Facing instability on several fronts, finance leaders are being asked to protect near-term margins without undermining investment that could improve future performance. Deloitte's Q1 2026 CFO Signals survey makes that tension concrete: among 200 CFOs at North American organizations with at least US$1 billion in annual revenue, 52% named cost management their most worrisome internal concern, up from 47% six months earlier. Supply-chain disruption became the leading external worry, also cited by 52%, rising from 35% in the prior quarter. The survey was fielded during the first two weeks of March 2026, shortly after the start of the Middle East conflict. These findings describe the surveyed executives, not every company or sector.
The central issue is not simply spending less. Deloitte reports that 49% of respondents identified pressure to invest in new technologies, including cloud and artificial intelligence, as a factor driving cost-management efforts. In a volatile environment, leaders have to decide which expenditure to reduce, which capability to preserve, and what evidence is strong enough to justify additional investment. Technology budgets are part of this same management problem: an AI project can create value, but its infrastructure and operating costs need to be visible and governed.
Scalable AI compute belongs in the cost-management conversation
Scalable AI compute means having access to enough processing capacity to support a workload as usage or model demands change, without committing blindly to maximum capacity at all times. For finance leaders, "scalable" should not mean "unbounded." It means that capacity, service levels, and spending can grow in a controlled way as a business case is demonstrated.
A useful investment proposal identifies the business process to improve, the people who will use the system, the expected volume of activity, and the measures by which benefits will be assessed. It also accounts for the full operating footprint: model usage, data movement and storage, application and platform services, security, integration, monitoring, and staff time. A pilot that reports only model-call charges can make an initiative look cheaper than it will be in production. Conversely, a project that yields measurable reductions in manual work or improves service may warrant investment even while other discretionary costs are being scrutinized.
Deloitte's finding that technology investment is a cost-management driver is a reminder to distinguish expense from value. A blanket freeze can defer useful modernization, while an unmeasured expansion can turn experimentation into a recurring obligation. Finance, technology, procurement, and business owners should agree on decision gates: a limited pilot, a review of adoption and unit economics, and explicit criteria for scaling, redesigning, or stopping.
How much does an LLM cost?
There is no single price for a large language model (LLM). Cost depends on the provider and model, the pricing unit, input and output volume, context length, throughput, and whether the model is accessed through a hosted service or operated on dedicated infrastructure. Hosted services commonly charge according to the amount of input and generated output processed; rates differ by model and can change. Self-hosting shifts the calculation toward hardware or cloud compute, utilization, engineering, reliability, and operations. Comparing only a quoted per-token rate therefore does not establish the total cost of a business application.
A practical estimate starts with a workload rather than a headline number. Estimate monthly requests and typical input and output sizes, then apply the current provider's rates for the model and service configuration under consideration. Add the cost of retrieval or other supporting components, storage, observability, safety controls, network transfer, and human review where applicable. Test the estimate against observed usage in a pilot, including busy periods and longer-than-average interactions. Report both total spend and a business-relevant unit such as cost per completed case, resolved support interaction, or approved document.
This approach makes the answer to "how much does an LLM cost?" useful for a budget decision: it is the cost of a defined workload and service level, not a universal model price. Teams should identify who owns the budget, set usage alerts or limits, and review changes in volume and unit cost. If demand rises, leaders can decide whether the resulting cost is justified by corresponding outcomes rather than allowing consumption to expand without review.
What is cloud cost?
Cloud cost is the total charge for the cloud resources and services an organization consumes, as well as the people and controls needed to operate them responsibly. Depending on architecture and provider, that can include compute, storage, databases, managed AI services, networking and data transfer, backups, observability, and support. The bill may be affected by capacity choices, usage patterns, discounts, licensing, and resources left running when they are no longer needed. Cloud is not automatically cheaper than owning infrastructure; its value may include flexibility and reduced need to provision for peak demand, but the economics depend on the workload and how it is managed. For a broader primer on the bill's components and the tooling used to manage them, see our overview of cloud cost management fundamentals and best tools for 2026.
For an AI application, cloud cost should be traced across the full path from user request to result. A model invocation is only one component if the system also retrieves documents, moves data, stores logs, runs application services, and performs evaluation. Finance and engineering teams can make this manageable by tagging resources to products or cost centers, monitoring usage and anomalies, and reviewing unit costs alongside service quality. Forecasts should state assumptions about adoption, request volume, model selection, and growth rather than treating one month's bill as a fixed run rate; our practical FinOps guide to cloud cost optimization in 2026 walks through how to connect that spend to business value.
Infrastructure choices: energy and the edge
AI infrastructure energy consumption is another planning consideration, particularly as workloads scale. The relevant financial question is not a generic claim that one architecture is always greener or cheaper; it is how energy, capacity, utilization, and service requirements affect the organization's actual deployment. Teams should include power and infrastructure constraints in vendor and architecture reviews where material, and compare alternatives using consistent workload assumptions.
AI edge infrastructure—processing closer to where data is generated or used—may be relevant when latency, connectivity, privacy, or local operation matters. It is not a universal cost-saving strategy. Edge deployments can introduce distributed hardware, software management, security, and maintenance requirements. A sound comparison weighs those costs against the cloud or centralized alternative and the specific operational benefit, such as responsiveness or continuity during network disruption. This is the same discipline required for other technology investment: define the use case and measure the trade-offs before expanding deployment.
Governance makes cost control compatible with innovation
Executives can turn a broad cost mandate into practical oversight without reducing every technology budget by the same percentage. Establish a baseline, assign owners to material workloads, and review forecasts against actual use. Require proposals to document expected business outcomes and full lifecycle costs. Use stage gates so that experiments can proceed at bounded scale, while further investment depends on adoption, reliability, and demonstrated value. Where an initiative fails its agreed tests, pause or redesign it; where it succeeds, scale deliberately and update the forecast. Shared practices help here: FinOps best practices for optimizing cloud costs across teams outline how finance, engineering, and business owners can split this responsibility without slowing delivery.
Interpretability can also inform purchasing and risk decisions. AI interpretability startups and established vendors offer tools and approaches intended to help users understand model behavior, but their presence does not itself establish that a system is safe, accurate, or cost-effective. Buyers should test whether a proposed tool answers a specific oversight need, integrates with the organization's workflow, and adds value commensurate with its cost. Evaluation should include the people and processes required to act on the information it produces.
Deloitte's survey does not prescribe a particular cloud budget, AI architecture, or cost-reduction program. Its signal is that CFOs are facing cost management alongside pressure to invest in technology, amid heightened external uncertainty. The practical response is to make technology economics legible: connect spending to workloads and business outcomes, expose assumptions, and revisit decisions as evidence changes. That makes scalable AI compute a managed investment rather than an open-ended commitment—and cost control a way to protect strategic capacity, not simply a brake on innovation.
Source: Deloitte, "Facing uncertainty on several fronts, North American finance leaders zero in on cost management," Q1 2026 CFO Signals survey, https://www.deloitte.com/us/en/insights/topics/business-strategy-growth/1q-2026-cfo-signals-survey.html.
