIMS, NVIDIA Inference Microservices. AMD has created something called AIMS, AMD Inference Microservices.
And the reason is both of these companies realize the complexity related to kernel revisions, operating system revisions, library revisions is too complex for average enterprise customers to debug and solve, and it slows the rate of AI deployment. So if we can containerize all of these models, put them into a store, allow the enterprise customer to bring whatever model he wants, a vision model, a speech model, a text model, bring it down in whatever size of model that they need, 7B, 70B, 405B, and make all of these available containers that the customer can deploy and implement immediately, that's how we're going to accelerate the pace of deploying AI. The way that we approach that is we have a plugin for Kubernetes that when you say, "Hey, give me Llama 7B as an example," we'll take the container, we'll bring it down, we'll crack open the container, we'll determine the exact amount of physical resources that that container needs.Well, this needs two GPUs and we want to put it on server number four. We'll pause that container, we'll go into the background infrastructure, we'll put two GPUs on server number four, and then we'll take that container and we'll launch it onto that server. We automate the entire process of deploying models down to a two minute instance. So you say, "Give me Llama 7B." Two minutes later, you're speaking to a chatbot with the exact amount of hardware resources required for that instance. And the moment that that server doesn't need that GPU anymore, we can delete that instance, remove those GPUs, put them back into a free pool so your next model that comes along always has a pool of GPUs to draw from. Liquid started, the way we started was around GPU pooling. That was the first thing that we went off and we accelerated and we composed and we pooled and we shared.
Now the next thing that we have just announced is memory pooling. So we are the first company to provide an end-to-end solution around DRAM pooling built around CXL. So now I can have my pools of GPUs and now I can have my pool of DRAM. The two most expensive things in the data center can now be dynamically allocated by workload. The days of over-provisioning your server with excess memory, those days are done now. And so we must find methods to precisely provision the exact amount of memory the application needs so we can reduce our memory costs inside the data center. And by doing that, it's back to that tokenomics. Yeah. Well, the company is surely in its growth phase right now. I'm a technologist at heart. I'm super lucky to have built out an amazing team. We're very lucky to have a new CEO on the team.
A gentleman named Rick Hagberg, who's just an industry veteran who's been phenomenal. We're raising additional capital. We're growing the team. We're bringing new executives on board. We are having our moment right now and we're super excited and blessed to be part of this industry. The way I feel about it is when we go around and we tell customers about our vision, which is stop putting static resources inside of a server. Let's take all of the GPUs, pull them inside of the rack, add a push of a button, dynamically allocate any GPU to any server that you want. You won't lose any performance. You'll scale up as big as you want. We're a driverless solution. We can drive to 100% utilization. When we tell this story, no one ever says, "Well, that's a terrible idea." That's not the reaction that we get, right? The reaction we get is, "Wow, this is incredible. This is obvious. Obvious is the one that we get more than anything else. This is obviously the right answer of how this should be done. Why isn't everybody doing it this way?" And me, that's my mission as the founder here is to make sure eventually everybody is doing it this way.
Ken Kaplan is Editor in Chief for The Forecast by Nutanix. Find him on X @kenekaplan and LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Related Articles
