The New FinOps Problem Isn't Cloud Bills  🔗 **Source:** [The New FinOps Problem Isn't Cloud Bills](https://share.snipd.com/episode/75b5bedd-5cc7-483f-98e8-bfb1980b1afc) ## Highlights - **FinOps Is About Value Not Just Cost** - FinOps is primarily about extracting business value, not just cutting cloud costs. - Roi Ravhon frames it as measuring ROI and reallocating investment toward high-value cloud, SaaS, on-prem, and AI projects. Transcript: Pathik Sharma Few times. So pretend I'm five and you have to explain FinOps to me. I think contrary to what many may think, FinOps is about saving money and optimization and cost reduction. I believe it's truly about like making more money for business. At the end of the day, it's about finding the value of your investment in cloud. So it all started with cloud, and then the scopes keep on increasing because there is SaaS, there is also on-prem, there is AI that comes into the mix as well. So I think at the end of the day, the idea here is, hey, are we investing in the right things or not? If we are, then how do we double down on it? If not, then how do we optimize it, be efficient, things of those things, right? So I think that's preliminary the basics of what Phenops is in the discipline. Yeah, build a discipline around these ideas that maybe wasn't available. Frédéric Lardinois People didn't think about it necessarily, right, in the early days of cloud yet. Yes, I think it's a proliferation of like, you have procurement at the edge. Pathik Sharma Anyone can create instances and services within seconds and minutes. And so now, how do you put your arms around that? Like, how do you put guardrails around that? So then we start to go into more of allocating costs by individual teams and driving accountability and having governance, guardrails, and optimization. ([Time 0:01:10](https://share.snipd.com/snip/adae27bd-88a3-46b8-985f-ad8090fc2d72)) - **AI Automates Repetitive FinOps Work** - AI enables automation of repetitive FinOps tasks previously solved by headcount. - Roi Ravhon highlights automating anomaly triage, ownership allocation, and right-sizing recommendations across thousands of teams. Transcript: Roi Ravhon So we used to solve problems with headcount, and now we have to start solving problems with automations. And FinOps is all about small problems that you need to fix. And just imagine a large enterprise, I have a thousand different development teams, and I keep on getting right-sizing recommendations and anomalies for a bunch of redundant stuff That I don't even care about, but it's all add up to a lot of money. And creating that culture, like Phenoms was always about culture. How can we get the engineers to actually care enough so they can start and create the optimization themselves? But now we have the ability to start to solve problems without a person actually changing stuff? Or how can we start to allocate in a smarter way? How can we find the actual person responsible for that resource? I think AI just started to open a whole bunch of new agentic possibilities in FinOps that will allow us to automate the dumb stuff and focus a lot more on the actual film of work on delivering The value of understanding what the business have to do with that with everything and not having to spend 80% of our days on like triaging tickets and understand who should who should Own what ([Time 0:03:34](https://share.snipd.com/snip/ce2b0260-a05b-4af4-bbbe-8be440c548b1)) - **Have AI Prepare Actions But Require Human Approval** - Use AI to analyze, synthesize, and create org-aware recommendations but keep human approvals for final actions. - Pathik Sharma recommends AI produce pull requests or tickets that app leads then approve or deny. Transcript: Roi Ravhon As well? Pathik Sharma Yeah, this is personally one of my energy creator, like my energy flow in this direction about, at the end of the day, like FinOps is a data science problem. You have data, you need to find insights, you need to find recommendations and then take action on it. And I think to what Roy was mentioning, like EA can actually help us do this at scale uh funny enough two years ago i was having a conversation about the intersection of ai and finops overall And the conversation was sort of brushed off it's like yeah ways away from that. You know, AI, not everyone is using AI. It's a very tiny piece of what we are experimenting. Two years later. That's why Google was behind on AI. No, I'm just kidding. Two years later, literally the top two out of three things from FinOps Foundation 2026 survey data is like, AI or FinOps? FinOps or AI? Roi Ravhon How do we get that? Yeah, 98% of FinOps teams are now managing AI. Frédéric Lardinois Yeah. Yeah, which makes perfect sense at this point, for sure. Yeah, for sure. Pathik Sharma And I think the other aspect is also, like, how do we drive action to your point, right? Like, it's, this is not rocket science but this is also you need many teams to take small actions to drive a larger impact and oftentimes fin ops might seem as it's not the priority for Me. I'm building new products, I'm trying to create more differentiation, but then I'm also trying to do phenops. How do I balance all that? And I think that's where AI actually plays a very important role of doing all the groundwork. Doing analysis, synthesis, insights, recommendations that are very organization aware. It is looking at my specific billing data, looking at my contractual agreements with cloud service provider. It's finding the right sizing according to my definition, because the definition of right sizing for app one versus app two could be very different. So like it to account for that. And I think that's where we see very interesting shift of like, AI is just like your team member doing some helpful work for you. Exactly. And when we show this demo to our customers of like AI coming in, finding insights and recommendations and actually creating a pull request in your repo and then sending it to the application Lead teams to either approve or deny. That's essentially what a normal team member would have done, essentially. So like that's how AI is able to take a lot of groundwork and be able to, you know. ([Time 0:04:56](https://share.snipd.com/snip/5cce460b-2151-4090-9f4c-71405440a13a)) - **Blend Deterministic Rules With LLM Enrichment** - Combine deterministic building blocks with LLM enrichment to avoid hallucinations. - Roi Ravhon advises using deterministic detection and observability signals, then letting agents enrich context before human review. Transcript: Roi Ravhon I think it's one of the challenges that the industry as a whole is starting to face right now. And it's a hard balance. How are you balancing it? So when we think about the agentic thin-ups flows, we're thinking about building blocks kind of fact nature. So if you want to, you can just throw away an LLM on a problem and ask you to solve it because they will solve it. But not necessarily the way that you want it. So you need to give it deterministic blocks that they can start to play around with it and then start to stitch everything in the middle. So if you want detection detection is deterministic like don't try to reinvent the wheel but if you want enrichment you want to understand the context and want to trash so like this is Great great workflows for upgrade for agents and you can use all that knowledge you can use everything you have and and build a lot more and then if i'm a second before i'm i don know, terminating A server because this is what I thought that I need to do, I want to make sure that the LLM is not hallucinating, right? Like I'm not going to just fire it in my environment. So again, I need a deterministic layer or human approval for everything. ([Time 0:08:42](https://share.snipd.com/snip/58737fbe-3f0f-49de-a65f-e78af8fcc1dc)) - **Onboard Agents Like New Team Members** - Onboarding an AI agent should mirror onboarding a team member with context, auth, and guardrails. - Pathik Sharma ties this to SRE practices: use Goldilocks signals, observability, and P99/Pmax metrics for right-sizing recommendations. Transcript: Pathik Sharma There's onboarding. Here's our standards. Here's how we do things. Here's our context. Roi Ravhon Context is key, by the way. Pathik Sharma And that's, I think, the similar process that needs to happen when you onboard an AI agent. And similar to how your team member has authentication, authorization to certain data and certain tools, the same process has to go through for AI agents as well. And if I were to pick an example based on what you were saying, right, like, I'm just going to pick, like, Kubernetes engine, right, and ability to right-size your workload, bin packing, The way SRE have done this and evolved over a period of years, like, Goldilocks is one of the avenues. We have golden signals within GKE. And to Roy's point, you're not saying LLM like go fix my Kubernetes, you're actually enabling LLM to say that, hey, here are the Goldilocks signals, you look at requests, you look at Limits, you look at used, you look at observability metrics for last 30 days, you look at P99, DCP utilization, you look at Pmax of memory, you come up with a very deterministic way of How an SRE would have gone about right-sizing your workload. AI agents does that exactly much faster rate. And then, now you instantly build that trust. Like, hey, I know where this recommendation is coming from. I know it's contextual, how it works and how our SRE team would have done. ([Time 0:10:12](https://share.snipd.com/snip/47d91324-2f9b-4b50-afb6-b87a0c689829)) - **Token Prices Fall But Per-Task Costs Can Rise** - Token economics are volatile: cheaper model prices don't always mean lower cost per task. - Roi Ravhon notes newer frontier models can consume 3x tokens for same task, so usage pattern matters more than headline price. Transcript: Frédéric Lardinois Not just, you know, AI for FinOps, but also FinOps for AI, because more and more, I think people need to figure out ways to manage the cost of those AI agents, of the token usage that they're Incurring every day now. How are you thinking about that and how does that change the discipline overall? Roi Ravhon So I think as we now love to call it, the token economics in general started to be like one of the hottest buzzwords in like very, very fast. And the reason is that, you know, we adopted into cloud for a decade, give or take. And now we need to do the same thing we did for cloud to AI, but we're doing it in a year. So like we started from, yeah, AI is cool and it's probably going to help at some point to like the past like six months. You see the level of evolution of adoption of enterprise with actual AI systems and it's not just the interference of my production system it's cloud card it's cursor it's like a tons Of different different parts of AI that is now all channeling into I into the phenops teams and CFO started with unlimited budgets that just you know let's be innovative because being Innovative is super important right now. But now we're starting to circle back into, okay, it's fun to be innovative, but let's start to showcase ROI. And like, is what we're doing makes sense? And can it scale? And everyone is talking about the price of tokens are decreasing. But they're not really, because, you know, Antropic just released Opus 4.7, which is like the newest model as for this minute. And OpenAI today, just as we're recording this, announced GPT 5.5, so yes. Yeah, which is a great model, and priced the same, but it's thinking, I don't know, 3x is much. So for the same task, you're just spending more tokens now. So while the token economics are to be cheaper, how you use them is higher. And suddenly you get into, and getting back to the under-deterministic nature of running AI, you ask the same question twice and you get different token usage for everything. So how can this scale? So we just started to evolve into FinOps very fast. And luckily for us, I think FinOps framework covers the basics for AI. We don't need to reinvent everything we're doing. We still need to show back and charge back. We still need to negotiate better rates. We still need to optimize. We still need to get folks to care about how much is spending and putting it in a centralized manner and showcase the cost. And we have the tool belt. ([Time 0:13:46](https://share.snipd.com/snip/fc5e4488-6471-4e33-a7b2-65286b976abf)) - **Orchestrate Model Selection For Cost Efficiency** - Route requests to the cheapest model that meets requirements instead of defaulting to frontier models. - Pathik Sharma recommends an orchestrator layer to pick Flash versus Pro or edge Gemma for appropriate use cases. Transcript: Pathik Sharma Higher. I compare this to like Thor's hammer. You don't want Thor's hammer all the time. Yes I do. I may have touched up on a soft note. We'll stick with it. Captain America's chill. Also a bad example. I got you doing this. And I was having this conversation with the customer. And they're like, hey, we started using Pro for everything. And then we're using Pro for, can you summarize this email for me? Can you help me write better emails for me? I'm like, okay, those are the use cases where you can get by using Flash, which is incrementally cheaper and gets you the same outcome. So how are you more thoughtful about using what what what agents uh for your use cases and i think ultimately it goes back to um you may not have to complicate it too much of like educating Everyone like use this for that use this for that i think if there is an orchestrator layer based on my question we will be able to figure out like hey this is a better fit for flash versus Pro and so and so forth um i know we've been talking about the first party model first party third party model from an lm perspective i think the other thing that is often lost in translation Is the cost that goes along with it right the infrastructure cost training cost entrance cost storage adapt layer cost the cost to maintain and support your AI systems, right? So you need to think about that holistically from a TCO perspective. ([Time 0:16:49](https://share.snipd.com/snip/601be0fe-21a4-4b68-92d6-a11f6fd9ecc4)) - **AI Costs Are Infrastructure Plus Organizational Change** - AI infrastructure costs include GPUs, storage, training, and big organizational change expenses. - Pathik Sharma warns every $1 spent on AI often requires ~$10 in org change, governance, and enablement. Transcript: Pathik Sharma For AI in a very different way? I think the thing that is often... Okay let me actually give you one more tidbits before i answer that question stand for study comes back with every single dollar you spend in ai you're spending 10 more dollars for organizational Change management right so the governance enabling people getting teams to use things more effectively, so on and so forth, right? So while it is very important on that $1 on how we spend it efficiently, it is also very important to think about those $10 that goes with that. And I think when it comes to managing infrastructure, it's the GPUs and TPUs, very rare resource nowadays, right? That's the brain of your inference or training. Like how do you think about driving more utilization for your GPUs or TPUs, for example, at scale? Anything that AI is using for storage, right? Maybe cloud storage buckets or structured unstructured data how are you thinking about life cycle management for that because those things tag on and actually have much higher value Value but also the cost of it right so like when you think about an infrastructure that way right like now you look at it holistically from your foundational perspective, your platform, And then your first or third-party models that you might be using. So you have the full viewpoint of how this is going to play out for my organization. ([Time 0:20:00](https://share.snipd.com/snip/386b3073-ef78-4689-891d-153cf276605b)) - **Gemma On Phone Showed Edge Models Are Practical** - Pathik Sharma installed Gemma on his phone and used it for OCR and translation at 4GB. - He uses this to illustrate edge models can handle many use cases cheaply without hitting frontier APIs. Transcript: Pathik Sharma Very cool. Hi. You mentioned Gemma. One thing that I did last week was install it on my phone. That's at 4GB. And it's mind-blowing. Like my edge device now is so much more powerful that I can do and interaction, OCR and translation and all that stuff. But I agree with what Roy mentioned. I think I was having a conversation with a customer literally just yesterday, but they're like, hey, we have use cases where we could just have Gemma solve those and we should be having An orchestrator that sends those requests directly to Jemma and not, you know, control our spending and be smart about it. Like be creative about this by deploying on an edge. But then there are other use cases where we are like, hey, we do want RAG and more scientific brains behind the scenes? And how do we start to think about it? ([Time 0:23:24](https://share.snipd.com/snip/ea78b834-be97-4042-83c1-30de9c8a25d5)) - **Start FinOps With Process Then Add Tools** - Start FinOps by focusing on organizational processes and culture before buying tools. - Roi Ravhon strongly recommends joining the FinOps Foundation to learn frameworks, then adopt tools like FinOut when ready. Transcript: Roi Ravhon But where should I start? I would recommend to sign up to the FinOps Foundation. I think FinOps is first and foremost an organizational problem that you're trying to solve. Just buying a FinOps tool is not going to solve the problem. You can optimize something, but it's not going to be like the main, this is not what you're starting FinOps for. So you need to understand what's the process going to look like. How can we design the organizational view of getting people to care about cloud costs? And it's a change. We need to change our relationship with cost and not just throwing a tool to solve the problem. Only when you understand that you need a tool to continue in scaling doing this, this is the time you need to talk to FinOps or an equivalent tool. But definitely start with the FinOps Foundation. They have a lot of resources. They're doing great work on organizing FinOps in the world right now. ([Time 0:24:50](https://share.snipd.com/snip/7fec85c5-7a6e-44f1-88d9-9a2a667acc84))