Gemini 5.0 is supposed to do more: More AI, less cash? Why Google's infrastructure bet is risky

AI agents instead of chatbots: Why Google is now investing billions

Google's AI plan: Billions for data centers, electricity, and Gemini 5.0

With Gemini 5.0, Google is promising a new generation of AI models designed to bring AI systems closer to autonomous digital employees. Instead of simply answering questions, such systems could plan tasks, utilize tools, and manage entire workflows over extended periods. However, a concrete release date and publicly available performance data are still lacking. At the same time, the financial investment is growing: Alphabet is investing heavily in data centers and infrastructure – so much so that its free cash flow was negative in the second quarter of 2026.

The crucial question, therefore, is not simply whether Gemini 5.0 will be more powerful. Google must demonstrate that the additional intelligence provides measurable benefits for customers and justifies the high costs of computing power, energy, and integration. Increased usage does not automatically translate into increased profits. Only when AI systems reliably perform tasks, integrate seamlessly into workflows, and the cost per successful outcome decreases will technological progress become a viable business model.

Gemini 5.0: Google's billion-dollar bet on the next level of productivity

Intelligence is increasing. So is the billing system

With Gemini 5.0, Google is promising a significantly more powerful model generation, while the economic consequences of the ongoing AI expansion are already becoming apparent: Demand is growing, computing capacity remains scarce, and infrastructure investments are increasingly consuming operating cash flow. However, a concrete release date for Gemini 5.0 and publicly verifiable performance data that would quantify the expected progress are still lacking.

The crucial economic question, therefore, is not whether artificial intelligence can deliver more impressive results. It is whether additional modeling power creates enough reliably usable added value to justify increased spending on data centers, energy, software integration, and control. There is a significant difference between a technically superior system and a profitable business model.

Gemini 5.0 thus represents a development that goes far beyond the competition between individual chatbots. At its core is the potential industrialization of intellectual work: systems should no longer merely formulate answers, but also plan tasks, operate tools, verify results, and continue work processes over extended periods. If this transition succeeds, it won't just change the AI ​​market. The cost structures, competitive advantages, and organizational foundations of numerous companies will also shift.

A promise of performance is not yet proof of productivity

Sundar Pichai bases his expectations for Gemini 5.0 on research advances and internal model evaluations. Google has not yet publicly quantified benchmark results or announced a separate computing budget for the next generation.

This distinction is important. A company may see compelling evidence internally of improved model properties without yet knowing the extent of the economic benefits in later operations. Improved mathematical problem-solving, more reliable tool calls, or longer, continuous work phases can be valuable. However, their benefit depends on which tasks customers can actually delegate and what costs are involved.

Even the concept of a major technological leap requires a precise definition. Progress can mean that a model solves difficult tasks more frequently. It can also consist of achieving the same quality faster, cheaper, or with less human intervention. For business evaluation purposes, these improvements are by no means synonymous.

A model that performs better in a demanding test but takes significantly longer to calculate is not necessarily the more economically attractive solution. Conversely, a seemingly unspectacular improvement in reliability can be particularly valuable if it enables the largely automated automation of a previously labor-intensive process.

The rationale is therefore that the economically relevant progress of Gemini 5.0 should primarily be reflected in lower costs per successfully completed task. Model rankings can provide indications of this, but they do not replace this proof.

Why Google fears underinvestment more

Pichai defends the expansion of the computing infrastructure with several signals: growing demand from the consumer and business sectors, improved modeling capabilities, and long-term contracts with guaranteed revenues. At the same time, he points out that the capacities available today are based on decisions made years earlier.

There's a sound strategic rationale behind this. Companies that provide insufficient capacity in a rapidly growing market risk losing more than just short-term revenue. They can also lose developers, customer relationships, and the opportunity to launch new applications early on. Such losses can't necessarily be compensated for later by adding more servers.

Computing capacity also has an option value. It enables faster responses to technological breakthroughs or unexpectedly high demand. Therefore, when a company invests in infrastructure, it is not only buying a currently utilized production facility, but also the flexibility to pursue uncertain business opportunities.

However, this logic is not an unlimited justification for high spending. If several large suppliers invest simultaneously out of fear of falling behind, what were each plausible individual decisions can lead to overcapacity in the overall market. Strategically necessary investments do not necessarily generate high returns for every market participant.

Google's position should therefore be understood neither as mere waste nor as a risk-free guarantee of future success. It is a bet that the costs of missed market opportunities will outweigh the potential burdens of excess infrastructure.

The investment wave is reaching the cash flow

How substantial this bet has already become is demonstrated by Alphabet's figures for the second quarter of 2026. The company generated $119.796 billion in revenue and $39.069 billion in operating cash flow, but invested $44.924 billion in property, plant, and equipment. This resulted in a negative free cash flow of $5.855 billion for the quarter.

A year earlier, the corresponding capital expenditures amounted to US$22.446 billion. The investment volume has thus roughly doubled within a year.

This development does not automatically mean that the business model has become unprofitable. Free cash flow, simply put, measures the cash remaining after operating activities and capital expenditures. A negative quarter can reflect a particularly intensive expansion cycle. However, it also shows that the pace of investment is now greater than the capacity of ongoing operations to fully finance these expenditures within the same timeframe.

At the same time, Google Cloud's revenue grew by 82 percent year-over-year in the second quarter of 2026, reaching approximately $24.8 billion. This development provides a substantial demand signal, but does not, in itself, prove the profitability of all additional AI investments.

A proper valuation must also avoid attributing all group investments to Gemini 5.0. The published figures encompass Alphabet's broader technical infrastructure and do not show separate expenditures for this single model.

The crucial question, therefore, is whether additional capacities will generate sufficient operating income and cash to cover their acquisition, operation, and subsequent renewal.

High usage does not equal high profit

Gemini 4 Argon processed nearly 25 trillion tokens in the first 72 hours after its wider release. This was roughly three times the volume achieved by Gemini 3 in a comparable launch period. At the same time, Google reported capacity bottlenecks and was working on more efficient alternatives to expand access.

This scale demonstrates considerable interest. However, its economic significance is easily overestimated. Tokens are units of processing, not uniform products with consistent customer value. A lengthy document, an extensive internal problem-solving process, and numerous short answers can each generate large quantities, even though their economic value varies considerably.

A triple token volume does not automatically translate into triple revenue or triple profit. Factors such as the distribution between inputs and outputs, applicable tariffs, potential discounts, the reuse of stored content, and the actual operating costs would be relevant.

Furthermore, the nature of a market launch must be considered. Developers often test new models intensively, compare capabilities, and repeat tasks. Such activities provide valuable insights into attention and potential demand. Whether this translates into sustained paid usage can only be assessed over longer periods.

Key performance indicators (KPIs) such as recurring revenue, customer loyalty, expenditure per productive work process, and contribution margin after computing costs would therefore be particularly meaningful. A high number of tokens can indicate a successful market launch, but it is not yet reliable proof of a sustainably viable revenue structure.

The price is not the only factor

For Argon, Google initially quotes $2 per million input tokens and $10 per million output tokens. After the introductory phase ends on November 30, 2026, the regular prices of $4 and $20 respectively will apply; cached inputs will be billed separately and at a significantly lower rate.

This pricing structure makes it clear why a comparison based solely on the price per million tokens remains incomplete. Crucial factors include how many processing steps a model requires, how often an attempt fails, and whether additional tools or models are needed.

A low-cost system can become economically expensive if employees have to correct numerous results. Conversely, a more expensive model can be advantageous if it reliably completes a task on the first attempt, thereby reducing the need for review, repetition, or delays.

For autonomous agents, the cost calculation becomes even more complex. In addition to the actual model execution, expenses arise for runtime environments, external data access, storage, monitoring, security mechanisms, and the integration of operational systems. The time of qualified personnel who approve results or correct errors must also be factored into the calculation.

For enterprise customers, a total cost analysis is therefore essential. The relevant unit is not the individual answer, but the fully completed business transaction. Those who ignore this distinction may choose a seemingly inexpensive AI solution that ultimately increases the overall cost of the workflow.

This, in turn, presents Google with the opportunity to sell not just model access, but an integrated execution and control environment. Whether customers will accept permanently higher prices for this depends on demonstrable benefits.

The infrastructure is lagging behind the model

According to Pichai, Google plans its computing capacity two to three years in advance. The current offering thus reflects decisions made in 2024, while planning for land and energy supply already extends to 2028 and 2029.

This time lag is one of the central problems of the AI ​​economy. Software capabilities can change rapidly. Land, network connections, buildings, cooling technology, and high-performance computing systems, on the other hand, follow longer development and deployment cycles.

This creates a structural forecasting risk. Companies have to tie up capital before they know exactly which models will be available later, which applications will prevail, and what prices customers will pay. They are investing in future demand whose technical prerequisites are still partly being researched.

The coordination of components with varying lifespans is particularly challenging. A plot of land can be used for decades, while the demands on computing hardware and power density change much more rapidly. Therefore, an economically viable location is not automatically optimal for every subsequent generation of equipment.

Even short-term peak loads do not solve this problem. A bottleneck can occur after a release despite adequate long-term capacity planning. Conversely, currently full capacity utilization does not prevent the development of excess capacity later on.

The quality of an investment decision is therefore not solely determined by the pace of expansion. It is equally reflected in scalability, technical flexibility, contractual security, and the ability to reallocate resources between different applications.

Electricity is becoming a competitive factor

According to the International Energy Agency's updated estimate, global data center electricity consumption was around 485 terawatt-hours in 2025. Their central projection anticipates approximately 950 terawatt-hours by 2030, representing roughly a doubling within five years.

This figure applies to all data centers and should not be equated with the consumption of a single provider or solely with generative AI. Furthermore, the projection describes an expected development, not a fixed consumption figure.

For the competitiveness of individual locations, however, the local availability of additional power is more crucial than the global share. A data center needs electricity in a specific location, at specific times, and with high reliability. A theoretically sufficient national electricity production does not guarantee a suitable grid connection.

This shifts part of the competition from model development to physical infrastructure. Energy supply, connection speed, permits, and cooling can determine how quickly a provider can deploy new services.

Economically, a distinction must be made between the electricity price and the total costs. A location with cheap electricity can be unattractive if connections are lacking for years or if additional safeguards become expensive. Conversely, a higher initial energy price can be partially offset by faster commissioning and better utilization.

This presents a strategic challenge for Germany and Europe. Those who want to promote AI applications must not only strengthen research and software expertise, but also create the conditions for their reliable operation.

Efficiency can increase consumption

Increased efficiency doesn't necessarily lead to lower overall costs. If a single AI task becomes cheaper, new applications that were previously too expensive can become economically attractive. Companies might then have more documents reviewed, more software variants tested, or additional processes monitored.

This relationship is particularly important for Google's investment logic. Lower costs per use can simultaneously improve margins and expand demand. Whether the overall infrastructure requirement increases or decreases as a result depends on how strongly additional usage responds to lower costs.

Technical efficiency and economic value creation should not be confused. A model can generate the same answer with less energy without the customer deriving any significant benefit. Conversely, an energy-intensive task can be economically advantageous if it replaces a high level of manual effort or accelerates a valuable development process.

Problems arise when additional computing power is primarily used for unproductive repetitions, uncontrolled agent loops, or oversized models. In such cases, usage increases without a corresponding increase in productivity.

The economically compelling development would therefore not simply be the highest possible computing volume. It would be an increasing number of value-adding tasks per euro invested, per occupied computing unit, and per kilowatt-hour consumed.

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

The skipped intermediate step is systematic

The development of Gemini 4 Argon involved a reorganization of the model strategy. Gemini 3.5 Pro was not released as originally planned; instead, insights gained from the rapidly successive Flash releases were incorporated into Argon's post-training. Once its pre-training was complete, the remaining lifecycle of an additional 3.5 Pro model appeared too short.

The previous lag in agent-based programming led not only to technical adjustments but also to a change in resource allocation. Within a few months, the further development of Flash and subsequently Argon became the focus.

From a business perspective, such a step is ambivalent. On the one hand, abandoning a product that is already well-developed can be perceived as a failure. On the other hand, it would also be problematic to invest additional resources in a publication simply because high development costs have already been incurred.

Past expenditures cannot be recouped simply by launching a product. The crucial question is whether the remaining necessary expenditures are justified by the expected future benefits.

However, this dynamic has a downside for customers. Rapid changes in model generations, performance levels, and product plans increase evaluation and migration costs. Therefore, companies need not only technological advancements but also reliable interfaces, transparent changes, and sufficient transition periods.

A rapid innovation cycle is only an advantage if it does not continuously undermine the stability of productive applications.

Chromium demonstrates the more interesting use case

When used in Chromium, Argon supported the separation of fundamental website functions from a legacy codebase of approximately six million lines within a codebase of around 40 million lines. Google describes the reduction of the necessary restructuring work from months to weeks.

The economic value of such an example does not lie in the fact that a model generates large amounts of code. More interesting is the ability to understand existing structures, to consistently implement changes across numerous files, and to embed the work in a sophisticated development process.

Up-and-coming enterprise software creates costs not only due to missing features, but also due to dependencies, legacy custom solutions, and difficult-to-maintain architecture. If AI can reliably modernize such systems, it can reduce investment backlogs and accelerate further development.

However, the transferability remains limited. A successful project in a technologically advanced organization does not prove that the same effect can be achieved in every company with incomplete documentation and inconsistent data structures.

Google's claim that agent-assisted code submissions have increased by more than 35 percent and change retractions have decreased by almost a third initially describes internal observations. This does not necessarily imply a similarly high increase in productivity across the entire software development landscape.

For a reliable assessment, additional factors would need to be considered: testing effort, error rates in operation, maintainability, safety quality and the speed at which actually usable functions reach the user.

Minecraft demonstrates perseverance, not efficiency

Four agents based on Gemini 4 Argon completed Minecraft up to the final boss without any human intervention. The documented run lasted 27 hours and 14 minutes; during this time, the players died a total of 605 times.

The computational costs mentioned in connection with the experiment were less than US$2,000. However, this figure does not describe the complete business expenses, including model development, experimental setup, and all accompanying infrastructure.

What's technically interesting is the longer-term coordination: The agents had to pursue intermediate goals, procure resources, exchange information, and react to setbacks. This is closer to complex work processes than a single question with an immediate answer.

The numerous failed attempts also demonstrate why a game is not directly comparable to a business. In a virtual environment, repeated errors can be part of a successful search strategy. In real-world business processes, however, similar failed attempts can trigger payments, jeopardize delivery dates, or expose sensitive information.

Economic progress, therefore, does not simply consist of a system eventually reaching a goal. It lies in achieving this goal within acceptable costs, timeframes, and risk limits.

The experiment thus indicates increasing capacity for action over longer periods. It is not proof that any operational task can already be delegated unsupervised and profitably.

Self-improvement begins with limited loops

Google is already using AI in the development of further models. Pichai describes collaborative loops between researchers and agents, but also emphasizes that fully autonomous recursive self-improvement has not yet been achieved.

The described functions include suggestions for training recipes and the evaluation of saved training progress. Such checkpoints help to assess how a model changes during its development and which variations should be investigated further.

This is economically significant because research itself is becoming an application area of ​​automation. If systems prepare experiments, improve code, process results, or identify promising approaches, the yield from limited research resources could increase.

However, this does not necessarily lead to an explosive, infinitely accelerating development. A faster partial step is of little help if the overall process remains limited by available computing power, data quality, practical experiments, or human decisions.

Evaluating new results is also challenging. A system may improve in known tests without necessarily performing more reliably in unknown situations. Therefore, as training recommendations and performance assessments become increasingly automated, the importance of independent audits grows.

The sober assessment is this: AI-supported research can create a significant competitive advantage. However, it should not be confused with the already achieved state of a machine that improves itself independently and indefinitely.

A research advantage can multiply competitive advantages

When high-performance models aid in the development of their successors, the leading position in the AI ​​market takes on added significance. A good model would then not only be a marketable product, but also a productive tool for further research.

This could create a cumulative advantage. Better tools facilitate experimentation, more successful experiments improve the next generation of models, and these in turn support further research. Such a mechanism would explain why vendors invest heavily in top-of-the-line models even when a large portion of current customer demand could be met with smaller systems.

The danger lies in treating this potential cycle as a proven development. It still needs to be demonstrated whether faster programming or improved experimental designs actually accelerate the overall research process more significantly.

Furthermore, knowledge can diffuse in various ways: through scientific publications, personnel changes, specialized tools, or alternative technical approaches. Therefore, an advantage need not be permanent, nor should it be the same across all fields.

The question remains central to market structure. If high-quality AI research increasingly requires advanced AI, the barriers to entry for new competitors could rise. Conversely, if efficient, specialized models enable similar research yields, development could be more widely distributed.

The assistant becomes part of the work environment

Google plans to further develop Gemini into a personal, proactive assistant that performs tasks in the background over extended periods. Services like Search and Gmail will be connected through a shared agent layer, without requiring all products to be merged into a single application.

Economically, this shifts the function of the user interface. A search engine primarily helps to find information. A permanently available agent could additionally decide which information is relevant, which steps to take, and which tools to use.

For users, this promises less coordination effort. A work order wouldn't need to be completely re-explained every time the process shifts between research, communication, and execution. The economic value lies in the continuity of the process.

This creates a strategically attractive position for the platform provider. Whoever accepts the order and organizes its execution can become the central access point for numerous other services.

This raises the question of the criteria an agent uses for selection. Does it prefer the cheapest solution, the highest quality, the already connected provider, or an economically linked offer? The more decisions users delegate, the more important transparency and traceability become.

This could shift the focus of competition from the visibility of individual websites to their consideration within automated decision-making processes.

Virtual workspaces only solve partial problems

Google envisions the same basic model and runtime environment for both business and private users, although the specific usage context differs. Business tasks are intended to be executed in isolated virtual environments with monitoring, control, and traceability.

A shared technical foundation can pool development costs and bring improvements to different products more quickly. However, it does not mean that all customer groups should have the same data spaces, permissions, or contract terms.

Isolating individual tasks is an important security measure. It can limit which resources an agent can access and what the consequences of an error are. However, it does not replace a clear decision about what a system is allowed to do in the first place.

An agent who solely generates a report requires different permissions than a system that modifies customer data or prepares payments. The greater the scope of action, the more crucial tightly restricted access rights, approvals, and auditable logs become.

The response to uncertainty must also be economically sound. Too many queries negate the benefits of automation. Too few queries can lead to costly mistakes.

The quality of an agent platform is therefore not only measured by how much it can do autonomously. Equally important is whether it reliably recognizes when it needs to stop and involve human responsibility.

The search business is facing internal competition

For Google, the agent strategy presents both an opportunity and a challenge. An assistant that directly performs tasks can enhance existing products. However, it can also change current search processes, on which a portion of the company's economic value creation is based.

If users perform fewer individual search steps and instead delegate a goal, monetization strategies must adapt. More subscription-based offerings, usage-based fees, or revenue models centered on actual transactions are all conceivable.

None of these options guarantees that the productivity of previous processes will remain unchanged. A particularly useful agent could combine several previous interactions and simultaneously incur higher computational costs.

The strategic challenge, therefore, lies in creating new uses without underestimating the economic transition. Not every improvement in the user experience immediately increases revenue. In the long run, however, it may be necessary to partially displace one's own existing offering before a competitor does.

Therefore, an isolated look at model subscriptions is insufficient for evaluating Gemini 5.0. It would also be relevant to consider how the system affects the interaction between search, cloud services, enterprise software, and other services.

Small and medium-sized businesses don't need a record-breaking machine

For medium-sized businesses, the most interesting question is not whether Gemini 5.0 is among the most powerful models. The crucial question is which specific processes can become more reliable, faster, or more cost-effective as a result.

In industry and logistics, examples include the preparation of technical documentation, the analysis of deviations, the preparation of customer communications, or the review of extensive documents. Tasks where results can be transparently monitored and incorrect decisions do not have immediate, irreversible consequences appear particularly suitable.

However, a more powerful model generation does not eliminate unclear responsibilities, conflicting master data, or missing interfaces. If information is poorly structured and processes contain numerous special cases, integration remains complex.

The business-oriented approach should therefore begin with a reliable baseline measurement. How long does a process take today? What errors occur? Which employees check which results? Only then can it be assessed whether the AI ​​solution actually leads to improvements.

A smaller or more specialized model can be more economically attractive than the most powerful system available. A well-designed architecture could handle routine tasks cost-effectively and only pass difficult cases on to a more powerful model.

The better strategy, therefore, is not to immediately deploy every new model generation across the board. It is to combine the appropriate performance level with well-designed processes.

Productivity doesn't happen automatically

Research on AI-supported software development demonstrates how difficult it is to measure actual productivity effects. METR initially found a slowdown for experienced open-source developers using data from the first half of 2025; however, a later study could not be interpreted as a reliable current impact estimate due to significant selection and measurement problems.

This does not imply that AI inherently slows down developers, nor does it mean that its economic benefits have been definitively proven. Rather, the results demonstrate how significantly the task, experience, tool usage, and experimental setup can influence the outcome.

Especially with agents, the division of labor also changes. Employees can take on other tasks during execution. Simply measuring the elapsed time may then fail to adequately capture either the additional coordination effort or the concurrent benefits.

Higher production volumes do not automatically translate into higher productivity. More code, reports, or analyses can be valuable, but they can also lead to additional testing and maintenance work.

From a macroeconomic perspective, time savings only reach their full potential when companies use the freed-up resources effectively. This requires demand, organizational adjustments, and expertise.

The most plausible expectation, therefore, is not a uniform increase in productivity across all industries. Significantly different effects between companies, activities, and quality requirements are more likely.

The decisive metric remains the benefit

Three economic directions can be distinguished for further development. In the best-case scenario, Gemini 5.0 improves reliability so significantly that companies can actually delegate more complex tasks. In that case, additional infrastructure and increased usage would be justified by a correspondingly higher customer benefit.

In a mixed scenario, technical capabilities increase significantly, while integration, control, and organizational adaptation progress more slowly. Google could then experience strong growth but simultaneously face high investment costs and limited additional return on capital.

In the worst-case scenario, computing needs and competitive pressure increase faster than willingness to pay. Technological progress would still be possible, but it would benefit customers more, while providers would bear a large share of the costs.

Which direction prevails cannot be determined based on a model announcement alone. More meaningful metrics would be sustained paid usage, cost per successfully completed transaction, reliable utilization, and the development of cash flow after investments.

Google's infrastructure bet has a sound strategic basis. The transition to longer-duty agents could open up new markets and transform existing products. Current demand signals are relevant, but they don't replace a long-term economic analysis.

Gemini 5.0 would therefore need to deliver more than simply outperform its predecessor in selected tests. The real breakthrough would be achieved if the additional intelligence not only consumed more computing power but also resulted in demonstrably better economic outcomes.

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