The Great Shutdown: Why companies want to get rid of their AI agents – When AI agents cost more than they are worth

When autonomous agents overtake control and data centers burn more capital than the market can earn

Billions invested without return: AI agents must prove their usefulness

Artificial intelligence is entering a new phase. The first phase was dominated by technological breakthroughs, spectacular demonstrations, and the prospect of making knowledge work significantly more productive. The second phase saw enormous budgets poured into models, data centers, graphics processors, cloud platforms, and initial enterprise applications. Now, the third phase is beginning: the economic and organizational reality check. This will determine not only which providers survive, but also whether companies will actually create lasting, measurable value from AI or merely build up new cost, control, and liability risks.

Two developments particularly highlight this turning point. At the application level, autonomous AI agents are proliferating faster than many companies can adapt their governance, access rights, cost control, and responsibilities. Simultaneously, at the infrastructure level, data centers and specialized cloud capacities are emerging on a scale for which there is currently no sufficiently large and willing end market to operate economically. This represents a convergence of two different overestimations: companies are overestimating the organizational maturity of autonomous systems in the short term, while investors and infrastructure operators may be overestimating the speed at which AI demand can be translated into profitable revenue.

The crucial question is therefore no longer whether AI works technically. It is where it functions reliably, controllably, and profitably under real-world conditions. This shift is changing the logic of the market. Maximum autonomy, maximum model size, or maximum computing power alone will not determine success. Competitive advantages arise where technical capability is combined with robust process design, clear accountability, verifiable decisions, realistic cost models, and differentiated customer value.

The great disillusionment begins in the companies

The prediction that more than 40 percent of agentic AI projects could be discontinued, downgraded, or fundamentally redesigned by the end of 2027 should be understood less as a judgment on the long-term significance of the technology than as a critique of its current implementation. Many projects begin as proof-of-concept trials under controlled conditions. In this environment, data is selected, tasks are limited, users are benevolent, and errors have relatively minor consequences. However, the transition to production operations changes all the parameters. An agent then has to deal with incomplete data, conflicting rules, legacy systems, special cases, changing permissions, and liability-related decisions.

It is precisely at this threshold that a significant number of technological pilot projects fail. A convincing demonstrator answers the question of whether a function is fundamentally possible. It does not answer whether it can be executed millions of times with acceptable overall costs, stable quality, and manageable risk. This problem is exacerbated with autonomous agents because they can not only generate content but also independently call up tools, modify data, trigger processes, and link decisions together. An error, therefore, is not necessarily limited to an incorrect text response. It can propagate as an erroneous order, unauthorized data access, a wrong customer decision, or a change to a production system.

The anticipated project cancellation is therefore economically plausible. Rising costs, unclear business value, and inadequate risk controls are not separate problems; they reinforce each other. The less reliably an agent operates, the greater the effort required for monitoring, auditing, and corrective action. The more manual intervention is needed, the smaller the actual productivity gain. The smaller the productivity gain, the harder it is to justify additional expenditures for models, integrations, monitoring, security, and compliance. A system originally planned for automation can thus create a new, largely invisible layer of manual oversight.

Autonomy becomes a new type of cost

The costs of agent-based AI are often viewed too narrowly. Companies compare model prices or calculate the fees for input and output tokens, but this only captures part of the economic burden. An agent often breaks down a task into multiple steps, accesses various systems, repeats failed attempts, checks intermediate results, and maintains longer contexts. Complex workflows can therefore require significantly more computing power than a conventional chatbot. Added to this are expenses for orchestration, data preparation, interfaces, logging, identity management, security controls, evaluation, and the operation of different models.

Even more significant are indirect costs. These include the time spent by specialists reviewing results, resolving errors, implementing additional internal controls, providing training, conducting legal reviews, and adapting workflows. If an agent performs a task for just a few cents, but the result requires regular review by a qualified individual, the model price becomes virtually meaningless as a metric. What matters is the total process cost for each correctly completed task. This must also include repetitions, the consequences of errors, exceptions, and the cost of human oversight.

The fact that only 21 percent of the CIOs surveyed have complete transparency regarding the costs of their AI agents and AI workloads reveals a fundamental lack of control. Without allocating costs by use case, business unit, and process, it is impossible to assess the economic benefits or manage a reliable budget. The risk is that AI spending will grow decentrally across cloud accounts, software subscriptions, and departmental budgets, while central IT only sees a portion of the total costs. The bill then only appears with a delay, when usage, data volume, and the number of agents have already increased significantly.

For a proper evaluation, companies must therefore move from considering only the technical usage costs to a comprehensive value model. The costs per successful process completion, the rate of correctly completed cases without human intervention, the error and repetition rates, the throughput time, the control effort, and the actual realized financial benefit should all be measured. Only when these metrics remain stable over a longer period can it be assessed whether an agent should be scaled. Autonomy is not an end in itself, but rather an investment variable: it is only economically viable if the additional benefits outweigh the increased control and error risks.

Shadow IT gains agency

The proliferation of AI agents is changing the classic problem of shadow IT. Previously, departments procured unauthorized software or stored data in uncontrolled services. Today, employees can quickly create their own agents that access company data and independently perform actions. A survey of 685 enterprise CIOs revealed that 84 percent report employees are developing AI agents faster than IT can implement appropriate governance rules. The problem, therefore, lies not only in an increasing number of applications, but also in an increasing number of digital actors.

These actors often possess technical identities, API keys, service accounts, or delegated user rights. This creates a new class of non-human identities that, in many organizations, are not subject to the same rules as employees. 59 percent of the surveyed organizations have not yet fully integrated AI agents into their existing identity and access management policies. At the same time, 72 percent report a medium to large gap between the access an agent is theoretically allowed to have and the access the organization can actually demonstrate.

Economically, this gap is particularly dangerous because it shifts the expected balance between efficiency and risk. An agent can act around the clock and at high speed. The very characteristic that promises productivity increases the scope of the damage in the event of an error. Broad permissions, long-lasting access keys, and inadequate logging can lead to a single incorrect decision being repeated systematically. The marginal cost of another incorrect action is almost zero for the agent, while the consequential costs for the company accumulate.

The consequence is clear: Every productive agent needs a formal identity, a documented purpose, a responsible human owner, and limited privileges. Access should be granted according to the principle of least privilege, time-limited, and continuously monitored. Particularly consequential actions must be human-approved before execution. Centralized shutdown mechanisms are equally essential so that agents do not have to be laboriously deactivated individually in multiple systems in the event of a malfunction. While these measures may reduce short-term speed somewhat, they increase the overall system's scalability by limiting damage, uncertainty, and regulatory friction.

Responsibility without actual control

The organizational tension is revealed in a global survey of 3,200 CIOs. Eighty-eight percent observe that the operational adoption of AI is progressing faster than the adaptation of governance. At the same time, 90 percent assume that central IT will be held responsible for security or compliance failures, even if the tools causing them were procured independently of other business units. This creates a disconnect between decision-making power and accountability.

Such a structure creates perverse incentives. Business units gain the short-term advantage of rapid experimentation, while a significant portion of the risk remains with IT, data protection, information security, and management. Central IT can become a bottleneck if it is required to approve each application individually. However, if it relinquishes control entirely, a complex landscape of models, agents, data accesses, and contracts emerges. Both extremes are inefficient. A purely centralized approach stifles innovation, while a completely decentralized approach multiplies risks and duplication of effort.

A federated governance model is therefore required. Central functions define binding technical and organizational guidelines, provide approved platforms, and monitor identities, costs, and critical activities. Business units may develop and operate applications within this framework, but must document benefits, data sources, risks, and responsible individuals. This keeps decisions as close as possible to day-to-day operations, while common standards prevent each team from having to solve the same security and integration problems repeatedly.

The personal pressure on CIOs demonstrates how drastically AI has evolved from an innovation topic to a career and governance factor. 76 percent of the IT executives surveyed believe their position is at risk if their company fails to achieve measurable business results with AI by the end of 2027. This pressure can accelerate sensible prioritization, but it also carries the risk of frantic scaling. Those under pressure for short-term success might prioritize visible implementation figures over robust process improvements. The paradoxical consequence would be that the very pressure to quickly demonstrate a return on investment increases the likelihood of costly mistakes.

When the controller himself becomes a cost factor

Initial experiences from governance and auditing indicate that the risks are not merely theoretical. In a survey of 417 governance, risk, compliance, and audit professionals, 34 percent reported that their organization had acted based on an inaccurate agent decision. 30 percent observed unintended actions, 25 percent a failure of control, and 24 percent impacts on customers. In 20 percent of cases, two agents came into conflict with each other. Nearly half of the employees required more time to review and correct AI results than before.

This data must be interpreted with caution because surveys reflect participants' perceptions and experiences, not automatically the frequency of all errors in the overall market. Nevertheless, they reveal a clear pattern: companies are delegating decision-making authority to agents before processes, data, and control systems are designed for this form of autonomy. AI doesn't automatically make a bad process good. It can merely make it faster, more opaque, and more widespread.

The additional auditing effort is particularly relevant from an economic perspective. Automation is usually justified by the argument of reduced personnel costs. However, if qualified specialists have to continuously check results, clarify exceptions, and reconstruct agent decisions, the work merely shifts. It doesn't disappear, but rather moves from the actual processing to monitoring and troubleshooting. These activities can be more demanding and expensive than the original process because they require technical, professional, and regulatory expertise simultaneously.

This does not mean that human oversight is inherently uneconomical. In high-risk areas, it is indispensable. The crucial point is to align the intensity of control with the potential damage. An agent categorizing internal documents requires different approvals than a system that initiates payments, prepares personnel decisions, or modifies customer accounts. Companies should therefore grant autonomy in stages. The more irreversible an action, the more sensitive the data, and the greater the potential external impact, the more stringent the prior approvals, logging, and independent audits must be.

Test paths are becoming part of the economic infrastructure

Two-thirds of CIOs are unsure whether they could provide a regulator with a complete, end-to-end account of an agent's decision. This is more than just a compliance issue. A lack of traceability increases the cost of every incident because causes are harder to determine, affected processes are more difficult to isolate, and countermeasures are slower to implement. Without robust records, a company can hardly prove that an agent acted within their authority or that a responsible human intervened appropriately.

Auditability must therefore be understood as a productive infrastructure, not as a documentation requirement added retroactively. The identity of the agent, the model and software version used, the data sources accessed, the available permissions, the intermediate steps, external tool calls, approvals, and the final result should all be recorded. Equally important is the ability to technically reproduce a process or at least plausibly reconstruct it. With probabilistic models, exact replication is not always possible, but the relevant inputs, configurations, and decision paths must be preserved.

The challenge increases in multi-agent systems. When multiple agents distribute tasks, mutually review results, or sequentially modify various enterprise systems, a chain of distributed responsibility emerges. An error can arise from the interplay of several seemingly plausible partial decisions. Conventional control systems that examine individual user actions or isolated software transactions are insufficient for this. Companies need continuous monitoring of the entire process chain.

Regulatory developments are intensifying this pressure to act. The European legal framework for artificial intelligence, data protection regulations, and industry-specific supervisory obligations are increasing the demands on risk management, documentation, and human oversight. Data protection authorities outside the European Union are also paying closer attention to how automated systems process personal information. The economic conclusion is that governance is not an additional cost center that can be added after successful scaling. It is a prerequisite for ensuring that scaling is legally compliant and sustainable.

AI risks move into the top group

In a survey of 1,330 board members and executives, new risks arising from the introduction of AI for the period from 2027 onward already rank second among short-term business risks, directly behind cyber threats. This ranking shows that AI is no longer viewed as an isolated IT initiative. It affects operational stability, regulation, data protection, third parties, personnel, reputation, and capital allocation simultaneously.

AI and cyber risks should not be analyzed separately. Agents increase the digital attack surface because they require access to data, applications, and tools. Attackers can attempt to manipulate input, induce agents to perform unauthorized actions, or misuse their credentials. At the same time, a properly authenticated agent can itself perform risky actions if its goal is incorrectly defined, its context manipulated, or its permissions too broad. Traditional IT security, therefore, only addresses part of the problem. Beyond determining whether an actor is authorized to access data, it is crucial to assess whether the intended action is permissible and appropriate within the specific context.

This entails expanded responsibilities for company management. They don't need to understand every model technically, but they must ensure that clear ownership, risk classifications, escalation procedures, and economic targets exist for key AI applications. Oversight cannot be limited to the number of projects launched or the size of the budget. Crucially, reliable information is needed about which processes are being automated in a productive manner, which data is affected, how frequently human intervention is required, what potential damage is possible, and whether the expected benefits are actually being realized.

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

A capital problem is growing beneath the software

While companies struggle to control their agents, the technology industry is expanding the physical foundation of AI on an unprecedented scale. Investments of nearly one trillion US dollars in AI infrastructure are expected for 2026. Between 2027 and 2030, another six trillion US dollars could be added. These sums are flowing into data centers, graphics processing units (GPUs), networking technology, storage, cooling, power supplies, and land. This expansion follows the expectation that the need to train and run AI models will grow significantly over the years.

Growth alone, however, does not guarantee an adequate return on investment. AI data center operators would need to generate cumulative revenues of approximately $8.5 trillion by 2030 to cover announced investments, financing costs, and an assumed 12 percent return on investment. Potentially attributable revenues from enterprise and consumer AI businesses are estimated at a maximum of $3.6 trillion for the same period. Even though these figures are subject to considerable uncertainty due to their long forecast horizons, the magnitude of the gap is remarkable.

However, it would be wrong to directly compare investments and AI revenues. Data centers have lifespans spanning several years, infrastructure can handle various workloads, and some of the economic benefits accrue to user companies rather than the infrastructure operator. Furthermore, decreasing computing costs can generate new demand. Nevertheless, the central message remains valid: The market must not only grow strongly but also produce monetizable applications exceptionally quickly for all planned capacities to achieve the expected return.

This condition is challenging. Many of today's AI applications reduce costs or improve existing products, but do not automatically generate additional revenue equal to the underlying infrastructure investments. Furthermore, productivity gains can be distributed between providers and customers. Intense competition drives down prices, open-source models lower barriers to entry, and more efficient models reduce the computational requirements per task. While this is beneficial for the overall economy, it can negatively impact the profitability of individual data center operators.

The depreciation clock is ticking faster than demand

AI hardware differs from long-lasting basic infrastructure in its short innovation cycles. New generations of accelerators often offer significantly more performance or better energy efficiency. As a result, older hardware can lose economic value, even though it still functions technically. An operator must therefore recoup their investment as quickly as possible through high utilization and attractive pricing. If demand falls short of expectations, book values ​​and financing models come under pressure.

By 2030, approximately $3.5 trillion in AI assets could be at risk of impairment; by 2032, this figure could rise to more than $4 trillion. An impairment charge doesn't mean the servers physically disappear. Rather, it indicates that the expected future cash flows no longer justify the asset's carrying value. For highly leveraged operators, this can tighten credit conditions, erode equity, and make further investment more difficult.

The dynamics are similar to previous infrastructure cycles. New foundational technologies often trigger expansion, initially creating excess capacity. Railways, telecommunications networks, and fiber optics have experienced comparable phases. Some of the invested capital was lost, while the readily available, inexpensive infrastructure later enabled new business models. Therefore, two statements can be true simultaneously: AI can trigger a profound economic transformation, and yet a significant portion of today's investors may still achieve inadequate returns.

For the overall economy, a market correction need not be entirely negative. Write-downs and bankruptcies can lower the cost of computing power and give more efficient buyers access to infrastructure. Companies may benefit from lower prices, even if original investors incur losses. The situation becomes problematic when overinvestment is linked to high debt, long-term energy contracts, or regionally concentrated capacity. In such cases, corrections can affect banks, energy suppliers, real estate markets, and local public budgets.

Neoclouds between growth and selection

The most vulnerable are specialized neocloud providers that supply computing power for AI workloads. Their advantage lies in their high level of technical specialization, rapid access to sought-after graphics processors, and often more flexible offerings than those of large, broadly diversified cloud companies. Companies like CoreWeave and Nebius exemplify this segment. The addressable market for AI-optimized Infrastructure-as-a-Service offerings could grow from $66 billion in 2027 to $174 billion in 2030. This would be an impressive increase, but it might still not be enough to profitably utilize the simultaneously built capacity.

The prediction that more than 80 percent of neoclouds could cease independent operation by 2030 does not necessarily imply the complete disappearance of their computing resources. Bankruptcies, acquisitions, mergers, long-term partnerships, or integration into larger platforms are all possibilities. The underlying economic factor is significant consolidation. Providers without favorable financing, a secure energy supply, high utilization, and a clearly differentiated service offering will struggle to remain independent.

The business model is fragile for several reasons. First, it requires enormous upfront investments in rapidly aging hardware. Second, the customer base can be concentrated, meaning that losing even a few large customers can drastically alter capacity utilization. Third, neoclouds stand between chip manufacturers and customers, without necessarily possessing the same level of bargaining power. Fourth, hyperscalers can compete with comprehensive platform services, global infrastructure, and existing customer relationships. Fifth, price pressure threatens as more capacity becomes available or business models become more efficient.

Not all neoclouds are therefore the same. Those most likely to survive are providers with structural advantages: long-term access to cheaper energy, particularly high technical utilization, access to scarce capital, specialized services, or a strong position in specific regions and regulated industries. Sovereign, nationally or European-controlled infrastructure can also be a differentiating factor if customers prioritize data privacy, data residency, and strategic independence over the lowest price.

Hyperscalers are buying themselves time

Large cloud and technology companies have a decisive advantage over specialized providers: They finance the expansion of AI from broad and profitable business areas. Cloud platforms, advertising, commerce, software subscriptions, social networks, and existing enterprise services generate revenue streams that can sustain investments over extended periods. This allows Amazon, Google, Microsoft, Meta, and other large providers to better weather a period of low direct AI returns.

Diversification, however, does not eliminate the return on investment problem. Even large corporations must explain in the long term how their investments create value. They simply have more time and additional monetization channels. AI can increase existing cloud usage, improve advertising products, retain customers on platforms, accelerate development processes, and defend other business areas. Therefore, some of the benefits do not appear as separately reported AI revenue. These strategic side effects make a direct comparison with neoclouds, whose economic success depends more heavily on renting out computing power, difficult.

The market power of hyperscalers could grow further as a result of the consolidation. When smaller providers exit the market, large platforms acquire customers, skilled workers, and attractively valued infrastructure. This can increase economies of scale but also exacerbate dependencies. Companies and governments must therefore weigh efficiency against resilience. A highly concentrated infrastructure is often cheaper in the short term but can influence pricing, innovation pathways, and digital sovereignty in the long run.

This creates a strategic dilemma for Europe. Building its own AI capabilities is energy- and capital-intensive, while American and Asian providers already possess significant economies of scale. Simply copying the hyperscalers would hardly be promising. A more sensible approach is to focus on specific industrial applications, trustworthy data spaces, regulated sectors, energy efficiency, European language and technical models, and sovereign operating solutions. The economic justification must stem from differentiated benefits, not from an attempt to indiscriminately imitate global capacity races.

Six trillion dollars is not a demand forecast

Another model illustrates the scale of the challenge. If annual spending on AI infrastructure is projected to reach approximately $1.5 trillion by 2031, and investments are expected to account for roughly a quarter of industry revenue, the global AI market would need to generate nearly $6 trillion annually. Existing consumer and enterprise applications could contribute between $1.2 and $1.8 trillion of this. This would leave a gap of approximately $4.2 trillion that would need to be filled by new markets.

This six trillion US dollar figure should be understood as an economic hurdle, not a reliable revenue forecast. The number describes the order of magnitude that would be needed to sustainably finance the expansion under certain return and investment assumptions. It does not predict that this revenue will actually be generated. This distinction is crucial because high potential market volumes are often treated like expected revenues in technology debates.

Additional growth markets include AI-powered search and advertising, autonomous vehicles, trucks and drones, industrial automation, robotics, digital twins, drug development, medical services, materials research, and energy. Autonomous systems are projected to have a potential of around US$400 billion, while physical AI, encompassing simulations, digital twins, and robotics, is estimated at approximately US$900 billion. These fields possess significant economic value but are developing more slowly than purely software-based products. They require hardware, regulatory approvals, safety testing, integration into physical processes, and often new liability regulations.

The time gap is therefore potentially more important than the long-term market potential. Even if robotics and industrial AI generate enormous revenues, they must scale quickly enough to justify today's investments. Data centers, chips, and financing costs are incurred immediately; new physical applications often take years to reach widespread market maturity. A fundamentally sound future expectation can therefore still turn into a poor investment if the price, financing structure, or timing are wrong.

Physical AI will determine the second half of the boom

The successes of generative AI to date have focused heavily on digital knowledge work. Texts, images, software code, research, and customer communication can be integrated into existing applications relatively quickly. The potential revenue from these markets is substantial, but hardly sufficient on its own to justify all planned infrastructure investments. The next growth phase must therefore extend deeper into the physical economy.

Physical AI connects learning systems with machines, vehicles, production facilities, and real-world environments. In factories, digital twins can simulate new production processes before expensive equipment is modified. Robots can react more flexibly to changing workpieces. In logistics centers, material flows, order picking, and autonomous transport can be coordinated. In the automotive industry, development times can be reduced, while autonomous systems enable new services. The economic value arises not only from reduced personnel costs but also from increased plant availability, lower scrap rates, faster factory start-ups, and new products.

This presents a strategic opportunity, particularly for Germany and Europe. Their industrial base, process knowledge, specialized machine manufacturers, and complex supply chains offer valuable data and application areas. However, this advantage doesn't arise automatically. Much industrial data is fragmented, proprietary, or insufficiently structured. Companies must modernize their data architectures, standardize interfaces, and digitally map operational knowledge. Without this groundwork, physical AI will remain limited to demonstrators.

Monetization also differs from typical cloud software. Industrial companies expect demonstrable reliability, long lifecycles, and clear accountability. They do not accept a permanent experimental environment in safety-critical processes. Therefore, providers must deliver robust, integrated systems that combine models, sensors, control systems, cybersecurity, and maintenance. This is precisely where a market for specialized European providers can develop, one that is less focused on maximum model size and more on domain expertise, integration, and trust.

The common core of both crises

The problems with AI agents and neoclouds initially appear separate. On the one hand, there are business processes, governance, and flawed decisions; on the other, data centers, financing, and depreciation. In reality, both are based on the same economic misconception: that technical capability will almost automatically translate into profitable, scalable benefits.

With agents, a successful pilot project is often mistaken for production maturity. With infrastructure, rapidly growing computing needs are equated with sufficient willingness to pay and a sustainable return on investment. In both cases, the intermediate stages are frequently overlooked. Between model performance and business benefit lie data quality, process redesign, integration, security, human oversight, and acceptance. Between computing capacity and investor return lie utilization, pricing, energy costs, hardware aging, financing, and competition.

This finding doesn't argue against AI. It argues for more professional capital allocation. Technological revolutions rarely proceed linearly. Excesses finance experiments and infrastructure, but subsequently lead to correction and consolidation. Winners are not necessarily the companies that invest first or the most. Often, those who prevail are those that manage costs prudently, learn from early mistakes, and align their offerings with actual willingness to pay.

The hype must translate into sound business practice

Companies should only deploy AI agents where a clearly defined process, reliable data, and a measurable target metric exist. The starting point should not be where the greatest possible autonomy can be demonstrated. The crucial factor is where a specific bottleneck can be eliminated, a lead time reduced, quality improved, or additional revenue generated. For each use case, it should be determined before the start what costs and error rates are acceptable and under what conditions the project will be stopped.

Deployment should be phased. Initially, an agent can provide suggestions, then act under human approval, and only after stable performance can it autonomously execute limited tasks. These maturity levels allow for separate monitoring of benefits and risks. A system that already generates no measurable value as an assistant will not become economically viable simply by being granted more privileges.

In parallel, a central inventory is required. Companies need to know which agents exist, which models they use, which systems they access, and who is responsible. Costs should be allocated to individual processes and business units. Critical actions require approval thresholds, and every production application needs a functioning shutdown mechanism. Governance should not be viewed as a rigid document; it must be technically embedded in platforms, identities, protocols, and development processes.

Procurement is also changing. Contracts with AI and cloud providers should include provisions on data usage, model changes, log access, security incidents, performance, liability, and switching options. Companies must avoid making core processes dependent on a single provider whose prices, models, or business strategy can change at short notice. Where possible, modular architectures and interchangeable models increase negotiating power.

For infrastructure investors, the customer is what counts, not the chip

For data centers and neoclouds, access to current hardware alone is not enough for a viable business model. Investors should primarily examine the quality and duration of demand. Long-term contracts are only valuable if customers are creditworthy and minimum purchase requirements can actually be met. Equally important are energy prices, network connections, cooling, utilization patterns, refinancing options, and the residual value of older hardware.

Particular caution is advised when growth is largely financed by additional debt and the expected capacity utilization is based on a small number of customers. A decline in demand can then simultaneously impact revenue, asset value, and financing capacity. Stress scenarios should combine lower prices, delayed commissioning, declining capacity utilization, accelerated hardware aging, and higher capital costs. Only when a business model remains viable even under less favorable assumptions is the return on investment reliable.

For specialized providers, the path from interchangeability leads through differentiation. This includes highly efficient data centers, sovereign European operating models, industry-specific platforms, guaranteed data residency, integrated security services, or a close connection between infrastructure and application expertise. As supply increases, pure capacity becomes a tradable commodity. Trust, integration, and demonstrable performance, on the other hand, remain scarcer.

The upcoming cleanup is not the end of AI

The anticipated project cancellations and the potential consolidation of neoclouds do not signify a refutation of artificial intelligence. Rather, they are typical of the transition from a technology-driven expansion phase to a market where return on investment, reliability, and institutional maturity are paramount. A large proportion of the early projects will disappear because they were poorly defined, badly integrated, or economically unviable. At the same time, successful applications will penetrate deeper into core processes.

Infrastructure assets don't simply become worthless. Computing capacity can still be used after depreciation, acquisitions, or lower prices. For users, market consolidation can even be advantageous if competition and overcapacity reduce costs. However, the distribution of profits changes: customers and later buyers can benefit, while early equity and debt investors bear some of the losses.

The most likely development path is therefore neither a complete collapse of AI expectations nor a smooth continuation of the boom. A two-pronged development is more probable. High-quality, clearly controlled applications will continue to grow and become standard business infrastructure. In contrast, undifferentiated agent projects, poorly secured business models, and overpriced capacities will come under pressure. The market is separating technological possibility from economic viability.

The real AI revolution doesn't begin with the next spectacular model, but with the sober ability to translate technology into reliable value creation. Companies that manage costs, control, and benefits together will emerge stronger from this correction. Those who scale autonomy without responsibility or computing power without profitable demand, on the other hand, are primarily financing the next wave of write-offs.

From visibility to trust: Your scalable path with Xpert.Digital - Image: Xpert.Digital

In industrial B2B, sustainable business relationships rarely emerge overnight. They develop step by step – through visibility, professional relevance, recurring touchpoints, and growing trust. Xpert.Digital's 4-stage model addresses precisely this: It offers a structured path that begins with a manageable entry point and can evolve into deeper collaboration in business development if needed.

Instead of relying on loud marketing promises, this model puts the relationship at the forefront. Companies start with clearly defined, easily calculable measures and then decide, based on their own experience, how far they want to expand the collaboration. A key factor for this undisturbed trust-building process: The platform completely avoids annoying advertising ads, so the editorial focus remains solely on the companies' expertise.

More information here:

• From visibility to trust: Your scalable path with Xpert.Digital

☑️ 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