Why Cloud Computing Costs Are Becoming a Bigger Business Concern

Cloud computing arrived with a promise about money. Instead of buying machines and waiting through a purchasing cycle, a company could rent capacity by the hour and pay for what it used. For most of two decades that promise held well enough that nobody treated the cloud invoice as a management problem. It was an engineering expense that happened to be large.

That framing has stopped holding. The bill has grown faster than the businesses behind it. The things it covers now stretch well past servers into software subscriptions, licenses and even staff time. And the people who approve the spending are not the people whose daily decisions cause it. The result is a cost line most US companies can neither predict nor control, arriving just as finance teams have less room to absorb surprises.

This article looks at how that happened using published survey data from the FinOps Foundation, forecast figures from Gartner, Flexera's annual cloud report, vendor pricing documentation, and one named company's public account of leaving the cloud. It runs no cost benchmark of its own, and it treats vendor statements about their own platforms as claims rather than measurements.

What the spending figures actually show

The scale is no longer disputable. Gartner put worldwide end-user spending on public cloud services at $723.4 billion in 2025, up from $595.7 billion in 2024, a rise of about 21 percent, in a press release forecast. The same release projected that 90 percent of organizations would adopt a hybrid approach through 2027, which means most large companies now pay two infrastructure bills at once rather than one.

The growth rate has not slowed where it matters. In an August 2026 forecast, Gartner projected total infrastructure-as-a-service spending of $287.3 billion in 2026, up 29.3 percent from $222.2 billion in 2025, and spending on AI-optimized IaaS rising 96.4 percent to $42.3 billion, from $21.5 billion. Its analysts attributed the shift to inference workloads overtaking training and to agentic systems that run many steps per request. Gartner's numbers are forecasts published by a research firm that sells the underlying data, so they describe an expected market, not invoices already paid.

For context on what a single company carries, the FinOps Foundation's 2024 survey summary reported an average annual cloud spend of $44 million across 1,245 respondents, with some companies reporting more than $1 billion a year. A figure that size does not survive as an engineering footnote.

Attention follows the size. Flexera's March 2025 release on a survey of more than 750 technical and executive leaders reported that 84 percent considered managing cloud spend their top cloud challenge. Cost had stopped being a line reviewed after the fact and become the problem named first.

The waste number stopped falling

Flexera's State of the Cloud report tracks what practitioners say they throw away. Its 2026 edition states that after five years of decline, wasted cloud spend increased slightly to 29 percent, and attributes the reversal to growing cost complexity from AI and newer infrastructure and platform services. In other words, the discipline of the previous half decade did not hold once the workload mix changed.

The same report carries a second finding that is more actionable than the headline. Despite clear savings, it notes, fewer than half of organizations use any one commitment discount. That is a gap between money already offered and money actually claimed, and it sits inside most cloud budgets. A company can negotiate a discount, sign a commitment, and then fail to operate in a way that collects it. The discount on offer is not trivial: AWS documents Savings Plans as saving up to 72 percent on compute workloads, though that is a vendor advertised maximum on eligible usage rather than a rate any workload collects automatically.

Both figures are self-reported estimates from survey respondents, not audits of invoices. They describe what practitioners believe they waste, which is a weaker claim than a measured loss and should be read that way.

Why the bill rises when nothing seems to change

Consumption pricing is the root mechanism. Cost follows usage, usage follows engineering freedom, and freedom is exactly what the cloud was sold as. Nothing has to break for a bill to grow. A team can ship a feature, leave a test environment running, copy a dataset to another region, and watch the next invoice move without any single decision looking expensive at the time.

Three specific channels do most of the quiet work.

Data movement. Cloud providers charge for moving data out, and the charges are published rather than hidden. Amazon S3 pricing documentation lists a Data Transfer OUT charge from S3 to the internet, and shows inter-region transfer billed per gigabyte, with a worked example at $0.02 per GB. The other major providers meter the same thing separately. Microsoft's bandwidth pricing page lists traffic from North America and Europe to any destination as free for the first 100 GB each month and $0.087 per GB for the next 10 TB, while Google Cloud's network pricing shows data transfer out to North America at $0.12 per GiB in the 1 to 1,024 GiB band. Rates vary by service, route and volume, and vendors quote in gigabytes or gibibytes depending on the page, so a comparison needs care.

An unexpected bill from this channel is documented rather than hypothetical. Security researcher Troy Hunt wrote in a first-party account that a December Azure invoice arrived far above normal and that a single egress line item accounted for 98 percent of his bandwidth costs across all services. His site had been crawled heavily, and the traffic it pushed outward became money. Architecture reviews that never mention dollars produce exactly this bill.

Over-provisioned containers. The Datadog container report describes measuring utilization by comparing the average CPU and memory a container actually uses against the resource request it was given, then compiling that across many days of workloads. Its 2023 announcement states that more than 65 percent of Kubernetes workloads use less than half of the CPU and memory they request. The report's own conclusion is blunt: there is still plenty of room for organizations to optimize resource utilization and reduce infrastructure costs. A cluster sized for a peak that arrives once a quarter is paid for every hour of the quarter.

Scope creep inside the budget line. The FinOps Foundation's 2026 survey data shows the discipline now reaching far beyond cloud infrastructure: 90 percent of respondents manage or plan to manage software-as-a-service spend, 64 percent manage licensing, 57 percent manage private cloud, 48 percent manage the data center, and an emerging 28 percent are beginning to include labor costs. The number labelled cloud in a budget review is often the visible part of a much wider technology bill.

AI moved the cost curve, not just the total

The newest pressure is not simply more spend. It is spend with a different shape. Gartner's August 2026 release describes inference as the dominant consumption model, with agentic AI raising compute intensity because a single request triggers many autonomous steps rather than one answer. That is a pricing model reacting to machine behavior, and it makes usage harder to forecast than a chat assistant with a per-seat fee.

The FinOps Foundation's 2026 data shows how fast the responsibility moved: 98 percent of respondents now manage AI spend, up from 31 percent two years earlier. AI management has become nearly universal inside organizations that already have a cost discipline practice, and it arrived before most of them had a method for it.

That survey also records the squeeze now visible in many finance reviews. Organizations report being asked to self-fund AI investment through optimization savings. The new line is meant to pay for itself out of the old line. This is where the two stories meet: the same teams asked to find savings in a 29 percent waste figure are being asked to hand the recovered money to a workload whose cost per request nobody has yet measured honestly.

Who owns a cost nobody owns

The structural problem is allocation. Engineering teams make hundreds of small decisions that are, in accounting terms, purchases. Finance sees a total and a vendor name. Neither group sees the link between a specific architectural choice and a specific charge, so neither can change the outcome deliberately. This is not a private-sector failing alone. The US Government Accountability Office reported in GAO-19-58 that agencies do not consistently track cloud-related savings, which makes it hard for them to decide whether cloud services are worth using, and that only 10 of the 16 agencies reviewed had established assessment guidance.

This is why the practice itself changed its definition. The FinOps Foundation reports that it updated its mission from advancing the people who manage the value of cloud to advancing the people who manage the value of technology, and describes its work as shaping technology decisions before commitments are made rather than only explaining spend afterward. The rewording is small. The implication is not. It concedes that the bill cannot be managed after the fact, because by the time it appears it has already been decided.

A cost with no owner does not get smaller. It gets explained. Most of the language around cloud cost is explanation: dashboards, showback reports, quarterly reviews that identify where the money went and change nothing about next quarter.

The exit conversation, and its real limits

The loudest response has been to leave. In a public post, David Heinemeier Hansson of 37signals wrote that the company was paying over half a million dollars per year for Amazon's managed database and search services, and argued that owning hardware would be cheaper for a business with predictable growth. He characterized the provider's margin as nearly 30 percent, citing $18.5 billion in profit on $62.2 billion in revenue, and raised a second point that was not about money at all: when a major region fails, an outsized share of the internet fails with it. An engineering follow-up put the company's 2022 cloud spend at $3.2 million even after a year of optimization work, and described moving services onto owned hardware as the next step.

The move has precedent at larger scale. Dropbox reported serving 90 percent of customer data on its own Magic Pocket storage system instead of a public cloud object store. Both cases share a condition most companies do not have: a workload large enough and stable enough that owning the hardware beats renting it.

The account is honest about its own conditions, and those conditions are the reason it does not generalize. 37signals had staff who knew the stack deeply, growth it could forecast, and a business model compatible with buying machines and writing them off over years. It also built its own deployment tooling to make the move survivable. Most US companies have none of those. They have a deadline, a compliance obligation, a scaling curve they cannot predict, and no interest in running a data center.

Repatriation is therefore a real option for a narrow band of workloads rather than a general answer. What transfers to almost everyone is narrower and less satisfying: know what you are paying for, collect the discounts you already committed to, and stop paying for capacity you do not use.

What mature teams actually do

The survey data hints at where the work stops paying off. In the FinOps Foundation's 2026 dataset, workload optimization and waste reduction was the single most common current priority at 42.48 percent of respondents, but the same question asked about the next twelve months dropped to 24.96 percent. Mature teams appear to be running out of easy savings, which matches what practitioners describe: the remaining waste is often left deliberately because acting on it costs more than it saves.

A practical sequence follows from the evidence rather than from vendor playbooks.

Allocate before optimizing. A total cannot be managed. Charges need to be attributed to the teams and products that create them, which is a bookkeeping problem before it is a technology problem.

Collect the discounts already bought. Flexera's finding that fewer than half of organizations use any one commitment discount suggests the cheapest available saving is the one already negotiated and unused.

Price the architecture. Data movement, retention periods and replication topology are cost decisions made at design time. Two cents per gigabyte is only cheap in the right direction.

Size for the median, not the peak. Container utilization data shows the gap between what is requested and what is used, and that gap is paid for continuously.

Treat AI as its own line with its own denominator. Cost per useful result, not cost per request, is the only comparison that survives contact with an agentic workload.

What to watch

Four things will determine whether this stays a management priority. Whether inference pricing falls as efficiently as hardware improves, since the AI cost line depends on it. Whether the waste figure turns back down or establishes 29 percent as a steady state. Whether commitment discount usage rises, which is the cheapest measurable improvement available. And whether cost data becomes comparable across providers, which the FinOps community's work on a shared cost and usage specification is meant to make possible.

Limitations

Every figure here comes from someone else's data. Survey results from the FinOps Foundation and Flexera are self-reported by practitioners, and their respondent pools skew toward organizations that already have a cost discipline practice, which likely understates waste across the wider market. Gartner's numbers are forecasts published by a firm that sells the underlying research. Vendor pricing pages document what a provider charges, not what a specific customer pays after negotiation. Datadog's utilization findings describe workloads on its own monitoring platform. This article performs no independent measurement of any cloud bill, and its recommendations are inferences from published evidence rather than tested results.