Written by Julien Ricciarelli-Bonnal

27 July 2026

Is AI Becoming Too Expensive to Deliver on All Its Promises?

For several years, competition in artificial intelligence was primarily described as a race for performance. Companies wanted the most powerful model, the largest training runs, faster responses, broader capabilities and increasingly spectacular demonstrations. The strategic assumption was simple: even if the immediate economics were uncertain, falling behind could mean surrendering a decisive share of tomorrow’s digital market.

That race is still underway, but another contest is beginning to matter just as much. It concerns not what artificial intelligence can do, but what it costs to build, deploy and maintain. Data centres, specialised chips, electricity, cooling systems, acquisitions, talent and software integration now absorb amounts that can no longer be hidden behind technological enthusiasm alone.

SAP recently offered a measured but revealing example. The company slightly lowered its 2026 non-IFRS operating profit outlook to reflect more than €100 million in costs related to the acquisitions of Dremio and Prior Labs. At the same time, it maintained its cloud revenue objectives and continued to forecast approximately €10 billion in free cash flow. This is not a crisis. It is a reminder that an ambitious AI strategy creates costs before it delivers all of its expected benefits.

Alphabet illustrates the tension more visibly. Strong revenue growth, particularly in cloud services, now coexists with a sharp increase in infrastructure spending. The company is generating greater demand for AI-related services while simultaneously committing enormous capital to the computing capacity required to support them.

The question is no longer whether artificial intelligence can create economic value. It is whether revenue, productivity gains and new business models will grow quickly enough to absorb the cost of the race already underway.

AI May Be Digital, but Its Economics Are Highly Physical

Artificial intelligence is often experienced as an almost immaterial service. A user opens an interface, enters a request and receives a response within seconds. That apparent simplicity conceals a heavy industrial system involving data centres, networks, processors, storage, cooling and a continuous supply of electricity.

Training advanced models requires extraordinary computing capacity, but deployment creates its own long-term burden. Once a model is integrated into a search engine, business platform, office suite or consumer assistant, the cost of inference becomes a recurring expense. A demonstration used by a few thousand people is economically different from a service accessed daily by hundreds of millions.

Every improvement in performance can increase the amount of calculation required. Providers therefore face a difficult balance between making models more capable and keeping them affordable enough to deploy at scale. A system that performs slightly better but costs several times more per request may be impressive scientifically while remaining difficult to commercialise.

Technology companies are not simply financing research teams. They are building an industrial capacity whose value depends on future demand, customer willingness to pay and the useful life of the equipment. Rapid innovation creates an additional risk: infrastructure requiring billions in capital expenditure may become less competitive sooner than expected.

Revenue Arrives Gradually, While Infrastructure Must Be Funded in Advance

Much of the current economic pressure comes from a timing problem. Companies must build capacity before they know exactly how much of it customers will use. They are investing today in data centres expected to support tomorrow’s demand, even though the dominant models, platforms and use cases remain uncertain.

This is why AI spending can increase faster than free cash flow. Revenue appears progressively through cloud consumption, subscriptions, enterprise licences and new product features. Infrastructure spending, by contrast, often needs to be committed several years ahead.

The financial advantage therefore belongs to companies capable of sustaining this imbalance for a long period. Large technology groups can use profits generated by advertising, cloud computing, software or commerce to finance AI infrastructure before the business model has fully matured. Smaller competitors must specialise, rent computing resources or depend on platforms whose pricing conditions they do not control.

Acquisitions Accelerate AI Strategies but Delay Their Financial Return

SAP’s adjustment is not driven only by the operating cost of models. It also reflects an acquisition strategy intended to secure data capabilities, technologies and expertise more quickly. Dremio strengthens data infrastructure, while Prior Labs specialises in foundation models for tabular data, a particularly important format for enterprise systems.

Acquiring a company can remove several years from a development roadmap. It provides immediate access to an established team, intellectual property and products that might otherwise take much longer to build internally. In a fast-moving market, that speed can justify a substantial premium.

The purchase price, however, is only the beginning. Integration requires technical alignment, organisational restructuring, product coordination and decisions about duplicated functions. A promising technology can take years to become a profitable component of a broader offer.

This challenge is especially significant in enterprise AI. A convincing chatbot is not enough. Systems must connect with financial, commercial, logistical or human resources processes, preserve access controls, maintain data traceability and operate in regulated environments. The most expensive work often begins after the demonstration has succeeded.

SAP’s cloud activity continues to grow, and its AI investments are supported by an established commercial base. Even so, the company illustrates the delay between acquiring a strategic capability and converting it into additional margin.

Falling Prices Are Making AI Easier to Adopt and Harder to Monetise

Competition is creating a second economic tension. As more models enter the market, their performance converges across many common tasks. Customers can compare suppliers, combine several systems or replace a service whose price no longer appears justified.

This encourages adoption, but it also limits the ability of providers to pass all of their costs on to customers. A slightly less capable model may become the rational choice when it processes significantly more requests for the same budget.

Providers must therefore continue funding frontier systems while also developing smaller, faster and cheaper models. The market still rewards maximum capability, but professional users increasingly evaluate the balance between quality, speed, reliability and cost.

The same pressure affects software companies integrating AI into existing products. A feature can be sold as a premium option, bundled into a more expensive subscription or included without an immediate price increase to defend the product’s competitive position. In the last case, AI may improve retention and perceived value while increasing the cost of delivering the service.

The paradox is becoming difficult to avoid.

The more ordinary artificial intelligence becomes, the more companies are expected to provide it. Yet the more widely available it becomes, the harder it is to charge for it as an exceptional innovation.

The Next Competitive Advantage May Be Cost Discipline

Model quality will remain important, but the next phase of competition will increasingly depend on economic efficiency. Providers will need to reduce inference costs, improve chip utilisation, design specialised hardware and reserve their most expensive systems for tasks that genuinely require them.

The strongest offer will not always be the model producing the best possible answer. It may be the one producing a sufficiently good answer at a cost that can be sustained across millions of interactions.

This changes how enterprises should choose technology. The most advanced model is not automatically the best operational choice. A smaller system can be more appropriate when speed, privacy, predictable pricing or local deployment matters more than marginal improvements in reasoning quality.

Cost control will also influence product design. Companies may route simple requests to inexpensive models, escalate complex tasks to more powerful systems and introduce limits where usage creates little measurable value. AI services will increasingly be managed as economic resources rather than unlimited features.

Return on Investment Can No Longer Remain a General Promise

The first phase of adoption gave companies room to experiment. They could finance pilots, train employees and test assistants without demanding an immediate financial return because the initial objective was to understand the technology. That phase was necessary, but it cannot become a permanent management model.

Calculating return on investment requires more than comparing the cost of a licence with the number of hours supposedly saved. Companies must also include data preparation, security, supervision, integration, training, change management and the consequences of errors. A tool that appears inexpensive at purchase can become costly when it requires months of implementation or continuous human validation.

Some benefits are also difficult to capture in a simple financial calculation. Faster decisions, improved service quality, reduced risk and stronger products may generate real value without appearing as an immediate saving. The challenge is therefore not to impose a single formula, but to define what success means for each use case.

Businesses need to distinguish personal productivity tools from operational automation, AI features embedded in products, structural investments and exploratory research. Each category has a different time horizon and should be evaluated accordingly.

That discipline requires organisations to define an economic and operational governance framework for artificial intelligence⁠, with clear responsibilities, success criteria, review stages and rules for ending initiatives whose value no longer justifies their cost.

Technology Giants Can Finance the Race, but Not Endless Irrationality

The companies leading the AI market possess an enormous advantage. They already generate substantial revenue from cloud services, advertising, software and commerce. This allows them to finance infrastructure on a scale that most competitors cannot approach and to tolerate several years of pressure on margins or cash flow.

That financial strength does not make every investment rational. Overbuilding capacity, overestimating demand or entering a prolonged price war can sharply reduce expected returns. Investors may accept high expenditure while growth remains credible, but their patience will ultimately depend on whether billions invested can be connected to recurring revenue and defensible margins.

Artificial intelligence can therefore be both a genuine technological revolution and the source of localised excess. The two ideas are not contradictory. A technology may transform the economy over the long term while also producing overinvestment, inflated valuations and projects that never recover their costs.

Previous digital transitions followed a similar pattern. Networks, cloud platforms and smartphones required vast investment before reaching large-scale profitability. Yet not every company that financed those transitions became a winner. Some built the dominant platforms, while others paid for infrastructure whose value was ultimately captured elsewhere.

AI is likely to follow the same logic. The total value created may be enormous without guaranteeing that every model, acquisition or data centre will generate the expected return.

AI Does Not Lack Promises; It Must Now Finance Their Delivery

Artificial intelligence is not becoming expensive because it is useless. It is becoming expensive because its promises require infrastructure, energy, data, acquisitions, expertise and integration on a scale that extends far beyond the visible cost of a software subscription.

The first financial signals do not prove that the industry is spending too much. SAP continues to grow in cloud services, while Alphabet benefits from strong demand for AI-related infrastructure. They do show that revenue growth may not immediately prevent pressure on profits or cash flow when expenditure is increasing even faster.

The next stage of competition will therefore separate companies that merely promise more from those capable of financing, industrialising and monetising what they build. Model performance will still matter, but so will cost per use, energy consumption, infrastructure efficiency, integration, customer retention and margin.

Business leaders must also abandon the idea that adopting AI simply means purchasing a tool. They will need to select use cases capable of justifying investment, organise their deployment and stop projects whose value remains theoretical.

Artificial intelligence may deliver on a large share of its promises. The unresolved question is whether those financing them can wait long enough to capture the benefits.

👉 Expensive technology creates value only when it serves precise priorities. Ricciarelli Partners helps organisations align their investments, offers and business development with a sustainable strategy⁠.