FinOps X 2026 | Pravir Gupta, Google Cloud

Clips
More from FinOps X 2026

Pravir Gupta

VP

Google Cloud

Google Cloud saves $30 million by applying AI to its own operations

As FinOps AI strategies continue to emerge, the familiar cloud cost management approach is breaking down — and organizations that fail to adapt risk runaway spending on workloads they barely understand.FinOps is rapidly evolving from a cloud-billing function into a strategic framework for governing the full technology stack, including AI, software-as-a-service and now autonomous agents. According to the “State of FinOps 2026 Report,” 98% of practitioners now manage AI spend — yet most organizations still lack the cost granularity needed to govern it effectively, according to Pravir Gupta (pictured), vice president and general manager of Google Cloud at Google LLC

play_circle_outline FinOps for AI: Implementing Budgets, Alerts, Spend Caps, Anomaly Detection, and Pre‑Deployment Cost Estimation

play_circle_outline FinOps for AI: tokenomics plus adjacent costs like VMs, storage, KV caches

play_circle_outline Google on Google AI: invoice reconciliation agents yielded 4x throughput, $30M savings

play_circle_outline Metrics: adoption, efficiency vs transformation, scopes (individual, organizational, product), revenue/productivity gains

play_circle_outline Agentic Workers: Cost Modeling, IO Tokens, Mutable Actions, Orchestration, Identity, and Audit Trails

play_circle_outline Cost Granularity, Explainability, and Attribution in Agent-Driven Workflows: Using Tags and Labels for Enterprise Chargeback

Info
Transcript

Pravir Gupta, Google Cloud

Pravir Gupta

VP Google Cloud

In this interview from FinOps X 2026, Pravir Gupta, vice president of Google Cloud, joins theCUBE's John Furrier and theCUBE Research's Paul Nashawaty to discuss the evolution of FinOps from cloud cost management into an enterprise-wide discipline for governing AI value. Gupta outlines Google Cloud's FinOps vision — ensuring costs are visible and allocated, continuously optimized through committed use discounts and properly managed through budgets, alerts and anomaly detection. He argues that FinOps for AI extends well beyond tokenomics: an agent completing a... Read more

explore Keep Exploring

What does the future of FinOps look like from a cloud service provider’s perspective? add

What should organizations focus on to manage the total cost of ownership (TCO) and operational costs of AI deployments beyond just tokenomics? add

How is Google Cloud using generative AI internally (the "Google on Google AI" effort), what are the main goals for these transformations, and can you provide a concrete example with measurable outcomes? add

How should organizations approach, evaluate, and prioritize Gen AI initiatives — what metrics, scopes, and outcomes should they consider? add

How will AI "digital worker" orchestrator agents drive business transformation by initiating workflows, coordinating sub‑agents, and involving humans for approvals (for example, in a quote‑to‑contract process)? add

What steps should an organization take to identify the potential revenue or efficiency uplift from a generative AI transformation and to attribute costs and usage across its business divisions? add

bolt Powered by CUBE AI

Pravir Gupta, Google Cloud

search

John Furrier

>> Welcome back over to theCUBE's live stream here in San Diego, FinOps X 2026. I'm John Furrier, your host of theCUBE, with Paul Nashawaty, my co-host with theCUBE Research. This next segment, we're going to unpack the future of FinOps with Google Cloud, Pravir Gupta, VP of Google Cloud. He's also internally called the Google Cloud Business Platform, which is basically the CIO. Pravir, great to see you. Thanks for coming on theCUBE.

Pravir Gupta

>> No, thank you, John.

John Furrier

>> All the way from Mountain View-

Pravir Gupta

>> Thank you, Paul....

John Furrier

>> you come down from, San Diego.

Pravir Gupta

>> Yeah, thank you for having me on the show.

John Furrier

>> As I said on the intro, the future of FinOps is bright. Got to wear shades, as they say, we're in San Diego. So it's very bright. It's changing, because this community is building the platforms and the tools to run in a full distributed computing hybrid environment, with AI coming onboard superfast. Cloud is still accelerating, but AI on top, you put those two together, explosive opportunities. And you're seeing the results. We're seeing revenue generation, not just cost takeout, with AI. And you're starting to see teams start to understand, "If we don't manage this, it's going to be out of control." So we're kind of in a managed chaos state. But unlike cloud, this is a business model transformation for companies. This is not an IT project. So I want to get your thoughts. What is the future of FinOps, from your perspective?

Pravir Gupta

>> You see, John, as a cloud service provider, the product area of FinOps is very important for us. Our vision for FinOps is to make sure that all your costs are visible and allocated, they are always optimized, and properly managed. What I mean by visible and allocated is that you have the right cost granularity, so that you can identify, and easily identify all your cost drivers. What I mean by always optimized, is that most of your spend should be covered by committed use discounts and cards. And what I mean by properly managed, is that you should have the right budgets, alerts, spend caps and anomaly detection. If I take a step back and zoom out a little bit, enterprises have been using cloud to move fast, innovate, scale and maximize value. But a big OpEx item for all the enterprises is their cloud cost.

John Furrier

>> Yeah.

Pravir Gupta

>> So managing that cloud cost is very important for every enterprise, whether it's small or big. Many enterprises, and many customers, are always hesitant to use new technology. Because they don't understand what would be the cost of using that technology, and how to manage that cost. We saw this when enterprises moved from on-prem to cloud, and we are seeing a similar trend when enterprises are looking to use gen AI for transformation. And that transformation is for either how they operate, to increase productivity and efficiency, or to improve their products. So the same trend will continue and every CEO is asking their teams, the whole organization, to innovate fast with gen AI. And that's where the FinOps is still very relevant to make sure that you have the right guardrails to better estimate the cost, to have explainability of those costs, as well as the guardrails that allow you to innovate faster.

John Furrier

>> Talk about the trend around cloud transition to AI, because that's really the topic here.

Pravir Gupta

>> Right.

John Furrier

>> Okay. You guys have done a great job, because you have years under your belt doing this.

Pravir Gupta

>> Yes.

John Furrier

>> Providing the kind of telemetry, kind of the data granularity, for the tools to work. Okay. That continues to progress. AI now is on the scene. What's different now? What should people be thinking about? Obviously, maintain the cloud, because it's all cloud spend now.

Pravir Gupta

>> Yes.

John Furrier

>> It's like, tokenomics is basically cloud plus tokenomics, is cloud spend.

Pravir Gupta

>> Yes.

John Furrier

>> It's an aggregate. What's your view on how people should think about that transition? How should they land, what's changed?

Pravir Gupta

>> I would say the big area to focus on is what I call FinOps for AI. Tokenomics is a big piece of that, but not the complete answer. And why I say that is, if you think about the TCO, or the total cost of ownership. If an agent is doing some task, yes, there are input tokens and output tokens. But the agent may spin up a VM, in a sandbox, to actually write scripts and do things. You may also have what you call adjacent AI costs from your KV, or the key value cache, that's there. How do you load in and progressively load in the context that is needed? What storage system has your context? How do you do RAG? All of those are additional adjacent costs outside just the input and output token. So tokenomics is a large piece of the FinOps for AI, but the real thing for every enterprise to focus on within the FinOps area, is the FinOps for AI.

John Furrier

>> It's like the iceberg-

Pravir Gupta

>> Exactly....

John Furrier

>> what's under the water. We see the cost because everyone's using the tools, as users. Adoption's certainly high. We see the numbers and the graph and the keynote. It's under the water.

Paul Nashawaty

>> Yes. Yeah. I actually think this is an important conversation. You talked about trends, you talked about where the customers are going, what they're asking for. I want to know what Google's doing. What is Google doing, what your business processes are doing? How are you changing this with Gemini? You're kind of a drink-your-own-champagne company. You want to execute something, but you want to try it internally first. And I want to know what you're doing, because I think that is your own best use case.

Pravir Gupta

>> So my role within Google Cloud is, I lead the Google Cloud business platform, and this is the CIO role within Google Cloud. As part of that, we end up being customer zero of trying all the different models and the platforms to innovate and transform how we run Google Cloud internally. Because we need to scale up, we need to transform how we run the business. And as part of that, we have been on this journey, and internally we call it Google on Google AI. And this is about how we do the business transformation of running our internal operations and systems using AI. Anytime we have wanted to transform with AI, we look at the opportunities that GenAI provides to us in three ways. First is about growing revenue with faster time to market, improved customer experience. Second is about increase in productivity, as measured by increasing efficiency. And third is about providing a competitive differentiation in everything you do. So any transformation we do should have one of these three benefits. And as part of that, we have done transformations in different business domains. Whether you look at marketing, sales, customer experience, all the way to back office in finance as well. Happy to also give an example. If I talk about finance, because the first thing that comes to people's mind is, "Oh, I can do something in marketing and customer support. Those are easy ones." And then people say, "No, no, don't touch finance, because I need 100% accuracy." So let me talk about an example that we have done in the finance area. And this was not just for Google Cloud, it was for all of Alphabet. So we have lots of suppliers who give us invoices that needs to be reconciled against the contracts that we have with the suppliers, before we pay them out. Earlier, we had a team of humans that would look at the incoming invoice, look up against the contract. If a discrepancy is found, they would initiate a contact with the vendor. If no discrepancy was found, they would do a payout to the vendor. So we had this aha moment that, instead of humans doing the task, let's have agents do the task, and shift the roles. And we have humans now review the task, right? So humans shift from doing, to actually reviewing the output of the agents. As part of that, we saw a 4X increase in throughput capacity, and a $30 million savings across all of Alphabet, which is hard actual measurable dollar amount-

John Furrier

>> Real cash....

Pravir Gupta

>> that we are able to see. That's actual cash.

Paul Nashawaty

>> So TK is happy then.

Pravir Gupta

>> Not just TK, this is Anat and Ruth, because this is Alphabet-wide.

John Furrier

>> It's Alphabet. Yeah, yeah.

Pravir Gupta

>> It's Alphabet-wide.

Paul Nashawaty

>> No, it's fantastic.

Pravir Gupta

>> But that same pattern applies in so many different ways, because the trick here was not to roll out with 100% accuracy. The trick is that you have human in the loop in the middle, where humans are reviewing the output of the agent and then providing the feedback.

Paul Nashawaty

>> So let's talk about the output, because I like to hear, that's really awesome. That use case really is powerful. You talked about some numbers there, which is very powerful. What metrics, for yourselves, but also what would you advise others to say, what metrics would you look at in order to say, "Is this a success? Is it not a success?" What are the things that move the needle?

Pravir Gupta

>> I think the first metric is, are you actually using AI, or are you sitting on the sidelines? That's the given. There are a lot of people who are sitting on the sidelines and still watching. If you're that, you'll have innovators dilemma, and that's not going to be a good recipe in the future. So the best way is to actually roll up your sleeves across the organization, where everyone is becoming a builder, and actually getting in and doing things. Second, I would say is, there are this analogy of those flowers blooming. Do you light up and support all of those efforts, but you can't just have thousand flowers blooming. How do you convert and curate them into these curated bouquets, that you say, "These are the blessed projects that we will really go after"? Any Gen AI transformation has three scopes. It is an individual scope, where you're improving the work that an individual does, an organization scope, where you're improving the workflow done by an organization, or you're improving the product. So individual, organization, or product. In either of these three scopes, you can have one of the two outcomes. It's either an efficiency outcome, which is improving the thing that you are doing, and you can measure, "Oh, 30% time saving or cost saving," or it's a transformation. So those are the only two outputs that you would have, efficiency or transformation. So as an example, in the individual scope, if you provide gen AI based suggestions and auto complete, that's just efficiency. Because now, developers who are writing code, they can just try better and faster code. But with vibe coding, non-developers can suddenly produce code. So you have democratized coding, so that's a transformation. So you have to look at from that matrix. So that's the second piece I would say. And the third is, really around any gen AI transformation you do should have the benefit of either it's growing revenue, it's providing a competitive differentiation, or it's increasing productivity. Or more than one of those benefits. So weighing in either of those benefits, to really say which are the real transformations that you'll go after. And you have to take some bigger bets in terms of making those transformations, and fail fast.

John Furrier

>> I want to tie this topic to the new foundation, Tokenomics Foundation, and actually the FinOps, as you point out. Cloud, super relevant, it's all cloud spend.

Pravir Gupta

>> Yes.

John Furrier

>> We were talking before you came on camera, here on the floor, about coding and examples versus token usage. And like, if I was a coder, I write long lines of code. "Wow, he's a good coder. He writes a lot of code." Versus, back in the old days, tight code, right? Now tokens are similar. You were sharing your thoughts on it. This is very important because, token efficiency, waste is a big topic here.

Pravir Gupta

>> Yes.

John Furrier

>> It's not about how many lines of code, or token maxing. I'm on the leader. "Look at me. I'm using a lot of tokens. I'm good." No, no, no, no. That's a false metric.

Pravir Gupta

>> Yes.

John Furrier

>> Explain. This is super, I think, relevant.

Pravir Gupta

>> Yeah, yeah.

John Furrier

>> And maybe an ongoing metric.

Pravir Gupta

>> So some companies had started, I would say in the last six months, to put up these leader boards with the whole notion of token maxing. And saying, "Look at the number of tokens being used and consumed in my organizations," as a proxy to say, "We are actually efficient in producing a lot of gen AI transformation." But it's a fallacy, because if I rewind back to even five years, 10 years, you can make the same argument around the number of lines of code produced, or the number of change lists that are produced. If you count developer productivity by the number of change lists produced, you have just given the wrong incentive for the developer. They would say, "Instead of my one change list, I'll break that up into five change lists, and ship those out." That doesn't make sense. Or, "I'll write long lines of code, which is less efficient code." Or, "I'll fork the code, and you'll see all the lines copied again." So it provides the wrong incentive.

John Furrier

>> Yeah, you get a good evaluation. "Well, you hit all your metrics."

Pravir Gupta

>> Yes, wrong metrics.

John Furrier

>> Reverse metrics, right.

Pravir Gupta

>> It doesn't measure the business outcome.

John Furrier

>> Okay. So let's take that to the next level. So you were talking earlier about adjacent costs.

Pravir Gupta

>> Yes.

John Furrier

>> Okay. That's humans that are triggering action, and there's some KV cache, you mentioned some of the KV cache things. Now with agents, they're not human trigger. So you're almost cascading the costs.

Pravir Gupta

>> Yes.

John Furrier

>> So now you have machines with agents. Agents are doing work. So take that metaphor to the next level. What is the metric for agents? How do you look at that? Because there's also adjacent costs to agentic, non-human triggered costs.

Pravir Gupta

>> Yes, yes.

John Furrier

>> What's your thoughts on that?

Pravir Gupta

>> And this is one area that we'll see more and more of. Today, a lot of transformation that we see are human initiated, that use an agent to get something done. Like in a marketing case, I'll drop in a campaign brief, and the agent produces all the assets to run my campaign. The shift that we'll start to see more and more, and we are starting to see that within Google, is this notion of your digital worker, your assistant. And those agents will often run as the orchestrator agent in a headless manner, and they are initiating the overall transformation. And a key piece of that is that agent is an orchestrator agent, which in turn will call sub-agents to do the tasks, as well as bring in the human where approval is needed. For example, we are building this, what we call a quote to contract agent. Every company that goes through the sales cycle knows the end number of conversations that need to happen in the quote to contract cycle to close that as quickly as possible. We are kind of flipping it on its head by saying the agent does the orchestration, and in that flow it may produce, "Here's an option A for a pricing model, here's option B." And the pricing team will need to validate those. As the agents do more and more of this work, and recently we announced also Spark, which is your personal agent that is in Workspace and will be rolled out to Gemini, Workspace and across the board, and having this notion of a digital worker. Once you set up this agent that also works on your behalf as a digital worker, identity becomes very important, because you need auditability and tracing on that. You need the end user credential, as well as the agent's own identity, when it does any mutable actions. Because it's not only about reading, it's about also mutating with crowd operations. So now you have to look at the overall cost of when an agent operates with sub-agents, and that's where the cost granularity comes very important, because I need to know not just the cost of the orchestrator agent, but the sub-cost of each of the underlying agents. As well as the cost by input-output token, the cost by different models, because each agent may not need to use the same model. Because if it's running a smaller task, it can use a Flash, a workhorse model, whereas a more complex task can use your sort of frontier models, and you have to balance that. So the cost granularity becomes really important, so that you have cost explainability of that overall cost.

Paul Nashawaty

>> Yeah, that makes a lot of sense. I think there was one question that kind of comes to mind. As you were talking, I was following, I was tracking. I was like, "Yeah, that makes a lot of sense." Right now it's kind of a big kind of overview of like, "Here's what the cost is," and it's not granular enough. It's not about even chargeback models, just to understand what's happening within your organization.

Pravir Gupta

>> Yes.

Paul Nashawaty

>> But let me ask you this question, because this event has been pretty dynamic, there's a lot of things happening, a lot of conversations going on. What's not being talked about that should be talked about? Because that's something that a lot of people don't even know.

Pravir Gupta

>> I think one piece that's being... Let me start with what is being talked about a lot, and a little bit of the trends that we see, and I'll talk about then what is not being talked about-

Paul Nashawaty

>> Sure....

Pravir Gupta

>> right? The three trends that we see a lot is AI cost explainability, better understanding of the cost and the cost granularity. AI cost guardrails, prevent me from having unexpected runaway spends. And AI cost estimation. Before I scale up my experiment or a prototype into production, help me better estimate the cost. So a lot of conversations on these three topics, and as they relate to tokenomics, and FinOps for AI, so lots of conversations on those. What is not being talked about is that all of these will not be human initiated. When agents start to do more transformative work, and that's where we are only seeing few enterprises out there that are at the bleeding edge. And within Google Cloud, we are trying to do a lot more internally as well, from large transformations. And remember, it's not just smaller efficiency plays, because those you can get. When you think about the large transformations, then it's the digital agent or the orchestrator agent that's running in a headless manner, and doing the overall orchestration. What does that mean? Because, let me give a productivity use case, which is a collaboration one that everyone can relate to. Imagine I'm getting a lot of emails. I'm traveling, or I don't have time. The agent, in a headless way, is able to look at all the important emails, create a draft of the response that I should have for all those emails, and have them ready. What is the cost of... The agent initiated it. I didn't ask it. Anytime an incoming email comes in, it decides if it's important or not. It generates the draft that is ready to go, that I can review, tweak if needed, change the tone and send. What is the overall cost model of managing those complex agents which are your digital workers, which are orchestrating the overall workflow? I think a lot of conversation needs to happen on those types of agents, because that's where I think the puck is going.

John Furrier

>> And you guys are doing that. I see my Gmail and Workspace is doing that, thank you very much. I edit it, sometimes I don't. But I want to share something with you and I want to get your reaction. So what I hear being talked about here, at a very elementary, is tagging. But what's not being talked about is how that ties to the revenue. And so, I want to get your thoughts because, not to date myself but back in the early internet days when routing, Cisco emerged, you had least path routing. So high priority policies were invented for high yield activities. "Here's the shortest path, least cost." So premium stuff went shortest path, low end stuff kind of went cost-wise. Apply that to workloads, and now agents and tokens can be applied to the same, because where FinOps I think is going to the next level, this is probably next year, dot connecting. Is, "Okay, great. This task is driving revenue. This person, this organization, their token usage, good code, good token usage, is driving revenue. They're writing marketing copy, which is a little bit elusive, abstract, but also contributes to productivity, which drives revenue downstream. Cool." That piece around revenue contribution, what's your thoughts on that? Because I think this is really going to be big.

Pravir Gupta

>> I think you have to do it in two steps. The first is identification of, what is the potential uplift, or the existing toil, even before you do a gen AI transformation. Because that helps to identify which areas you should do the gen AI transformation. So early on, that identification to say, "Here is the overall revenue lift that we expect," because you always want to go in with the hypothesis, and not after the fact and say, "This is the revenue." Because you may need to do a time motion study to figure out either the efficiency gain from the humans who are doing the task, or the revenue uplift that if the product does something, that you can charge more. So that's the first piece of identification upfront. The second is, now as the agents are doing certain tasks, the labels and the tags that you want to use. It's important for a few aspects. First is, even if when the CSPs and Google Cloud, we provide the cost granularity, that can give you the granularity at all the different dimensions. But what it will always miss is the enterprise who's using those tokens. The enterprise may have their own CRM system, where their business divisions are organized in certain ways. How do you allocate that and attribute that this usage belongs to this division, and this usage belongs to this other division? That's where tags and labels come in, in the sense that you can now do the attribution, which then the FinOps teams in those enterprises can do a chargeback and show back to those business divisions.

John Furrier

>> So that's the integration of what you guys provide, with the teams.

Pravir Gupta

>> Yes.

John Furrier

>> Okay, great. Awesome.

Pravir Gupta

>> And with the tags that then represent the business schema in the organization, that's on the other side of the experience.

John Furrier

>> And they have the business logic. They can see that-

Pravir Gupta

>> Exactly....

John Furrier

>> and they match that up.

Pravir Gupta

>> Exactly. And then you can set up guardrails on those as well that, "For usage that has attribution of this tag, I want alert, I want guardrails," right? It allows all of those controls that you want to prevent that unexpected or other things.

John Furrier

>> Yeah. And that explains why the agentic side, not to change gears, but the projects are focused on the biggest problems, because they drive revenue.

Pravir Gupta

>> Yes.

John Furrier

>> And so people are taking a different approach than the old IT, it was, "Hey, we'll do it sandbox, and test this use case." Here, it's like, "Let's go with the revenue."

Pravir Gupta

>> Exactly.

John Furrier

>> And they tag it up, and they learn. All right, great conversation. Great to have you on. I have to ask one final question. As AI builders is now a persona, generically, call it AI builders. That's an engineer, that could be a business user, AI is everywhere. How does that change the tooling? There's a lot of tools. Is it going to look like a consolidation, and will there be a platform approach? How do you see customers thinking about this, from a Google perspective and from their perspective? Because they want the value.

Pravir Gupta

>> I think first, for builders, we are seeing exactly that. Some people call it builders, some people call it technologists. Within Google, we are seeing a merge of these role profiles as well. It's not just your engineers, but your UX designers, your product managers, your program managers. They all become builders, that they can easily do the end to end, and do the thing on their left and the right. Which is, a product manager can easily design the UI. The UX person can easily convert that into both the code, and sometimes also produce the artifacts that are still needed. See, the overall SDLC, the software development lifecycle will change. Internally, we have started to call it, whether you call it ACDC, or there's another, NAIL, which is Native AI Lifecycle. There are all sorts of acronyms that are popping up. Where, while you change the whole model from a waterfall to this agentic way of driving it, for different sorts of regulations, you'll also produce the artifacts. So in many of the transformations and the way that we are producing, we have our product managers, and the UX, who are creating the prototype first. From their prototype, they will create the UX mark, they'll create the PRD from the prototype. Because remember, it's not that you needed the PRD first, but the PRD is still needed as an artifact for all sorts of SOX compliant, and all the regulatory audits that you need. So you're producing that paper trail and audits as well. How that consolidation happens is through a few sets of tools that allow you to do the transformation across the different set of underlying tools. So we are seeing on Google Gemini Enterprise and Google Antigravity, as the two applications upstream. One for coding and development, and the other for interfacing with all the underlying applications. So you don't need to go in an underlying SaaS application to actually configure a quote, to add a tag or a comment. You can just use it through Gemini Enterprise, and it will not only do the reads, but also provide the cred operations under all the fragmented SaaS applications that you have. And that interface is available not just as a standalone, but also where you are, because you have to meet the user where you are. So you may be in your workspace in Gmail, with the personal agent, the Spark agent that talks to Gemini Enterprise. It just executes the command, and it creates that task.

John Furrier

>> It's a great time to be an engineer, technologist, and a builder. Pravir, thanks for the insights. Really appreciate you coming on theCUBE here at FinOps X. Great time.

Pravir Gupta

>> No, thank you.

John Furrier

>> Thanks for sharing.

Paul Nashawaty

>> Thank you. All right.

Pravir Gupta

>> Thank you for having me, Paul and John. Thank you.

John Furrier

>> You. I'm John Furrier with Paul Nashawaty here, getting all the data, sharing what people are doing, and really kind of where the puck is going and connecting the dots, here in theCUBE. Thanks for watching.