The “your cloud bill went up” conversation is so familiar it’s almost background noise. What’s been different in the first half of 2026 is the shape of the increases. The cost pressure isn’t from the usual suspects — compute and storage growth — it’s from a set of more subtle drivers that didn’t show up in last year’s budget planning at most enterprises.

After looking at several real enterprise cloud bills over the past few months and comparing notes with finance and IT leaders, the pattern is clear enough to be worth writing about.

Where the Cost Is Actually Coming From

The big-ticket compute and storage spend is still the largest line. What’s surprising people is the velocity of growth in three smaller-but-fast-growing categories:

• AI model inference costs — even with per-unit pricing falling, total spend is up because volume is rising faster than prices are dropping
• Data movement and egress — particularly across regions and out to edge or on-premises systems
• Observability and security tooling — the unit costs of logging, monitoring, tracing, and security analytics have crept up on workloads that have grown

These three categories are now meaningful line items at most enterprises. They were rounding errors three years ago.

The AI Inference Trap

The AI inference cost trajectory is the most interesting story. The per-token cost of frontier model inference has dropped substantially over 18 months. CIOs anticipated this and budgeted accordingly. What was harder to predict was the rate at which usage volume would grow.

The pattern at most enterprises is that a few AI use cases that started as experiments have become embedded in workflows. Once embedded, usage scales with business activity rather than with whatever the original pilot budget assumed. A customer service team using AI summarisation across every ticket generates inference volume that scales with ticket volume. A sales team using AI on every call generates volume that scales with calls. The user base grows. The per-user usage grows. The total bill grows faster than the unit price falls.

This isn’t a problem to be solved by optimising the inference cost per call. It’s a budgeting problem — the AI spend is now an operational cost that grows with the business rather than a project cost that can be capped. The CIOs who haven’t reframed this are facing finance conversations they’re not enjoying.

Egress Has Become a Strategic Issue

Data egress costs were tolerable when most workloads stayed within a single cloud region. They’ve become significant as architectures have got more distributed — multi-region deployments, hybrid cloud-edge architectures, integrations with on-premises systems, and the rise of analytics platforms that need data movement between transactional and analytical layers.

The cost trajectory here is harder to forecast because egress is a function of architectural decisions and usage patterns that aren’t always centrally controlled. Application teams making sensible local decisions can produce cumulative egress patterns that the IT finance team didn’t anticipate.

Some enterprises are responding by establishing internal architectural review processes that explicitly factor egress costs into design decisions. Others are tolerating the higher bills because the alternative — restricting architectural flexibility — feels worse. The right answer probably depends on the specific cost magnitude and the strategic value of the flexibility.

Observability Costs Are Now a Real Line Item

The observability stack — logs, metrics, traces, security analytics — has become a more significant cost than most CIOs were tracking. The trend toward more granular telemetry, longer retention, and more sophisticated analytics has pushed the cost up steadily while the actual operational use of the data has often lagged.

The honest conversation in many enterprises is whether the volume of observability data being collected is actually being used. In many cases, the answer is no — it’s collected because the tooling makes collection easy and because nobody wants to be the person who deletes the log line that would have caught the next incident.

The smarter observability programs are now doing serious work on what to collect, what to retain, and how long to keep it. This requires engineering effort and produces uncomfortable conversations about which signals matter. The payoff is meaningful cost reduction without losing the operational visibility the observability program was meant to provide.

What’s Quietly Helping

A few things have moderated the cost growth:

• Reserved capacity and commitment-based pricing — savings that require longer-term forecasting commitments but produce real reduction
• More aggressive workload right-sizing — automated tooling has improved enough to make this less manual
• Workload migration to lower-cost regions or providers for specific use cases — politically harder than it sounds
• Selective workload repatriation to on-premises or co-located infrastructure for predictable, steady-state workloads
• Better unit economics tracking — knowing what each business activity actually costs in cloud terms

The last item is the most underrated. The enterprises that can attribute cloud cost to specific business activities, products, or teams make better optimisation decisions than the ones who can only see aggregate cloud spend. The instrumentation work to enable this is non-trivial but pays back consistently.

The Multi-Cloud Reality

Multi-cloud commitments made for sovereignty, regulatory, or strategic reasons have produced cost complexity that wasn’t always factored into the original decisions. Running workloads across two or three cloud providers means licensing, support, training, and tooling costs that don’t fully consolidate.

A few enterprises are quietly walking back ambitious multi-cloud strategies as the cost reality becomes clearer. Others are committed and managing the complexity. The right answer depends on the specific drivers for multi-cloud in the first place — regulatory requirements remain valid, while “avoiding lock-in” looks weaker when the operational cost of the alternative is clear.

What CIOs Are Doing Tactically

The tactical responses I’ve seen working in mid-2026:

• Establishing cloud cost ownership at the team level with monthly accountability
• Investing in unit economics instrumentation to make cost decisions data-driven
• Reviewing observability costs critically and rebasing the retention and ingestion strategy
• Setting AI inference cost budgets at the use-case level rather than aggregate
• Building cost review into architectural decision processes rather than treating it as a post-hoc finance concern

These aren’t dramatic interventions. They’re the kind of operational discipline that compounds over time. The CIOs who have been doing this consistently for a couple of years are in a much better position now than those starting from scratch.

The Broader Picture

Cloud cost pressure isn’t unique to 2026. It’s a constant feature of running production infrastructure on someone else’s hardware. What’s specific to this year is the velocity of growth in the newer cost categories that didn’t exist or were small a few years ago.

The CIOs who treat this as a strategic finance issue rather than a tactical IT optimisation problem are positioning their organisations for the next several years better than those who view each spike as an isolated problem. The cost pressure is going to continue. The organisations that build the muscles to manage it are going to find it manageable. The ones that don’t are going to keep being surprised.