Thoughts of 2026-03-20

Andy Jassy just doubled his AWS revenue forecast to $600 billion over the next decade, driven by AI demand [1].

Meanwhile, only 26% of CFOs describe their cloud spend as "highly predictable" [2]. AI workloads already consume 22% of cloud budgets according to the same survey, and that number is climbing fast.

This gap between vendor ambition and finance reality is where most AI strategies fall apart.

I have been writing about Top-down FinOps for months now, and AI makes the argument even stronger. Bottom-up optimisation alone will not get you to predictable AI costs. The starting point has to be business outcomes.

Asking "How many GPUs do we need?" optimises for infrastructure.
Asking "What business outcome does this enable, and what's acceptable cost per outcome?" optimises for strategy.

72% of CFOs are willing to accept short-term cost increases for features that drive user growth [2]. The ones getting this right define acceptable outcomes first, then let technology follow.

Some observations:

Range forecasting works better than false precision. I typically think of it as Conservative (run-rate +10%), Expected (current projects +25%), Aggressive (all experiments +50%). Leadership can work with ranges. A single number that fails by Q2 helps nobody.

The "AI or Automation?" filter is underrated. Does this task require understanding context and ambiguity? That is AI, and it is expensive. Does it follow rules and patterns? Traditional automation costs a fraction.

Finance-led governance actually improves forecast accuracy. The Cloud Capital survey found 32% achieve predictable forecasts with finance involvement, compared to 16% when engineering manages costs alone [2]. Policies enforced by systems beat policies that live in PowerPoint.

The Duke CFO survey found something worth noting: AI investment is not expected to significantly reduce headcount or generate measurable cost savings in 2026 [3]. Asking "will AI save money?" is probably the wrong question. We heard the same thing about cloud a decade ago—marketing said yes, reality said it depends. The difference now: almost no one has the capability, money, appetite, or risk tolerance to run AI on-premises, so we are stuck optimising what we have rather than debating the location.

Context is everything. A high-margin tech firm where the CTO rules behaves entirely differently from a low-margin retailer where the CFO holds the budget. AI cost strategy has to fit your actual organisation.

More on this thinking at frankcontrepois.com where I have been exploring what Top-down FinOps looks like in practice.

Sources and first draft AI, getting out of mediocrity Frank.

[1] siliconangle.com/2026/03/17/amazon-ceo-andy-jassy-forecasts-cloud-revenue-hit-600b-2036-thanks-ai/
[2] cfotech.co.uk/story/cfos-tighten-grip-on-soaring-cloud-ai-spend-risks
[3] fuqua.duke.edu/duke-fuqua-insights/CFO-outlook-for-2026-tariffs-hiring-prices-AI-impact