If you’re struggling to contain cloud costs in this suddenly volatile AI-fixated environment, it might be time to consider FinOps as an exercise in granular allocation and unit economics, with a focus on outcome.
Two recent insights highlight this shift:
“Because we track costs at hourly granularity … we can tie cost drops directly to the rollout window.” That’s from Shelby Wengreen, a software engineer at Duolingo, in an article describing FinOps processes using CloudZero on Duolingo’s blog.
And: “Cost per outcome is now the number-one margin driver for outcome-based software.” That comes from David Gross, Research Director at GPU Economics LLC, in a LinkedIn post.
Together, they reveal the new FinOps reality: Your margins depend on granular cost tracking and precise unit economics.
Unit economics, in this context, means understanding your cost per relevant output, whether that’s a user session, API call, or completed task. It’s about measuring spend in terms that directly map to business value.
Across hundreds of organizations, we see the same pattern. Simply: You can’t manage what you can’t segment. Without granular breakdown, you’re staring at a single overwhelming number with no way to optimize it.
In the AI era, where a single model inference can cost anywhere from $0.001 to $10, this segmentation becomes existential. One enterprise customer might suddenly generate 100x the infrastructure cost of another due to their query complexity, yet both pay the same subscription fee.
For instance, granular segmentation might reveal your ML-powered search feature costs $50K monthly for SMB customers but $500K for enterprise. Quickly learning the cost of a new feature is critical insight for architecture, business, and pricing strategies.
Why Granularity Drives Confidence In FinOps
The power of granularity becomes clear in practice. Tag and track at hourly intervals especially during deploys, scaling events, and traffic surges.
The difference between hourly and daily cost tracking becomes clear during your first cost spike. With daily granularity, you’re investigating 24 hours of changes across dozens of services. But, with hourly data, you’re already able to isolate the problem to a two-hour window when a misconfigured autoscaling policy triggers unnecessary compute.
Duolingo has that down to a science with CloudZero, according to Shelby.
“When something goes wrong and costs spike, we can narrow the investigation to just a few hours of changes,” she writes, “rather than sifting through an entire day or week of deployments.”
Granularity isn’t just about time, service, or team. High-performing teams also tag by feature, user segment, or environment to allow for drill-downs that drive fast, cross-functional decision-making.
The value becomes clear through:
- faster feedback loops: cost impacts visible within hours
- pattern recognition: revealing hidden optimization opportunities
- precise accountability: changes map directly to teams
In your own approach, start with your top five cost drivers. Track them hourly. Once confident in the data quality, expand from there.
From SaaS-Era Metrics To Outcome-Driven Margins
David Gross connects this granularity to a larger transformation in software economics. He emphasizes that SaaS-style seat-based pricing is increasingly outdated. AI, API, and ML apps demand a different kind of pricing and tracking, based on per interaction, per query, per resolution.
Traditional SaaS enjoyed predictable unit economics: infrastructure scaled linearly with users.
However, as David suggests, AI changes the game. Pricing based on API calls will see extreme cost variation, even as much as 100x between customers despite comparable revenue from each. One complex inference driven by agentic AI processes may cost several dollars while another, simpler genAI exchange costs a fraction of that.
This cost variability makes traditional FinOps practices obsolete. Traditional metrics like LTV and CAC are now insufficient on their own. You’ll need to start monitoring cost-per-outcome if you want to protect and manage your margins.
“Unlike seat-based 2010s SaaS products, agents are increasingly priced based on the outcomes they produce, not how many employees have access to them,” David writes. “And the primary resource used to create an outcome is not a salesperson, but the memory, power, and GPUs supporting it.”
FinOps teams must therefore define their own ‘unit’ first, be that DAU, inference, API call, checkout, etc., and then they build reporting around that. The key here is to choose units that intuitively make sense to both engineering and finance.
Shelby confirms this approach at Duolingo.
“Unit economics are another way to make costs visible by putting dollar amounts in terms of units your stakeholders care about,” she writes. “Everyone’s speaking the same language, just mapped to their priorities.”
This shared language unlocks self-serve insights for engineering and forecast accuracy for finance. And it enables execs to model business impact.
That’s the new world: when everyone understands how “cost per daily active user” impacts margin, then optimization becomes a company-wide exercise, not isolated within a function.
5 Best Practices For Accurate Unit Costing
Making this company-wide optimization work requires disciplined execution. Here are five best practices that separates teams that succeed from those struggling with granular FinOps.
1. Define meaningful units
A streaming service tracks cost per stream-hour by content type, device, and region, not just ‘streaming costs’.
Start with business KPIs, work backward to technical metrics, create clear mappings, then validate with both engineering and finance.
2. Tag for outcomes, not just ownership
Layer multiple dimensions: environment (prod/staging), team (payments/search), feature (checkout-v2), customer segment (enterprise/SMB), and revenue impact (high/medium/low).
This enables questions like “What’s the cost of serving enterprise customers for our ML-ranking feature?”
3. Track deploy-time changes
Hourly granularity unlocks deploy-to-impact correlation. This means injecting deployment markers into cost tracking and correlating git commits with cost changes to enable automated regression alerts.
This establishes a direct feedback link between code changes and financial impact.
4. Visualize unit drift over time
Monthly averages hide critical patterns. Build views showing 30-day rolling averages, statistical anomaly detection, week-over-week comparisons, and cohort analysis.
You want to watch for unit costs increasing faster than feature velocity or sudden distribution changes across services.
5. Build multi-stakeholder views
Tailor your dashboards by audience. Finance wants to see DAU. Engineers want cost per endpoint. Product needs cost per feature or user segment. Executives need margin by product line and cost as a percent of revenue.
These practices compound; each one reinforces the others. Teams that implement all five see optimization cycles accelerate from months to days, with cost decisions happening at the speed of deployment rather than quarterly reviews.
The FinOps Discipline Is Expanding
“[Cost-per-outcome] is now the number-one margin driver for outcome-based software,” writes David in his LinkedIn post. “The side FinOps practice that was set up to manage the AWS bill needs to become front and center in managing infrastructure margins.”
Modern FinOps extends far beyond traditional cloud infrastructure. AI/ML infrastructure introduces GPU cost per inference ranging from cents to dollars, with tenfold differences between batch and real-time processing. SaaS platforms require managing per-integration pricing and API rate limits. Embedded costs from Stripe, Twilio, or CDNs must be tracked per transaction.
Top FinOps teams now track costs across multiple dimensions: AI infrastructure at the inference level, SaaS platforms per integration, API billing per resolution, and data pipelines per TB processed.
If you want to be an organization ahead of the curve, build a comprehensive FinOps practice and implement a reliable platform that tracks cost ingestion across all services, real-time allocation engines, unit economic processors, ML-driven anomaly detection, and automated optimization recommendations.
As Shelby puts it: “Engineers can think in terms of cost per request or instance… Finance in terms of cost per DAU or session.”
This unit-level clarity turns cost from a finance problem into a company-wide metric.
Granular Allocation Pitfalls To Watch For
Even with the best intentions, teams stumble in predictable ways. Here are the five most common pitfalls and how to avoid them:
Over-tagging paralysis: Over-tagging paralysis happens when teams create 50+ tags but use just five. Start with 5-7 essential tags, add only when needed.
Missing the forest: Obsessing over small savings while ignoring major opportunities. Always sort by impact, tackle top 20% first.
Engineering-only adoption: FinOps becomes a silo. Mandate monthly business reviews with product and finance.
Static thresholds: Fixed alerts create fatigue. Use dynamic thresholds based on historical patterns.
Retroactive only: Always investigating last month. Build predictive models and pre-deployment estimation.
The ROI Of Granular FinOps
Organizations implementing these practices will typically see 20-40% cost reduction within six months. They’ll also experience dramatically faster anomaly resolution, improved engineering cost awareness, and reduced friction between finance and engineering.
A typical fintech startup might discover 40% of GPU costs from forgotten experiments, overspending by threefold on peak traffic handling, and tenfold variance in customer acquisition costs across channels. If those can be caught and removed, this potentially translates to millions in annual savings.
These aren’t edge cases. They’re now the norm. And they signal a broader transformation in how modern software companies must operate.
Granular Allocation: The New Default
We’re witnessing a fundamental shift in how software companies must operate. A quick recap of main takeaways:
- Granularity isn’t optional — it’s how you validate change
- Unit economics are no longer a “nice to have” — they’re your margin model
- FinOps isn’t a side project — it’s now a strategic lever
Test your readiness with five questions:
- Can you trace cost shifts to deploys within hours?
- Do you know your cost per session, per user, per inference?
- Can finance and engineering speak the same “cost language”?
- Are unit costs improving faster than competition?
- Is cost a first-class architectural metric?
If the answer to any or all of these is “no”, it’s time to raise your allocation standard.
The companies mastering granular allocation and unit economics won’t just optimize margins. They’ll reimagine what’s possible when every dollar is understood, attributed, and optimized. In the outcome era, this isn’t just good practice. It’s survival.
Start tomorrow: pick one service, track it hourly, define its unit cost. The returns compound from day one.
The Cloud Cost Playbook
The step-by-step guide to cost maturity
