Why 91% of Companies Planning New AI Infrastructure Lack Finance-Reviewed ROI

Artificial intelligence has created an unusual business problem. Companies are investing in computing power to accelerate automation, improve productivity, and compete in an increasingly digital economy. Yet new survey data suggests that much of the infrastructure they already have is not being fully used. Instead of resolving that mismatch first, many organizations are preparing to spend even more.

According to VentureBeat Intelligence's July–September 2026 surveys, 79% of companies running AI in production that reported their GPU utilization said their GPUs were busy half the time or less. Among those companies, 70% planned to adopt a new, additional, or replacement AI computing platform within the next 12 months. The findings raise a question that extends beyond technology procurement. Are businesses building sustainable AI operations, or are they accumulating computing capacity faster than they can turn it into economic value?

The answer matters to executives, investors, cloud providers, and anyone trying to understand whether the AI investment boom can deliver returns that justify its enormous costs.

The News Behind the Numbers

The findings come at a moment when the economics of artificial intelligence are receiving increasing scrutiny. On October 13, 2026, the International Monetary Fund warned that a sharp decline in AI-related share prices could threaten global financial stability. Much of the public debate has focused on whether companies supplying AI infrastructure are building too much capacity relative to what customers will ultimately pay for.

The survey results reveal another dimension of the problem. The potential imbalance is not confined to data centers, semiconductor manufacturers, or cloud computing providers. It may also exist inside the organizations buying and deploying AI.

Across three monthly VB Pulse surveys, the proportion of respondents reporting that their GPUs were busy half the time or less remained between 77% and 81%. The consistency suggests that underutilization was not an isolated observation in one month.

Of the 218 respondents who supplied GPU utilization figures, 172 reported that their GPUs were busy half the time or less. This represents 79% of the surveyed group.

These figures require careful interpretation. The surveys do not establish that 79% of all businesses worldwide waste their GPU capacity. They describe respondents running AI in production who provided utilization figures, not the entire enterprise market. Nevertheless, the persistence of the pattern provides a useful warning about how companies are managing their AI investments.

What Is New About the Enterprise AI Problem?

The apparent contradiction is straightforward. Why would a company with substantial unused GPU capacity plan to adopt another computing platform?

The answer may lie in the difference between possessing computing resources and having access to the specific AI capabilities employees need.

Enterprises are effectively dealing with two related but distinct requirements. The first is computing capacity. The second is access to highly capable AI models.

Companies that feared GPU shortages during 2024 and 2025 may have secured capacity in anticipation of future demand. That strategy made sense in an environment where access to advanced computing resources was uncertain. However, acquiring capacity did not automatically create the applications, engineering systems, or workloads necessary to use it efficiently.

Meanwhile, employees and developers increasingly rely on frontier models from providers such as OpenAI, Anthropic, and Google. These models can be accessed through application programming interfaces, subscriptions, and cloud platforms.

The difficulty is that the most capable closed models cannot simply be installed on GPUs a company already controls. Organizations wanting to use their own infrastructure for those workloads must consider alternative models, including open-weight systems, and undertake the engineering required to deploy them.

The result is a mismatch. A company can have available GPUs while continuing to pay external providers for model inference. Buying more infrastructure does not automatically eliminate those API bills.

Why This Matters for Business Economics

The central issue is the difference between technology spending and productive investment.

An investment creates economic value when the benefits it produces exceed its total cost. In AI, those costs include more than hardware or cloud invoices. They also encompass engineering salaries, integration work, monitoring, security, data preparation, and the ongoing expense of running models.

The survey found that 70% of companies with underutilized GPUs planned to adopt a new AI computing platform. The comparable figure among companies whose GPUs were busy more than half the time was 74%. The difference was not statistically significant.

In other words, companies reporting substantial unused capacity were almost as likely to pursue new platforms as companies reporting higher utilization.

That finding suggests that existing infrastructure utilization is not necessarily the principal factor driving procurement decisions.

The financial picture is even more revealing. In the September survey, 91% of companies planning a new AI computing platform had no return-on-investment figure reviewed by their finance team.

This does not prove that their projects lack value. Some investments may be strategically justified, and some capacity may deliberately remain available for demand spikes. A planned platform may also replace outdated infrastructure rather than add net capacity.

But the absence of finance-reviewed ROI makes it difficult to distinguish a necessary investment from an expensive experiment.

For businesses under pressure to control operating costs, that distinction is increasingly important.

How the AI Infrastructure Stack Works

Understanding the problem requires looking at the different layers of an enterprise AI system.

At the infrastructure layer, GPUs perform the computations required to train models and generate responses. Businesses can obtain these resources through public cloud services, specialized GPU providers, colocated infrastructure, or their own on-premises clusters.

At the model layer, systems interpret requests, generate text, analyze information, and support other applications. Closed frontier models are generally operated by their providers, with customers paying for access through usage-based charges or other commercial arrangements.

At the application layer, businesses integrate those models into software, internal workflows, customer support systems, research processes, and AI agents.

These layers have different economic characteristics.

A company may pay for a GPU cluster regardless of whether the hardware is being used continuously. Its costs may also include electricity, maintenance, and infrastructure management. By contrast, API-based model inference generally incurs charges tied to usage or contracted capacity.

Consequently, an enterprise can simultaneously have idle hardware and rising AI expenses.

Moving work from a commercial API to an internal GPU cluster may reduce some costs, but it is not automatic. The company needs a suitable model, sufficient engineering expertise, acceptable output quality, reliable deployment infrastructure, and enough workload volume to justify the transition.

The cheapest architecture on paper is not necessarily the cheapest architecture to operate.

Where the Data Reveals the Opportunity

The survey contains several findings that help explain the choices enterprises are making.

Source: VentureBeat Intelligence, pooled July–September 2026 VB Pulse surveys.

The figures for self-managed infrastructure are particularly interesting. Among 32 companies running AI on on-premises or colocated GPU clusters, or self-managed open-source stacks, 18 reported GPUs busy half the time or less. That represents 56%, compared with 83% among the other 186 companies running AI in production.

The survey did not establish why the difference exists. One possible explanation is that teams operating their own infrastructure have greater visibility into unused capacity and more engineering control over how workloads are assigned.

However, the sample of self-managed operators was relatively small. The results should be treated as an indication of a potential operational advantage, not proof that self-hosting will improve utilization for every business.

The broader lesson is that infrastructure ownership can create opportunities for optimization, but those opportunities depend on technical capability and workload design.

How AI Companies Make Money and Create Value

The economics of AI are shifting from a simple race to acquire computing power toward a more complicated competition over inference, productivity, and the cost of completing tasks.

Cloud providers monetize computing capacity and managed services. Model providers generate revenue through subscriptions, API usage, and enterprise agreements. Specialized GPU infrastructure companies sell access to computing resources that customers may not want to purchase and operate themselves.

For customers, however, the relevant question is different. How much does it cost to complete a useful business task?

Consider an AI system that summarizes documents, extracts information from invoices, or classifies customer inquiries. A company might pay for a premium model even when a smaller, less expensive model can perform the task adequately.

Conversely, a cheaper model may require additional prompts, longer reasoning, repeated attempts, or more human correction. Its lower price per token could therefore translate into a higher cost per completed task.

This distinction becomes more important as businesses adopt AI agents that execute multistep workflows.

According to the article's account of Gartner's forecast, inference costs per agentic workflow are expected to increase more than fivefold through 2028. The underlying concern is that lower prices for individual tokens may not offset the larger number of tokens consumed by agent-based systems.

For AI vendors, increased usage can create a valuable revenue opportunity. For customers, it can create a budget problem unless the additional expenditure produces measurable improvements.

Platform or Product. Where Is the Competitive Advantage?

The distinction between an AI platform and an AI product is becoming more consequential.

A platform supplies the underlying infrastructure, models, and tools. A product uses those capabilities to solve a specific problem for a customer or business function.

Enterprises may be tempted to invest heavily in platforms because infrastructure decisions appear strategic and future-oriented. Yet the platform itself does not guarantee productivity gains.

A well-designed AI product can deliver substantial value using a relatively modest infrastructure footprint. For example, an internal document-processing workflow may combine a small model, a retrieval system, and a human review step to automate repetitive work without requiring a large dedicated GPU cluster.

By contrast, a company could operate an advanced computing environment and still achieve limited returns if its applications are poorly integrated or employees do not use them effectively.

The competitive advantage increasingly comes from matching the right model and infrastructure to the right task.

That requires understanding which activities need frontier-model reasoning, which can be handled by smaller open-weight models, and which do not require generative AI at all.

Competitors and the Importance of Model Choice

The enterprise AI market offers several broad approaches.

Frontier-model providers compete on capability, reliability, and access to advanced reasoning. Cloud providers bundle models with infrastructure, data services, security controls, and enterprise procurement arrangements. Open-weight models offer businesses greater control over deployment, although they require appropriate technical expertise.

The original report identifies DeepSeek as the least expensive of five open models in VentureBeat's October tests. It also notes that some businesses remain reluctant to depend on Chinese models as their primary systems because of various concerns.

Those concerns can include organizational policy, regulatory obligations, data governance, security reviews, and geopolitical considerations. The source does not quantify how much each factor influences procurement decisions.

This illustrates an important trade-off. A model's token price is only one element of its total economic value.

Businesses must also assess performance on their actual tasks, the cost of operating the model, data-handling requirements, integration complexity, and the consequences of errors.

The most competitive architecture may combine several models rather than depend exclusively on one provider.

The Hidden Costs of AI Agents

AI agents can increase productivity by breaking complex work into steps, retrieving information, using tools, and coordinating subtasks. But the same capabilities can multiply inference costs.

The article cites an analysis from Vals AI in which coding-agent teams, with a lead agent running up to five subagents simultaneously, cost 1.8 to 5.1 times as much as a single agent. Only one of four tested team configurations improved benchmark scores by a statistically significant margin.

This is a warning against assuming that more agents necessarily produce better outcomes.

Each additional agent may process instructions, documents, intermediate results, and conversation history. Some of that information is repeated across multiple model calls.

Context caching can help. In a technical guide published in March, two Nvidia engineers reported that subsequent calls after the first call in a Claude Code session reused 85% to 97% of their context from a cache. The article also notes that Anthropic's cached input-token prices can range from one-fortieth to one-tenth of normal input pricing, depending on the model.

These figures show how software architecture can influence AI economics.

Efficient context management, selective use of agents, and intelligent caching can deliver savings without requiring businesses to abandon advanced models.

The Risks Facing Enterprise AI

Several risks emerge from the survey and the economics of AI deployment.

Financial risk. Companies may continue expanding AI infrastructure without a clear, finance-reviewed understanding of expected returns. If productivity gains arrive more slowly than anticipated, budgets may become difficult to justify.

Operational risk. Idle GPUs can represent unnecessary expenditure, but eliminating all spare capacity may also leave a company unable to respond to sudden increases in demand. Utilization must be balanced against reliability and performance requirements.

Model dependency. Heavy reliance on a single provider can limit flexibility and expose a business to pricing changes, service disruptions, or changes in product availability.

Engineering risk. Moving workloads to self-hosted models may reduce API spending, but poorly managed deployments can introduce new reliability, security, and maintenance costs.

Governance risk. AI agents may access sensitive information, make mistakes, or perform actions beyond their intended scope. Greater automation therefore requires appropriate permissions, monitoring, and human oversight.

The central challenge is not simply to reduce expenditure. It is to ensure that every layer of the AI system contributes to a clearly defined business objective.

How the Global AI Economy Could Be Affected

The implications extend beyond individual technology budgets.

Cloud providers and GPU infrastructure companies depend on sustained demand for computing capacity. If enterprises discover that they can complete more work with existing resources, some planned infrastructure purchases could be delayed, redesigned, or redirected toward more efficient systems.

That would not necessarily mean that AI demand is collapsing. Companies could continue increasing their use of AI while spending more selectively on hardware and infrastructure.

Model providers may benefit from growing demand for advanced reasoning, coding, and agentic applications. Yet rising inference costs could encourage customers to route routine tasks to smaller models or move certain workloads onto their own infrastructure.

The result may be a more segmented market. Premium models could handle difficult, high-value tasks, while cheaper systems process repetitive, high-volume work.

For businesses in Europe and the United States, this shift could make AI cost management a competitive differentiator. Organizations that measure outcomes carefully may achieve better margins than competitors that prioritize the size of their infrastructure budgets.

For smaller businesses, efficient model selection and cloud-based access may reduce the need to invest in expensive hardware. For large enterprises, workload routing, caching, and internal infrastructure management could become increasingly important.

The key economic question will be whether AI generates enough additional value to justify its total cost, not whether businesses can purchase more computing power.

What Businesses Should Do Before Buying More AI Infrastructure

The survey points toward a practical strategy. Enterprises should seek more value from the computing resources and model usage they already pay for before committing to another platform.

First, measure the cost of completed tasks across both GPU infrastructure and API spending. Token prices alone do not reveal whether a workflow is economical.

Second, evaluate models against real business requirements. Compare output quality, processing time, human correction rates, and total cost per successful task.

Third, control context. Send each agent only the information it needs, reuse repeated context through caching, and avoid repeatedly processing large documents when a smaller relevant excerpt will suffice.

Fourth, introduce additional agents only when the work can genuinely be divided into parallel tasks. More agents should be justified by better results, faster completion, or another measurable benefit.

Fifth, consider moving routine, high-volume work such as classification, extraction, and summarization from frontier-model APIs to suitable open-weight models running on available infrastructure.

Finally, use spare GPU capacity for workloads that can tolerate flexible scheduling, including batch inference, model evaluation, and fine-tuning. This can turn otherwise unused computing time into productive work, provided the additional operational costs are justified.

These measures do not eliminate the need for investment. They establish a stronger basis for deciding where investment belongs.

The Next Phase of AI Will Reward Efficiency

The enterprise AI boom is entering a more demanding phase. Early adoption often rewards experimentation and speed. Mature deployment requires something more difficult: consistent performance, reliable operations, and defensible financial returns.

VentureBeat Intelligence's survey data does not prove that the AI market is a bubble. Nor does it establish that companies should stop expanding their infrastructure. The findings instead reveal a disconnect between reported GPU utilization, planned platform purchases, and financial validation.

That disconnect deserves attention.

If businesses continue treating AI capacity as a strategic asset without measuring how effectively it is used, they risk building expensive systems whose economic benefits remain uncertain. If they instead connect infrastructure decisions to completed tasks, measurable productivity, and verified ROI, they can make AI spending more sustainable.

The next competitive advantage may not belong to the company with the largest GPU cluster or the most agents. It may belong to the organization that consistently delivers better results with fewer wasted computing cycles, less unnecessary context, and a clearer understanding of what each AI workflow actually costs.

Why 91% of Companies Planning New AI Infrastructure Lack Finance-Reviewed ROI

enterprise AI costs, GPU utilization, AI infrastructure, AI computing platforms, AI return on investment, AI inference costs, AI agents, open-weight models, cloud AI platforms, GPU optimization, AI context caching, enterprise automation, AI model economics, AI productivity gains