The AI ​​boom and the billion-dollar question: Demand is real, but returns are far from guaranteed – Who earns enough to pay for the infrastructure?

AI is becoming cheaper – but why are costs still rising for companies?

Cloud or own AI server: Which solution is really worthwhile for companies?

The AI ​​boom is real: Companies are increasingly using artificial intelligence, the demand for computing power is growing, and technology companies are investing billions in data centers, chips, and cloud capacity. But rising revenues, full order books, and impressive applications don't yet answer the crucial question: Will this growing use translate into sufficient profits and reliable cash flow in time?

There are significant differences between technological progress and a profitable business model. An AI application can be useful without being profitable for its provider. Data centers can be fully utilized without the investments generating an adequate return. And even if the costs of individual model queries decrease, a company's overall expenses can increase—for example, because AI is used more frequently, additional checks are required, or complex processes run in the background.

For companies, therefore, the goal is not to predict the next stock market boom or a potential correction. What matters is whether a specific application delivers measurable business benefits, what the actual costs are, and what dependencies arise. Success is not determined by the number of models used, but by the relationship between total costs, quality of results, and economic impact.

AI is becoming more productive – but also more profitable? The key risk for companies

AI in business: Why economic benefits are more important than the next model

The crucial question in the AI ​​market is not whether artificial intelligence is economically beneficial. It is whether its benefits will translate into cash flow in time, precisely where the largest investments are currently being made. Significant discrepancies exist between technological progress, increasing usage, and profitable business models. These discrepancies explain why an industry can simultaneously experience strong growth, produce productive applications, and yet come under increasing financial pressure.

The most economically accurate description of the current situation is therefore, first and foremost, an amortization risk in an exceptionally capital-intensive investment cycle. This can lead to a monetization trap, regional or technical overcapacities, and ultimately a valuation correction in the capital markets. However, these developments are neither identical nor necessarily linked.

For companies using AI, this presents a different challenge than for investors. They don't need to reliably predict when technology stocks will correct. They need to understand which AI applications generate tangible economic benefits, what the actual costs are, and what dependencies they create. Success isn't determined by the number of models used, but by the relationship between total costs, quality of results, and operational impact.

Billions in investment are not proof of billions in losses

The financial results published at the end of October 2026 by major technology companies illustrate the wide gap that capital expenditures and ongoing cash flow can now create. Alphabet reported investments of approximately $51.4 billion in the third quarter and a negative free cash flow of about $9.1 billion. At the same time, Microsoft reported Azure revenue growth of 48 percent and continued positive free cash flow of $18.6 billion. Investments of roughly $170 billion were projected for the calendar year. These companies are thus operating within the same investment cycle, but by no means in the same financial position.

For Amazon, the timing is particularly important. The reported negative free cash flow of $42.4 billion relates to the twelve months ending September 2026, not just the third quarter. Investments in property, plant, and equipment amounted to approximately $77.3 billion in the third quarter. Therefore, figures from different reporting periods cannot simply be added together to create a single quarterly funding gap.

Equally crucial is the distinction between revenue, operating cash flow, profit, and free cash flow. If a company invests more than it receives in cash from its ongoing operations, this does not necessarily mean that its revenues are lower than its investments. Nor does it necessarily mean that it is incurring correspondingly high operating losses.

A data center is paid for today but is expected to provide services over several years. The payment immediately impacts liquidity. In contrast, in the profit and loss statement, an asset is generally depreciated over its useful life. Therefore, a company can be profitable and still report a negative free cash flow after investments.

The term "financing gap" can describe this situation if used precisely. It becomes misleading as soon as it is interpreted as an uncovered operating loss or an imminent threat of insolvency. The crucial question is whether existing liquidity, other business areas, or external financing can support the investment phase and whether the new facilities will subsequently generate sufficient cash inflows.

Why genuine demand does not preclude an investment bubble

The idea that a bubble must be based on a useless product or completely fabricated demand is too simplistic. Economically significant technologies can also trigger inflated valuations and unprofitable waves of investment.

The expansion of railways or telecommunications networks generally provided societal benefits. Nevertheless, individual companies may have overestimated construction costs, overly optimistic utilization rates, or made their capacities available too early. The infrastructure sometimes remained valuable, while the original investors suffered losses.

The same analytical distinction applies to AI. First, the technology can be productive. Second, the demand for computing power can grow significantly. Third, prices for certain applications can fall. Fourth, investors may still underpay. These four developments are not mutually exclusive.

Even a rapidly growing order backlog doesn't eliminate the risk. Google Cloud recently had an order backlog of approximately $561 billion. Such a backlog represents substantial contractually agreed-upon future revenue, but it is neither immediately available cash nor a guaranteed profit. The economic quality depends on factors such as contract durations, performance commitments, cost trends, and customer concentration.

A long-term contract can provide security if prices and cost structure are well-aligned. However, it can also become a burden if a provider has committed to services under unfavorable conditions or if the necessary infrastructure becomes more expensive.

The relevant question is therefore not just whether there are enough orders. It is whether these orders generate a reasonable return after operating costs, depreciation, financing and further investments.

Four terms describe four different risks

Amortization risk concerns the recovery of the invested capital. A company invests today and expects future cash flows. Problems arise when these cash flows are lower, arrive later, or are tied up more quickly in replacement investments than planned.

The term "amortization dilemma" further emphasizes a conflict of objectives. Those who don't invest risk losing market share, technological competitiveness, or important customers. Those who invest tie up capital under considerable uncertainty. Both decisions can be costly.

The monetization trap lies elsewhere. Here, a service is used intensively, but customers' willingness to pay doesn't sufficiently cover the long-term costs. High usage figures can then create a false impression of financial strength. A free assistant with millions of users is not yet a profitable business.

The overinvestment problem, on the other hand, concerns the available capacity. If many providers react simultaneously to similar growth expectations, more computing power, buildings, or power connections can be built than are needed at profitable prices. In this context, overcapacity in a particular chip generation can exist concurrently with bottlenecks in other components.

The fixed cost trap describes the pressure of a business model with high tied-up costs. If capacity utilization decreases, these costs must be distributed across fewer units sold. The calculated unit costs increase. Whether this translates into higher market prices, however, is another question.

These terms are helpful as analytical descriptions, but not all are equally standardized subject categories. In particular, the monetization trap and the price elasticity trap should not give the impression that they are clearly defined, universally measurable crisis states.

Why high investments don't force higher prices

The statement that investments must eventually be recouped through pricing contains a kernel of truth: A sustainably viable company needs sufficient revenue. However, this does not mean that every supplier can actually pass on their historical investment costs to customers.

Prices are not solely determined by a company's costs. They also depend on competition, alternatives, willingness to pay, product quality, and negotiating power. Therefore, a company that has invested too heavily does not automatically have the right to demand higher prices.

On the contrary, high sunk costs can encourage aggressive price competition. If a data center is already built, it can be advantageous in the short term to accept additional orders, as long as their revenue exceeds the additional operating costs. While this contributes to covering fixed costs, it does not guarantee a full return on the initial capital.

In times of overcapacity, particularly capital-intensive suppliers may be the ones to lower their prices. They are trying to increase capacity utilization and generate liquidity. What seems rational for the individual supplier in the short term can put pressure on the profits of the entire industry.

Investment calculations can fail in several ways: through prices that are too low, insufficient utilization, higher operating costs, or a shortened useful life. The loss can then be incurred by shareholders, lenders, or sellers of used equipment. It doesn't necessarily have to be passed on entirely to the end customer.

This distinction is crucial for users. A financially strained AI market can lead to both price increases and decreases. More likely than a uniform trend is a stronger differentiation between standardized services and scarce premium offerings.

Data centers and chips follow different economic clocks

One of the biggest challenges is that AI infrastructure consists of assets with very different lifespans. Land, buildings, power connections, cooling, network technology, and accelerator chips do not age at the same rate.

A well-located site with a reliable power supply can remain valuable in the long term. Conversely, a particular accelerator may come under economic pressure, even if it continues to function technically. The crucial factor then is not its defect, but its relative performance compared to new alternatives.

When a new generation of hardware delivers significantly more usable power per kilowatt-hour, older hardware can lose its competitiveness. This is especially true for energy-intensive applications and high workloads. For less demanding tasks, however, the same hardware can remain economically viable for a long time.

This creates a double uncertainty. On the one hand, companies must estimate the technical lifespan of their equipment. On the other hand, they must assess how long customers are willing to pay a sufficient price for its performance.

Depreciation periods in the financial statements do not solve this problem. A longer depreciation period can reduce the expenses reported annually. However, it neither extends technological competitiveness nor automatically improves the actual cash flow.

The situation becomes particularly critical when a plant only reaches full capacity late in the investment cycle and the next wave of investments is imminent. In such cases, there is only a narrow window of time in which the initial investment can generate sufficiently high profits.

The race for market positions can be economically rational and collectively risky

The major providers are not investing solely because they expect the highest possible short-term return on investment from additional servers. They are also securing strategic positions.

Having your own infrastructure allows you to control customer relationships, launch new products faster, reserve capacity, and limit dependence on competitors. Furthermore, AI can strengthen existing business areas, such as enterprise software, advertising, commerce, or cloud services.

For a diversified corporation, not every single AI offering needs to be immediately profitable in isolation. It can contribute to the economy by retaining customers or protecting the competitiveness of other products. However, these indirect returns are harder to measure and carry the risk of overly generous justifications.

Furthermore, strategic investment pressure arises: If several large providers assume that only a few platforms will shape the future market, waiting seems risky. Everyone wants to be among the winners. This can lead to investment decisions that are understandable from the perspective of the individual company, but which, overall, exceed the economically sensible capacity level.

This mechanism is neither proof of irrational management nor a guarantee of a future crisis. However, it explains why sharply increasing investments should not be interpreted solely as a precise response to already secured end-customer demand.

Part of this expenditure is a bet on market structure, technological development, and future bargaining power.

Why cheaper AI can still lead to higher corporate bills

Technological advancements can reduce the cost of a single model query. At the same time, a company's overall AI budget can grow. This apparent contradiction resolves itself once one considers the usage.

When a service becomes cheaper and more readily available, it is often used more frequently. Occasional text drafting evolves into automated customer responses, document reviews, software testing, and multi-stage analysis processes.

Furthermore, a visible user request can trigger numerous model calls in the background. A system searches for information, evaluates matches, generates intermediate versions, checks results, and starts another attempt if uncertain. The number of billable processing steps can therefore be significantly higher than the number of human-initiated processes.

The requirements are also increasing. A simple model may suffice for a rough summary. For a robust contract review or a complex technical analysis, larger models, more extensive inputs, and additional verification steps may be necessary.

Lower prices per processing unit do not automatically protect against rising overall costs. A company can operate more cheaply per process and still pay more because it automates significantly more processes.

This isn't necessarily problematic. If the economic benefits outweigh the costs, a growing AI budget makes sense. It only becomes critical when the growth in usage is no longer linked to a corresponding increase in value creation.

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

Hardware prices are warning signals, but not a universally valid cost curve

Sharp price increases for individual graphics cards or specific memory configurations can indicate shortages and a change in willingness to pay. However, they do not allow for a direct conclusion about the total costs of local AI.

A specific retailer price depends on availability, model variant, delivery terms, and the point in time. A conspicuous individual price can distort the market. Similarly, the manufacturer's surcharge for additional RAM is not the same as the general market price of memory modules.

For local language models, available storage capacity is indeed an important factor. It influences which models and context sizes can be practically used. However, this does not mean that having as much storage as possible is automatically the most economically sound decision.

Model size, computing power, memory bandwidth, quantization, response speed, and number of concurrent users must be compatible. One configuration might be able to load a large model but be too slow in operational use. Another configuration might be able to handle a clearly defined process more reliably and cost-effectively with a smaller model.

Investment decisions are therefore not based on the price of a particularly sought-after top-of-the-line product, but rather on the price of a suitable overall solution. This also includes less spectacular alternatives: smaller models, used systems, specialized accelerators, rented dedicated servers, or a combination of different operating modes.

Those who focus solely on the price of the most powerful hardware available confuse top technical performance with economic efficiency.

The correct reference point is the reliably completed task

For small and medium-sized enterprises (SMEs), the most important metric is neither the price of a token nor the number of model parameters. What matters is the cost of completing a task to the required quality standard.

This includes model usage, technical integration, data preparation, access management, quality control, rework, and ongoing support. For in-house infrastructure, acquisition, financing, energy, maintenance, and contingency planning are additional factors.

A simple calculation illustrates the difference. Let's assume an automated document check incurs direct model costs of five cents. If, on average, two minutes of subsequent human review are required, and this internal labor hour is valued at 60 euros, that alone adds another two euros. The visible model calculation then represents only a small portion of the actual process costs.

Even more important is the question of cost. An inappropriate wording in an internal draft can easily be corrected. However, an incorrect delivery approval, a flawed contract interpretation, or an inaccurate technical recommendation can trigger significantly higher consequential costs.

Therefore, not only the average quality is economically relevant, but also the distribution of errors. Rare but costly errors can devalue an otherwise convincing average result.

The appropriate unit of comparison is therefore, for example: cost per correctly processed customer case, per approved document, or per successfully completed order. Only this perspective connects technical performance with operational benefits.

Productivity gains only translate into savings through process changes

Many AI projects promise to save working time. However, faster processing based on calculations does not automatically lead to lower costs or additional revenue.

If employees save a few minutes each day, but cannot use this time for additional tasks or increased throughput, the financial effect remains limited. While the time saved can still be valuable, it should not be recorded as a fully realized reduction in personnel costs.

The situation is different if a company can process additional orders without hiring proportionally more staff. Similarly, AI can have an economic impact if it shortens lead times, reduces error rates, or relieves scarce skilled workers of standardized tasks.

The entire process chain must be considered. Faster quote generation is of little use if technical approval still takes several days. Automated document capture is only of limited help if unstructured master data subsequently has to be corrected manually.

The greatest return on investment often comes not from replacing individual work steps in isolation, but from reorganizing a coherent process.

This also changes the investment decision. A more expensive model can be economically superior if it requires significantly less rework. A less expensive model may suffice if the task is clearly defined and an effective testing mechanism is in place. The evaluation must be carried out at the process level.

When cloud usage is economically superior

Cloud offerings have a significant advantage: they transform a portion of fixed upfront investments into variable expenses. Companies can test applications without having to initially purchase hardware for uncertain demand.

This is particularly attractive for occasional use, changing requirements, and highly fluctuating loads. Quick access to different models can also be economically valuable. If the best technical solution changes regularly, a flexible supply option reduces the risk of premature commitment.

The time factor also comes into play. A solution that can be put into productive use within a few weeks can be more profitable, despite higher ongoing operating costs, than a cheaper in-house solution that is only available after months.

However, cloud costs are only truly flexible if contracts and architecture allow for it. Minimum purchase requirements, reserved capacity, proprietary interfaces, and complex integrations can limit the options for switching providers.

Even a seemingly simple usage-based billing system requires oversight. Lengthy data entry processes, unnecessary repetitions, unlimited agent loops, or unsuitable default models can generate costs without correspondingly increasing benefits.

Cloud computing is therefore not inherently cheap or expensive. Its economic value lies primarily in mitigating uncertainty and providing capacity on demand. This advantage must be weighed against long-term usage costs and potential dependencies.

When owning your own infrastructure will actually become cheaper

Owning your own hardware can be economically attractive for recurring, well-planned tasks. However, the decisive factor is not just the total volume, but its distribution over time.

A system that is regularly used productively spreads its fixed costs across many processes. In contrast, a high-performance machine that is only needed occasionally can incur high costs per task.

A deliberately simplified example: A local solution costs €24,000 to purchase and is intended for three years of use. Without any residual value, this results in an annual capital expenditure of €8,000. If an additional €10,000 is allocated for operation, administration, and maintenance, the total annual costs amount to €18,000.

If a comparable cloud solution, after its own control and integration costs, were on average €0.30 more expensive per successfully completed task than variable local processing, the calculated cost threshold would be 60,000 tasks per year. Only above this threshold would the local solution become cheaper under these assumptions.

This calculation is not a market price forecast. It shows which factors are related. Even a different usage period, lower utilization, additional redundancy, or increased administrative effort will significantly alter the result.

Furthermore, the quality must be comparable. A local model that requires more frequent human corrections can completely lose its infrastructure advantage.

Therefore, having your own infrastructure is particularly convincing when tasks are stable, data flows are controllable, quality requirements can be met, and technical support is reliably available.

Data sovereignty is valuable, but not free

Sensitive data may justify local or highly controlled operation. However, data protection, security, and sovereignty are not inherent characteristics of server location.

A dedicated server requires access management, security updates, logging, data backup, and clearly defined responsibilities. Incorrectly configured local infrastructure can create significant risks. Conversely, a carefully designed cloud operation, both contractually and technically, may be suitable for certain applications.

Companies must therefore make specific distinctions: What data is being processed? What legal and contractual requirements apply? Who is authorized to access the data? Which subcontractors are involved? How is data deleted and security incidents handled?

Sovereignty encompasses more than just physical data storage. A locally operated model may still depend on specific chips, software libraries, or licensing terms. Complete independence is rarely achievable.

Therefore, an assessment of the additional control gained is economically sensible. If local processing protects confidential development data, bridges outages of external services, or improves long-term negotiating power, this benefit can justify additional costs.

However, it should be explicitly named. Otherwise, there is a risk that economically weak projects will be legitimized with a vague promise of independence.

Hybrid architectures avoid the false all-or-nothing decision

For many companies, a combination of different operating models is the most robust solution. Standardized, frequent tasks can run locally or on dedicated systems. Rare, complex requirements are offloaded to more powerful external models as needed.

Such routing can also be based on data classes. Particularly confidential content remains in a controlled environment. Less sensitive tasks utilize external services. This requires that the classification and technical implementation function reliably.

Similarly, the model size can be adapted to the task. Not every query requires the most powerful model. A smaller system can sort information or handle simple cases. Only complex processes are escalated.

The economic advantage arises from the combination of high base load and flexible peak coverage. Companies then do not need to design their entire infrastructure for rare peak loads.

However, a hybrid architecture also incurs costs. Multiple models and operating environments increase the demands on monitoring, testing, and fault diagnosis. Therefore, it is not automatically cheaper, but primarily a way to address different requirements more effectively.

The key factor remains the simplicity of actual operation. A manageable solution with clearly defined responsibilities can be economically superior to a technically elegant but difficult-to-maintain model landscape.

The critical point arises from multiple stresses simultaneously

There is no universally applicable point in time at which the AI ​​investment cycle tips. It typically becomes critical when several unfavorable developments coincide.

On the supplier side, this includes declining capacity utilization coupled with continued high investment commitments. Added to this can be rising financing costs, weaker customer payments, and a rapid depreciation of existing equipment.

Customers experience pressure when suppliers raise prices or tighten contract terms without a corresponding increase in added value. Applications with limited or easily replaceable benefits are particularly sensitive to this.

For business-critical processes, the willingness to pay may be higher. At the same time, the requirements for availability, traceability, and liability limitation often increase in these processes. Higher revenues therefore do not automatically translate into higher margins.

A warning sign is when revenue growth can only be achieved through disproportionately increasing capital expenditure. Another is when growing order backlogs do not translate into corresponding operating cash flows.

For users, their own tipping point is usually more concrete: The total cost per reliably completed task exceeds the economic benefit. This point can also be reached with falling model prices, for example through increased monitoring work or unproductive use.

Conversely, an application can remain highly profitable despite rising supplier prices if it eliminates a valuable bottleneck.

Three developments are more plausible than the simple bubble question

One initial scenario is an orderly productivity trajectory. Computing power becomes more efficient, applications are better integrated, and demand grows sufficiently to utilize large parts of the infrastructure effectively. Returns then arise primarily where providers control costs and customers receive measurable value.

A second scenario involves a margin problem despite continued growth. Usage continues to increase, but competition and falling prices prevent sufficient returns for all parties involved. Users benefit from cheaper services, while investors have to accept lower returns.

A third scenario is a more severe shakeout. Individual providers come under financing pressure, investments are postponed, and capacity changes hands at lower valuations. The technology remains relevant, but some of the originally invested capital is written off.

These scenarios can occur simultaneously in different market segments. Highly specialized systems may remain scarce and expensive, while standardized word processing faces intense price competition. Individual regions may experience power shortages, while elsewhere unused capacity emerges.

The likely development is therefore neither a uniform price explosion nor a complete collapse. It is a stronger economic selection process between applications, providers, and infrastructure types.

For medium-sized businesses, capital discipline counts more than technological euphoria

A robust AI strategy begins with a clearly defined business problem. Only then do the choice of model and the decision between cloud or in-house infrastructure follow.

Before a widespread implementation, a company should measure its existing processes: processing time, error rate, throughput, personnel resources, and costs of delays. Without this baseline, the subsequent benefits remain difficult to verify.

A pilot project must then demonstrate how these same parameters change under realistic conditions. Demonstrations with selected examples are insufficient for this purpose. Typical cases, complex exceptions, and an assessment of the ongoing monitoring effort are required.

For investment decisions, at least one conservative, one medium, and one favorable usage scenario should be considered. Especially with in-house hardware, it's important to examine the impact of lower utilization or earlier replacement. For cloud services, increasing volume, contract changes, and provider switches must be taken into account.

Equally important is an exit option. Data, interfaces, and processes should be designed, as far as economically feasible, in such a way that a model change does not require rebuilding the entire business process.

The goal is not to avoid all technological dependencies. The goal is to consciously accept dependencies and to know their cost.

The real decision is not made on the stock exchange

The AI ​​boom represents real demand, considerable productivity potential, and significant investment risks. This is precisely why simply comparing it to a bubble or lasting success is insufficient.

The most plausible economic diagnosis is this: The industry must prove that its future cash flows can keep pace with the rate of its current capital commitment. Growth alone does not answer this question. Large order backlogs and technological advances are also no substitute for a sound return-on-investment calculation.

This does not imply either an argument against AI or a recommendation for indiscriminate expansion for companies. What makes sense is an application-oriented approach to capital management that combines quality, total costs, and actual process impact.

The main danger is not that artificial intelligence will suddenly become useless. It is that providers and users will pay too much for a technically impressive but economically inadequately integrated service.

Those who understand this difference don't need to celebrate every investment announcement or interpret every stock market correction as the end of the technology. They can make sober decisions about where AI creates value, what infrastructure is necessary, and what price is truly justifiable.

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs