AI Broke the Old FinOps Model
A Q&A with Steve Cole, Former Leader at Adobe and AWS, on the Future of Enterprise FinOps
We recently sat down with Steve Cole — architect, engineer, and former leader at Adobe and AWS — to discuss where enterprise FinOps is heading in the age of AI.
Steve has spent decades building and modernizing large-scale compute systems where performance, resilience, and economics must coexist. At Adobe, he helped build the SaaS foundations that enabled its transition to cloud-delivered services. At AWS, he partnered with large enterprises navigating modernization under real operational and financial constraints.
His perspective is technical, pragmatic, and grounded in firsthand experience.
Here’s what he shared.
Q:You’ve seen Cloud evolve from early migration to large-scale modernization. What feels different right now?
Steve Cole:
AI fundamentally changes cloud economics.
Traditional FinOps focused on compute, storage, and networking. You could model those. You could reason about utilization curves and reservation strategies.
Now you’re layering:
Vendor-specific AI services
Token-based pricing
Agentic workflows
Data lake modernization across clouds
Costs don’t behave linearly anymore.
Organizations adopt multi-cloud to avoid lock-in — but that increases architectural complexity. Add AI services on top, and predictability becomes much harder.
Q: When we talked about forecasting GenAI costs, you called it “voodoo rocket science.” Why?
Steve:
Because agentic systems don’t scale predictably.
They pass context back and forth. They spawn multiple processes. Token usage compounds in ways teams don’t anticipate.
A team estimates a new AI feature will cost $10,000 per month.
They launch.
The first bill is $50,000.
At that point, forecasting becomes lagging trend analysis rather than a forward-looking decision tool.
The real challenge is modeling economic behavior before production — not explaining it afterward.
Q: So, is better visibility enough?
Steve:
Visibility is necessary. It’s not sufficient.
You can have dashboards that explain last month’s spend. That’s helpful.
But with AI workloads, context size, token loops, and multi-agent interactions can escalate costs quickly.
With the right system: engineering responds rather than finance reacting
FinOps has to move earlier in the lifecycle.
Q: What about automation? Enterprises want execution, not just reporting.
Steve:
Automation is powerful. But it must be governed.
In large enterprise environments, I’ve seen tools that could automatically remediate issues — change permissions, adjust infrastructure, scale resources.
Technically impressive....even more so when it's inferred rather than pre-defined in code.
But enterprises are cautious. They don’t want uncontrolled automation making changes inside critical systems.
You need:
Guardrails
Permission layers
Explainability
Human-in-the-loop controls
The future isn’t AI running wild. It’s automation operating within defined economic and governance boundaries.
Q: You emphasized the difference between potential and realized savings. Why is that important?
Steve:
Because leadership cares about realized impact.
Many tools surface optimization opportunities. That’s useful.
But mature FinOps programs track:
What could be saved
What was implemented
What was actually realized
It also acts as a lighthouse within large organizations. When one team demonstrates measurable impact, it creates momentum for others to follow.
Tracking potential → implementation → realized value provides hard data on where to focus next. For instance, or on quick optimization wins versus longer-tail structural improvements.
Without that discipline, cost management becomes reporting instead of financial control.
In large organizations, that distinction is everything.
Q: Are AI budgets still expanding unchecked?
Steve:
They’re being scrutinized more now.
There was a period of heavy experimentation. Some of that was necessary.
Now executives are asking:
What’s the unit economics of this AI initiative?
How does this tie to revenue or margin?
Can we reduce waste elsewhere to fund AI responsibly?
AI isn’t going away. But it’s entering a phase of accountability.
And that accountability requires economic clarity.
Q: Cloud providers now offer native optimization tools. What must a FinOps platform do to truly matter?
Steve:
It has to offer something beyond what I can get natively or build myself.
That means:
Economic visibility
Deep workload-level attribution
AI-specific cost modeling
Business value correlation
Governance frameworks enterprises trust
KPIS that tie outputs and costs to better decision-making
If it’s just another dashboard, it won’t move the needle.
Q: If you had to place a strategic bet for the next phase of FinOps, what would it be?
Steve:
AI cost discipline.
Organizations are launching agentic systems, running inference at scale, modernizing data platforms.
The economic behavior of these systems must be understood — not just technically, but financially.
Enterprise FinOps works best when underlying systems are designed with economic intent.
That means performance-per-dollar, resilience-per-dollar, and long-term sustainability — not just raw capability.
Final Takeaway
Our conversation with Steve Cole reinforced a fundamental shift:
The old FinOps model focused on explaining last month’s cloud bill.
The next model must:
Forecast before deployment
Govern automation safely
Track realized savings
Tie infrastructure decisions to business outcomes
Bring discipline to AI economics
AI didn’t just increase spend.
It broke linear cost assumptions.
The organizations that adapt will treat FinOps not as tooling — but as financial control infrastructure for cloud, data, and AI.
Learn how FinOpsly governs AI spend with AI cost governance.

