Why DigiUsher Rebuilt Its Cost Engine on FOCUS — and Why That Decision Is Now an Enterprise Advantage (Blog 3/5 Series)

You ran the utilisation report on Friday. The numbers looked reasonable — average GPU utilisation sitting around 68%, committed spend accounted for, no obvious runaway jobs. Then Monday morning, the cloud bill arrived $340,000 over forecast. The culprit: a distributed training run that spun up 200 on-demand A100 instances over the weekend, ran for 11 hours, produced a model checkpoint the ML team ultimately discarded, and terminated — leaving no anomaly flag, no alert, and zero attribution to any cost centre anyone could find.

This isn’t a technology failure. It’s a process failure with a structural cause: most FinOps workflows were designed for predictable, stateful workloads, and AI/ML infrastructure is neither.

Why the standard playbook fails when AI workloads enter the picture

The core problem with GPU cost management is that the traditional FinOps playbook — tag everything, rightsize instances, buy reservations, review weekly — was built on assumptions that simply don’t hold for AI infrastructure. Traditional applications have relatively stable, knowable compute profiles; a web server cluster might spike during peak traffic, but the envelope is predictable. AI agents and training jobs behave completely differently. Our own live TCO benchmarking finds AI agents generating 40% more bursty compute than traditional enterprise applications, and that burst isn’t randomly distributed — it clusters aggressively around ad-hoc experimentation cycles, model evaluation windows, and automated pipeline reruns, frequently outside business hours and entirely outside the visibility window of a weekly governance review.

The direct result is a specific, brutal failure mode: reserved instance commitments purchased against a projected steady-state GPU demand get stranded the moment actual workloads shift to on-demand for engineering flexibility — and then they never rebalance. Reserved instance utilisation averages just 61% at enterprises without active, specialised FinOps governance, meaning 39% of committed GPU spend generates zero compute return. For an organisation carrying $2M in annual GPU reservations, that’s $780,000 in committed spend producing nothing at all.

The three-layer remediation playbook

The fix isn’t to stop buying reservations or strangle innovation — it’s to instrument GPU infrastructure at a granularity matching the actual decision cycle of AI teams, rather than the reporting cycle of finance.

Enforce job-level attribution upstream. Attribution has to happen at the job level, not the cluster level — a Kubernetes node running mixed training and inference workloads can’t be meaningfully attributed to a cost centre at the node level, since the cost signal is far too coarse. Pod-level and job-level tagging, enforced automatically at deploy time rather than applied retroactively, is the minimum viable attribution schema for AI infrastructure. This isn’t a tooling problem so much as a policy one — FinOps practitioners need to work upstream with MLOps teams to bake financial governance directly into the orchestration pipeline before a job is scheduled, not after the bill arrives.

Shift to tokens-per-second-per-dollar unit economics. The right unit economics metric for AI workloads is tokens per second per dollar, not cost per hour — cost-per-hour only tells you what you spent; TPS/$ tells you what value you actually extracted. An H100 running an inference workload at 60% utilisation can look efficient on a traditional input-cost dashboard while simultaneously delivering worse TPS/$ than a smaller, fully saturated, lower-cost hardware configuration. Without an output-normalised metric like this, optimisation decisions get made blindly on input cost alone — systematically underinforming every hardware and reservation decision downstream of it.

Normalise multi-cloud data through FOCUS. Multi-cloud AI spend needs a unified cost schema before it needs a unified dashboard. Most organisations running GPU workloads across AWS, GCP, and Azure are dealing with three completely disparate pricing structures, three different reservation models, and three conflicting tag schemas that don’t map cleanly to one another. FOCUS exists precisely to solve this — a vendor-neutral cost and usage schema making cross-cloud GPU spend directly comparable at the line-item level. Implementing FOCUS ingestion is unglamorous backend work, but it’s the absolute prerequisite for any multi-cloud AI cost governance producing actionable numbers rather than approximations.

Closing the visibility gap

Applying all three layers together — job-level attribution enforcement, TPS/$ unit economics, and FOCUS-normalised multi-cloud ingestion — as a single interconnected governance framework rather than three separate initiatives is what actually closes the gap. A weekend overrun like the one above can’t be solved by any single isolated capability; it’s solved by systematically shrinking the delta between when a cost is generated and when it becomes visible — from an industry-average 18-to-24 hours down to minutes.

There’s one genuinely actionable step available to any FinOps practitioner today: pull your current GPU reserved instance utilisation rate. If it’s below 75%, you have a commitment rebalancing problem that no amount of dashboard investment alone will fix without upstream job attribution data behind it. Start there.

The GPU cluster doesn’t know it’s Saturday. The governance model has to catch up to that reality, not the reverse.

See what job-level GPU attribution looks like against your own estate. Book a 30-minute session and we’ll map real reserved-instance utilisation and TPS/$ against what you’re actually running. Request a demo →

Frequently asked questions

Why do standard FinOps practices fail for AI and GPU workloads specifically? Because they were designed around predictable, stateful compute — tag everything, rightsize, buy reservations, review weekly. AI agents and training jobs generate roughly 40% more bursty compute than traditional applications, clustered around ad-hoc experimentation and pipeline reruns that frequently happen outside business hours and outside a weekly review’s visibility window entirely.

What is the “stranded reservation” problem in GPU cost management? It’s when reserved instance commitments purchased against a projected steady-state demand get abandoned as actual workloads shift to on-demand capacity for engineering flexibility, and never rebalance. Reserved instance utilisation averages just 61% at enterprises without dedicated FinOps governance, meaning 39% of committed GPU spend is generating no return at all.

What is tokens-per-second-per-dollar, and why does it matter more than cost-per-hour? It’s an output-normalised unit economics metric measuring value extracted per dollar of GPU spend, rather than raw cost. Cost-per-hour only shows what was spent; TPS/$ shows whether that spend actually produced proportionate throughput — a GPU running at 60% utilisation can look efficient on a cost dashboard while delivering worse TPS/$ than a smaller, fully saturated, cheaper configuration.

Why does multi-cloud GPU spend need FOCUS normalisation specifically? Because AWS, GCP, and Azure each price GPU compute differently, structure reservations differently, and use incompatible tagging schemas — none of which map cleanly to one another. FOCUS provides the vendor-neutral schema making cross-cloud GPU spend directly comparable at the line-item level, which is the prerequisite for any multi-cloud governance producing real numbers rather than approximations.

What’s the single most actionable first step for governing GPU spend better? Pull your current GPU reserved instance utilisation rate. If it’s below 75%, that’s a commitment rebalancing problem, and no amount of additional dashboard investment fixes it without upstream, job-level attribution data feeding into the decision.

Related reading

• GPU Cost Governance for Azure OpenAI, AWS Bedrock & Google Vertex AI — March 19, 2026
• FinOps for Kubernetes: The Ultimate Guide to Rightsizing, Bin Packing, and GPU Optimization — April 14, 2026
• AI Cost Governance: How to Prevent Runaway GenAI Spend — January 14, 2026
• FOCUS 1.3 Decoded: What Changed, What It Governs, and What It Means for Your FinOps Stack — May 26, 2026

© 2026 DigiUsher. All rights reserved. Privacy Policy · Terms of Service · Trust Center