Microsoft's withdrawal from Oracle: Why the AI boom is now experiencing its economic maturity test – Creative image on the topic, featuring AI: Xpert.Digital
The AI boom is becoming more demanding: Microsoft is examining the value of expensive computing power
AI market facing a maturity test: What Microsoft's Oracle decision means for providers
Microsoft, Oracle and the AI infrastructure: Why growth alone is no longer enough
Microsoft is apparently letting its agreement to lease AI computing power from Oracle expire and has been gradually shutting down the relevant capacity over several months. This is more than just a typical contract decision: the process demonstrates that the standards in the AI infrastructure market are changing. It's no longer just the available computing power that counts, but also whether it's technically suitable for the specific application, efficiently utilized, and delivers sufficient economic benefits.
However, this doesn't automatically mean that Microsoft is scaling back its AI expansion or that Oracle's infrastructure is fundamentally no longer competitive. A targeted reassessment of specific resources is more likely. Older Nvidia chips, which provided additional capacity in an earlier development phase, may no longer meet the requirements of this major customer. Oracle, on the other hand, maintains that there is still demand for older graphics processors. Both assessments could be correct simultaneously, as different applications and customers require different performance levels.
The decisive shift, therefore, does not necessarily lie in an end to growth, but rather in a greater differentiation of the market. From a phase in which additional AI capacity was in demand almost regardless of its specific configuration, a market is evolving with varying quality levels, return profiles, and technological lifecycles. For cloud providers, infrastructure investors, and enterprise customers, one question is becoming increasingly important: Which computing power is not only available, but can also be used effectively and profitably on a sustained basis?
Related to this:
Not every graphics processor is gold: In the AI business, it's no longer just computing power that counts, but what it actually earns
Microsoft is apparently letting its agreement to lease AI computing power from Oracle expire and has been gradually shutting down the relevant capacity over several months. Economically, this is more than a typical contract decision. The process illustrates how the benchmarks in the artificial intelligence market are shifting: The mere availability of computing power is becoming less important compared to its technical suitability, its integration, and its concrete economic benefits.
The obvious conclusion that Microsoft is withdrawing from AI expansion would be premature. Nor can it be inferred from this process that Oracle's infrastructure is fundamentally no longer competitive. What is apparent is a targeted reassessment of certain resources. Older Nvidia chips, which provided valuable additional capacity in an earlier development phase, apparently no longer meet the changed requirements of a single major customer. At the same time, Oracle maintains that there is strong demand for older graphics processors as well. Both statements could be economically accurate.
The decisive shift, therefore, does not necessarily lie in an end to growth, but rather in its differentiation. A market in which additional AI capacity appeared attractive almost regardless of its specific design is evolving into one with varying quality levels, return profiles, and technological lifecycles. This marks the beginning of a more challenging phase for cloud providers, infrastructure investors, and enterprise customers.
An expiring agreement does not necessarily constitute a break between two corporations
The multi-year agreement between Microsoft and Oracle was announced in 2023. Oracle was to provide additional AI infrastructure, which Microsoft would use, among other things, to run the AI-powered version of its Bing search engine. The financial details were not disclosed. The exact number of graphics processors affected, their specific models, and the complete shutdown date are also not publicly known.
These information limitations are essential for any serious economic evaluation. Without the contract value, Oracle's immediate revenue loss cannot be quantified. Without information on the cost structure, it is impossible to determine the lost contribution to earnings. Furthermore, without knowledge of potential new customers, it remains unclear whether the freed-up capacity will remain unused for an extended period or be leased to other customers relatively quickly.
According to the information available, Microsoft did not renew the agreement in question and gradually reduced its use. This differs from an abrupt shutdown or a proven extraordinary termination. Microsoft was initially unavailable for comment regarding the underlying reporting. The specific reason for the withdrawal therefore remains based on information from sources close to the agreement, not on a detailed public explanation from the company.
Nor does this prove the end of all business relations between Microsoft and Oracle. A single infrastructure agreement should not be analytically equated with the entire relationship between two broadly diversified technology companies.
From rapid capacity acquisition to targeted optimization
To understand the economic logic of the original collaboration, one must distinguish between two different tasks: the short-term securing of additional computing power and the long-term design of the most efficient infrastructure possible.
In a rapidly evolving technological landscape, it can be advantageous to lease resources from a direct competitor. Companies that want to quickly launch a new service or meet unexpectedly high demand value immediately available capacity differently than those with several years of lead time. The benefit then arises not only from lower operating costs but also from avoiding delays.
For Microsoft, additional external infrastructure could therefore serve as a bridging function. Such a solution buys time, increases flexibility, and reduces the risk of new applications failing due to a lack of resources. However, this does not mean that the same infrastructure will always be the most economically viable solution.
As development progresses, models, software architectures, user numbers, and internal capacities change. What made sense at the start of a service may later be too expensive, too inflexible, or technically unsuitable. The economic value of an agreement therefore depends not only on its initial suitability but also on whether it continues to fit the current production system.
Microsoft's growth contradicts the simple retreat thesis
The sheer scale of Microsoft's business helps put this into perspective. In fiscal year 2026, the company generated revenue of $331.8 billion, an 18 percent increase year-over-year. Operating income reached $155.2 billion, a 21 percent rise. Revenue from Microsoft Cloud increased by 27 percent to $214.4 billion. Azure surpassed the $100 billion revenue mark for the first time.
These figures demonstrate a rapidly growing cloud business. However, they are not synonymous with pure AI revenue. Microsoft Cloud encompasses offerings beyond Azure, including commercial cloud services from Microsoft 365, parts of LinkedIn, and Dynamics 365. Even within Azure, revenue is generated from numerous applications outside of generative AI.
Nevertheless, the data clearly demonstrates why a single non-renewal of a capacity agreement is not sufficient evidence of a general strategy shift against AI. A corporation of this size can simultaneously build new infrastructure, reduce certain external services, and reassess its supplier relationships.
Strong growth increases the pressure to correct inefficient structures. An avoidable cost block, which is still manageable in an early development phase, can have a significant impact on the bottom line when used on a large scale. Growth and cost control are therefore not opposites, but increasingly intertwined.
The decisive factor is not the chip price, but the application costs
In public discourse, AI infrastructure is often judged based on the number of chips produced or investment sums. However, other questions are crucial for its actual economic viability: How many usable results does a system deliver? How quickly are they available? What is their quality? And what are the total costs?
A cheaper graphics processor can be economically disadvantageous if significantly more devices are needed for the same task. Conversely, older hardware can remain attractive if it operates reliably, offers sufficient memory, and can handle an application without incurring high additional costs.
Therefore, the decisive factor is the cost per successfully completed task. For a search service, this might be a answered query; for a document management system, a correctly evaluated file; and for an industrial application, a quality defect detected in time. These results are more economically meaningful than theoretical peak performance.
Quality assurance is another important factor. A seemingly inexpensive model response is not inexpensive if it requires frequent correction, recalculation, or review by staff. Similarly, a technically powerful infrastructure can lose value if its integration is complex and costly.
For Microsoft, this means that external capacity can continue to function and still be excluded from the preferred infrastructure portfolio. Technical usability and business viability are different categories.
Why older Nvidia chips don't automatically become worthless
The statement that older hardware no longer meets the requirements leads to a false generalization. It initially only means that a particular configuration is no longer optimal for certain tasks of a particular customer.
AI applications vary considerably. Some require very large storage capacities, others particularly short response times. Still others process large volumes of relatively simple tasks where low operating costs are more important than maximum speed. Additionally, model size, software optimization, and workload influence the selection of suitable hardware.
An older graphics processor can therefore become less attractive for demanding new models while remaining suitable for smaller models, specialized enterprise applications, or less time-critical calculations. Its market value then changes without disappearing completely.
Oracle may therefore be correct in its assessment of continued demand for older graphics processors, while Microsoft no longer wishes to use the same devices. This is not a necessary contradiction, but rather an expression of differing customer requirements.
The crucial economic question, however, is under what conditions this demand exists. Another user is only a comparable replacement if the contract duration, price, occupancy rate, and maintenance costs allow for a comparable contribution to the bottom line. Demand alone does not guarantee unchanged profitability.
Training and ongoing operations follow different cost logics
Another key to understanding this lies in distinguishing between training an AI model and its subsequent use. During training, model parameters are adjusted using large datasets. In operation, the trained model processes new inputs and generates results. This second process is called inference.
Both tasks require infrastructure, but with different operational priorities. Extensive training runs place high demands on the collaboration of many accelerators. For large, interactive applications, the reliability and speed with which a system can handle fluctuating user requests is also crucial.
For AI-powered search services, it is particularly relevant that costs are incurred anew with each use. Increasing user numbers therefore not only represent a revenue opportunity but also increase ongoing resource requirements.
A small improvement in efficiency can have a significant financial impact when dealing with a high volume of requests. Similarly, a slight decrease in response time can negatively affect the perceived value of a service.
Microsoft describes its own AI infrastructure as a complete system where hardware and software are jointly optimized for performance, scalability, and energy efficiency. Different processors and accelerators are intended to cover various use cases. However, this does not reveal which specific architecture will replace the capacities previously used at Oracle.
The network becomes an economic production factor
Graphics processors are the focus of attention, but their performance depends on the surrounding system. Data must be transferred between devices, storage, and applications. If these connections are not suited to the workflow, expensive accelerators can be waiting for data instead of performing productive calculations.
This creates an economic utilization problem. The company pays for high-performance hardware but only receives a fraction of its potential benefit. A bottleneck in the network can therefore negatively impact the return on investment of an entire installation.
Against this backdrop, the changes at Iren are significant. The company announced that it had postponed infrastructure services after Microsoft requested changes to the network design. Additionally, it was revealed that Microsoft is increasing its control over certain data center projects belonging to Iren, Nebius, and Nscale.
These developments suggest a more thorough examination of the technical and operational implementation of external resources. However, they do not prove that all the aforementioned providers have comparable problems or that Microsoft fundamentally questions their services.
The connection with Oracle lies on a higher level: Microsoft is apparently not only evaluating how much computing power is available, but increasingly how precisely it fits into its own architecture. Standardized capacity is thus becoming more of a customer-specific, tailored production resource.
Technical changes can be expensive in the short term but beneficial in the long term
Changes to already planned infrastructure projects often cause additional work. Components need to be adapted, installations rescheduled, and dependencies re-coordinated. This can lead to higher costs and delayed revenue.
From the infrastructure customer's perspective, such a delay can still be beneficial. A system that is completed on time but performs poorly for years is not necessarily more economical than a better-tuned solution that becomes available later.
The relevant assessment therefore compares the short-term additional costs of a change with the long-term avoidable operating costs. The greater the eventual scope of use, the more significant initially small technical differences can become.
At the same time, there is a risk of over-optimization. If requirements are constantly changing, completion can be delayed indefinitely. Furthermore, a highly customized system can be more difficult to adapt to other users.
This creates a classic conflict of objectives between perfection and speed. Success lies not in the theoretically best architecture, but in the solution that delivers a reliable economic benefit within realistic timeframes. For operators and customers, jointly managing this conflict of objectives becomes a crucial competitive factor.
A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) - Platform & B2B solution | Xpert Consulting
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:
Microsoft's strategic realignment in the AI market
The billions in investments are increasing the pressure on the return on capital
The scale of these expenditures explains why Microsoft is taking a closer look at external resources. In the fourth quarter of fiscal year 2026, the company reported capital expenditures, including related lease effects, of $41 billion. According to the company, roughly two-thirds of this was for shorter-lived assets, primarily CPUs and GPUs. The remaining portion involved longer-lasting equipment.
This composition is economically important. A data center building can be used for a long time, while servers and accelerators need to be replaced or repurposed more quickly. The invested capital is therefore tied up in equipment with very different technical and economic lifespans.
At the same time, Microsoft Cloud's gross margin reached 65 percent in the quarter in question, falling short of the previous year's level. Microsoft cited factors including the business mix, investments in AI infrastructure, and increasing product usage as contributing to the decline; efficiency gains partially offset these effects.
This does not allow for the derivation of an isolated profitability metric for individual AI applications. However, it becomes clear that sharply rising revenues do not eliminate the cost issue. The more capital flows into the infrastructure, the more important utilization, technical suitability, and the avoidance of redundant capacity become.
The actual write-off risk begins before the technical defect
A server doesn't have to fail to become economically problematic. Its attractiveness can decrease simply because newer systems can perform the same task faster, with less energy, or with lower operating costs.
This is where technical functionality and economic competitiveness differ. A plant can operate reliably for years and still come under pressure because the benchmarks for comparison change.
This creates a timing risk for infrastructure operators: the purchase is paid for today, but the returns are spread over several years. If achievable prices deteriorate faster than expected, the actual return will be lower than planned.
A shorter economic lifespan can therefore be more burdensome than a moderate increase in construction costs. It shortens the period in which an investment must recoup its capital outlay.
The non-renewal of an agreement can reveal such a risk without automatically proving a need for an accounting write-down. This would require information on residual book values, alternative uses, and expected cash flows. In the case of the Oracle capacities in question, these details are not publicly available.
Utilization determines whether expensive hardware earns money
For an operator, it's not just the price per usage hour that matters. The actual number of hours paid for is also crucial. Capital commitment, personnel costs, and numerous infrastructure expenses remain regardless of whether a graphics processor is used productively.
A simplified, explicitly hypothetical example illustrates the relationship: If the paid utilization of an unchanged plant increases from 60 to 80 percent, the same fixed cost block is spread over one-third more productive operating time. The fixed costs per paid hour decrease accordingly by one-quarter.
However, the total unit costs do not develop identically, because variable costs can also increase with higher utilization. Nevertheless, this example shows why capacity utilization is crucial for profitability.
For Oracle, the crucial factor is therefore how quickly the freed-up resources can be re-let. A short-term change between customers would be viewed differently from an economic perspective than a longer period of vacancy.
The type of subsequent usage also matters. High nominal occupancy is of little use if it is achieved solely through substantial discounts or particularly service-intensive customers. What is economically relevant is the combination of occupancy, net revenue, and the actual remaining contribution margin.
Oracle's challenge lies in remarketing, not just in customer loss
The loss or reduction of a major customer initially represents a demand shock for the affected capacity. However, its actual significance depends on the breadth of the remaining customer base and the flexibility with which the offering can be deployed.
Oracle emphasizes continued strong demand even for older graphics processors. Additionally, it was reported that the company achieved a 20 percent premium over the original price for certain contract renewals of GPUs that are approximately four years old.
This information is interesting, but must be interpreted with caution. It does not describe a price trend that can be automatically applied to the Microsoft agreement. Without knowledge of the original contract terms, the included services, and the respective capacity constraints, the comparison remains limited.
Furthermore, a higher price compared to an older, original contract does not necessarily mean a higher profit margin. Energy, financing, maintenance, and other operating costs may have changed in the meantime.
The reliable conclusion, therefore, is not that older hardware will inherently become more valuable. Rather, it is that its market value can vary depending on the application and the customer's situation. Oracle's economic strength is demonstrated by its ability to successfully leverage these differences and profitably reallocate freed-up capacity.
Why specialized AI cloud providers are being watched particularly closely
Companies like Iren, Nebius, and Nscale belong to a group of providers whose business models are strongly linked to the expansion of specialized AI infrastructure. For such operators, large customer relationships can enable the rapid development of extensive capacities.
This also presents a structural risk. If a significant portion of the planned business depends on a few customers, their technical changes and timing decisions will have a particularly strong impact on the company's own development.
The problem isn't limited to a potential loss of revenue. Even a postponement can increase financing costs, delay commissioning, and postpone the start of operational cash flow. Meanwhile, construction work and components may already have been paid for or placed on firm orders.
A large customer agreement therefore does not automatically guarantee complete planning security. What is crucial is what conditions apply to actual use and how technical acceptances, adjustments, and delays are regulated.
From an investor's perspective, it's important not to focus solely on announced capacities. More meaningful indicators include operational facilities, actual usage, reliable cash flows, and a customer mix that limits dependence on individual decisions. A large order backlog and a robust, ongoing cash flow are economically distinct metrics.
The bargaining power of large buyers increases with the selection available
As a major infrastructure customer, Microsoft has several leverage points: its own data centers, diverse hardware architectures, and relationships with multiple external providers. The more credible these alternatives are, the greater its negotiating power.
During an acute capacity shortage, a significant portion of the strength lies with suppliers who can deliver quickly. This balance can shift as more suitable offers become available or a customer expands their own resources.
Then, performance indicators, flexible adjustments, and economic conditions become more important. Providers must not only demonstrate their ability to supply infrastructure, but also prove that their solution is superior to realistic alternatives.
However, this does not mean unlimited interchangeability. Large AI systems cannot be moved arbitrarily between locations and operators. Software adjustments, data transfers, security checks, and tests all require effort.
Negotiating power therefore doesn't stem solely from the number of potential suppliers. It also depends on how practical switching suppliers is. Those who prepare credible alternatives can achieve better terms. Those who are technically tied to a single solution remain limited despite large purchasing volumes.
In-house operation and external procurement fulfill different strategic tasks
The decision between in-house infrastructure and leased capacity is often treated as an either-or question. For large cloud companies, it's more of a portfolio management issue.
Owning their own facilities allows for close coordination of hardware, software, and operational processes. With consistently high utilization, they can offer economic advantages. At the same time, they tie up capital and transfer construction, operational, and obsolescence risks to the company.
External capacity creates additional flexibility. It can accelerate growth, close regional gaps, or absorb peak loads. However, this also entails ongoing payments and dependencies on the partner's execution quality.
The ideal mix changes with demand. A stable base load over the long term can be covered differently than an experimental service with uncertain growth. Similarly, the optimal solution during a market launch may differ from that in later regular operation.
Microsoft's approach can therefore be understood as a possible reorganization of this portfolio. Certain external resources are becoming less important, while others may remain necessary or attractive. A decline in one area does not necessarily mean a decrease in overall infrastructure requirements.
Power, cooling, and site quality determine the value of computing power
The economic benefit of a graphics processor depends not only on the device itself. Power supply, cooling, network connectivity, and reliable operation are prerequisites for transforming theoretical performance into a marketable service.
Especially with older hardware, location can become more important. If a task requires more energy or more devices, differences in electricity costs and cooling requirements have a greater impact. However, it cannot be assumed that every new generation results in lower overall costs in every application.
A low electricity price alone does not automatically make a location attractive. A delayed grid connection can delay commissioning. Inadequate connectivity can limit applications. Complex operating conditions can create additional risks.
This creates a multidimensional competition. An economically viable AI location combines available energy, suitable infrastructure, reliable implementation, and access to demand.
This has an important consequence for evaluating older facilities: their future does not depend solely on the age of the chips. A well-integrated system in a suitable location can remain competitive, while a newer but poorly configured installation will be economically disappointing.
Efficiency gains can accelerate growth, rather than slow it down
Decreasing costs per application are sometimes seen as an argument against the need for further infrastructure. However, this correlation is not necessarily true. If AI becomes cheaper and more reliable, additional applications can become economically viable.
A company that has previously only automated the analysis of a small number of documents can now include significantly larger datasets at a lower cost. A digital assistant can handle more tasks if each additional calculation costs less. This allows more efficient technology to expand overall demand.
Whether this effect more than compensates for the savings per task cannot be determined in general terms. It depends, among other things, on willingness to pay, application quality, organizational hurdles, and available data.
This distinction is crucial for infrastructure investors. Efficiency improvements can enable greater utilization in the long term, while simultaneously putting certain hardware classes under price pressure in the short term. Overall market growth and challenging conditions for individual systems can occur concurrently.
The case of Microsoft and Oracle fits this logic. A more economically efficient distribution of tasks would not be evidence of waning interest in AI. Rather, it could be a prerequisite for offering AI services profitably in the long term, even with increasing usage.
What corporate customers can learn from this development
This process is also relevant for companies outside the cloud industry. It demonstrates that AI infrastructure should not be evaluated in isolation from the respective business process.
When selecting a solution, it's essential to first clarify the required outcomes. Internal knowledge retrieval, automated invoice verification, and interactive customer service each have different requirements regarding speed, accuracy, and availability.
The complete cost structure must then be considered. In addition to model and computation costs, this includes integration, data preparation, monitoring, security measures, and human review. A low usage price can be more than offset by high organizational effort.
Adaptability is equally important. Models and infrastructures change rapidly. Therefore, a solution should not be unnecessarily tied to a specific hardware generation if the business objective can be achieved with other configurations.
This can lead to a pragmatic perspective for small and medium-sized enterprises (SMEs): not every application requires the latest and most expensive technology. What matters is sufficient, reliable performance at a sustainable overall cost. Even older infrastructure can find meaningful applications when faced with differentiated requirements.
Three development paths show what matters now
Based on the current state of information, three plausible development paths emerge, although none of them can yet be considered certain.
In the first case, it is primarily a normal adjustment of the infrastructure portfolio. Microsoft replaces unsuitable capacity, while Oracle finds new users. The immediate financial consequences then remain limited.
In the second scenario, a stronger market segmentation begins. Particularly demanding applications concentrate on more modern, tightly integrated systems. Older capacities are increasingly used for specialized or less time-critical tasks. The overall market continues to grow, but prices and returns develop differently depending on the infrastructure class.
In the third scenario, the process becomes part of a broader investment correction. Several large customers simultaneously reduce their external demand or postpone commissioning. Suppliers with high debt, low customer diversification, and assets that are difficult to repurpose then come under greater pressure.
Individual headlines are insufficient to differentiate between these development paths. Meaningful indicators would include the speed of re-letting, actual revenues realized, changes in occupancy rates, and the time lag between announced and productively used capacity.
The AI market won't shrink just because its standards become stricter
The most economically compelling assessment at present is that of increasing selectivity. Microsoft is apparently scrutinizing more closely which external resources are a technical and financial fit for its own needs. Oracle must demonstrate that freed-up capacity retains viable value with other customers. At the same time, specialized providers are under pressure to reliably translate their growth promises into functioning infrastructure.
This does not automatically contradict further AI growth. More mature markets are characterized precisely by the fact that not every product or service is in equal demand. Differences in quality, cost, and capital commitment become more apparent.
The rationale, therefore, is that the central competition is shifting from the fastest possible accumulation of computing power to the most productive use of the capital employed. Availability remains important, but is no longer sufficient on its own.
For investors, this means separating technological enthusiasm from reliable cash flows. For operators, the ability to profitably deploy hardware across multiple lifecycles is crucial. For enterprise customers, translating technical performance promises into measurable business benefits will be essential.
Microsoft's withdrawal from the relevant Oracle agreement is therefore neither a sure sign of the end of the AI boom nor an insignificant footnote. It highlights what will likely determine its next phase: not how much computing power is built, but which of those units actually earns its keep.
🎯🎯🎯 Data-driven B2B industry hub as a quasi-in-house solution
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:
Your global marketing and business development partner
☑️ 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 wolfenstein@xpert.digital:or simply call me at +49 7348 4088 965. My email address is
I'm looking forward to our joint project.
