Share this:
Cloud teams get daily cost data. Data center teams get a monthly allocation report. For most enterprises, that gap sits on top of the majority of their IT spend.
That was the starting point when Victor Garcia, host of the FinOps Weekly Podcast, sat down with Tom Mack, Lead Technologist at Visual One Intelligence®, to talk about bringing FinOps to on-prem and making hybrid cloud work at the cost level. Below are the biggest takeaways for IT and FinOps leaders.
Listen to the full episode here
Why on-prem FinOps matters now
The FinOps Foundation now includes the data center in its framework. Tom sees a bigger shift driving it: more IT is customer-facing every year. Curbside pickup, app ordering, and in-store kiosks all run on infrastructure that now belongs in a P&L conversation, not just a back-office allocation.
The spending picture backs this up. As Tom put it, cloud accounts for roughly 30% of total IT spend, and infrastructure alone splits about 50/50. That means most IT dollars still go to on-prem resources.
The problem is speed. In the cloud, a team piloting a new service can check cost three days in. On-prem, they wait for a monthly report. “All you can do with on-prem is a technological feasibility study,” Tom said. Cloud lets you ask the business question too: what will it cost, and is it worth it?
Making on-prem speak cloud’s language
Visual One Intelligence® weighed about a dozen ways to report on-prem costs. The team chose the one closest to how AWS and Azure price infrastructure, so cloud FinOps practices carry over directly. It comes down to three daily unit metrics, tracked per system:
- Storage: cost per terabyte per day
- Compute: cost per CPU per day
- Memory: cost per GB of memory per day
The per-system detail matters. A cloud provider might offer one or two storage tiers. An on-prem environment can run 14, 20, or even 30 storage platforms, each with its own cost dynamics. Seeing cost and outcomes by platform helps inform the next purchase decision.
The second piece is an imputed menu price. Cloud providers roll power, networking, and physical ports into the price of a VM. Visual One does the same, folding those costs into a fully loaded rate. The result is a direct, like-for-like comparison between an AWS VM, an Azure VM, and any on-prem configuration you run. In most workload-to-workload comparisons, Tom noted, on-prem comes out cheaper.
Closing the last mile of on-prem tagging
On-prem tagging can reach about 70% coverage without much trouble. VMware, for example, supports tags on many resources. The last 30% is the hard part. How do you tag a physical host, or a NAS volume presented over NFS?
Visual One Intelligence® handles this in two ways:
- Read tags at the source where they already exist, such as VMware
- Tag any resource inside the platform where the source can’t, for complete coverage
Tag normalization (treating “HR” and “Human Resources” as the same thing) works across environments. Integration with tools like Apptio Cloudability brings in AWS and Azure tags too. The outcome is universal tagging: one tag, queried across cloud and on-prem, with near real-time usage and cost.
Why you can’t just point AI at your gear
A common idea right now: connect an AI assistant to your hardware’s APIs and let it compare the numbers. Tom has seen it tried many times. “Every person I’ve ever seen try to do this, this is where it fails,” he said.
The reason is normalization. HPE, Dell EMC, and Hitachi each report a terabyte differently, largely because of how they count compression and deduplication. Without normalizing that data first, the comparisons are junk. For some workloads like VDI, data reduction can reach 13 or 14 to 1, which changes the cost picture entirely.
There’s also a math problem. AI inference may not answer the same math question the same way twice. Financial reporting needs deterministic code that gives the same answer every time.
That’s why Visual One Intelligence® built on 16 years of infrastructure monitoring, with well over 100 device adapters and agentless, API-based collection. Once the data is normalized, teams can query it through the platform, a REST API, or an MCP gateway for tools like Claude and Copilot.
Sweating assets through the capacity crunch
With hardware supply tight, many teams are keeping equipment past its five-year depreciation schedule. That’s an option cloud doesn’t offer, and it raises a practical question: how many more workloads can my current gear handle over the next 12 to 24 months?
Tom described the target as running between 60% and 80% utilization:
- Below 60%: poor asset utilization
- Above 80%: queuing problems start to stack up, like holiday checkout lines
Visual One Intelligence® shows how many more VMs a cluster can take before it hits that 80% line. It also handles a FinOps quirk unique to on-prem. A fully depreciated asset shows $0 on the books, which distorts unit costs. The platform tracks both the book value and a normalized operating cost, so business decisions rest on realistic numbers.
Getting executive buy-in
On-prem often reports to IT operations while FinOps covers cloud. Victor asked how to bring the two together. Tom’s answer was direct: there’s no shortcut around executive involvement.
Where IT faces customers, the case is easiest to make. Restaurants now run Kubernetes clusters in stores. Regional banks keep on-prem systems in every branch so tellers can work offline. Tom sees the balance swinging back toward on-prem after years of cloud-first thinking.
The winning pitch is a P&L case for the business owner: daily usage and cost feedback for on-prem, customer-facing IT, the same way they already get it for cloud. Not every company needs that. One energy customer, for example, uses the platform mainly to put dollar figures on capacity planning.
What’s next: GPUs and on-prem AI infrastructure
Many companies want to run small, specialized language models on hardware they own. When those models carry core business know-how, they don’t want them running on someone else’s infrastructure.
The challenge is getting the most out of that hardware. Tom noted that the same GPUs, configured differently, can produce up to a 60x difference in tokens per second. Utilization alone won’t reveal it, since GPU and CPU often sit at 100% either way. Wasted memory allocation is where the gains hide.
Tom expects that configuration problem to fade within about two years, much like early virtualization growing pains did. In the meantime, Visual One Intelligence® is building more visibility into the infrastructure these workloads run on. His framing on the trade-off stuck: cloud has no limit on spend, while on-prem has no limit on time. A job may run eight hours instead of ten minutes, but the cost is predictable.
Hear the full conversation
Tom and Victor cover much more in the full episode, including a screen-share walkthrough of how Visual One Intelligence® collects, normalizes, and prices on-prem data.
- Listen to the episode on FinOps Weekly
- Watch the video version on YouTube
- Explore more episodes of the FinOps Weekly Podcast
Ready to see your data center the way you see your cloud? Learn how Hybrid FinOps™ works or take the self-guided tour.
