Progress Is No Longer Limited to Generating Text or Images
Recent advances in artificial intelligence are no longer simply about producing better-written answers or more realistic images. The most advanced models can now chain together operations, use tools, browse files, modify code, conduct research and pursue an objective through a sequence of steps.
This evolution is particularly visible in software development, where systems can now handle tasks that go far beyond generating a few lines of code. Evaluations conducted by METR show a rapid increase in the duration and complexity of tasks that agents can complete at a given level of success.
In its “time horizon” metric, which measures how long a task would take a human expert when an agent has a 50% chance of completing it successfully, METR historically observed an approximate doubling every seven months. In 2026, the most advanced models began reaching the limits of some benchmarks used to measure this progress, making further evaluation increasingly difficult.
These results should, however, be interpreted with caution. The evaluations focus primarily on software development, machine learning and cybersecurity tasks with relatively well-defined objectives. They do not mean that a model can replace a professional for several hours in any occupation or independently manage a real-world project of equivalent duration.
Agents Can Now Complete Some Tasks That Once Required Hours or Days
The change is nevertheless becoming difficult to describe as merely incremental. In its report published in May 2026, METR found that the highest-performing agents evaluated earlier in the year were capable of completing some software projects that would have required hours or days of human work. On an early version of another benchmark, MirrorCode, some systems even completed software reimplementation tasks estimated to require several weeks of human work.
This does not mean that they can now autonomously perform several weeks of ordinary professional work. The evaluated tasks are precisely defined and their results can be verified. As the environment becomes less structured, success criteria become harder to measure or decisions require more judgment, performance declines.
More Capable Does Not Always Mean More Reliable
This is one of the defining characteristics of the current situation. Capabilities are advancing rapidly without reliability necessarily improving at the same pace across every domain. The same system may successfully complete a complex technical task and then fail at an apparently simpler step, misunderstand an instruction or produce a convincing but incorrect result.
METR's research illustrates this gap. Agents evaluated in 2026 perform particularly well when tasks are verifiable and clearly defined, but their results remain significantly less robust on more ambiguous problems or those requiring greater judgment. AI capabilities therefore remain highly uneven depending on the nature of the work being performed.
This distinction is essential to understanding what comes next in the race. The industry is no longer focused solely on producing models capable of generating better content. It is seeking to improve their ability to reason, use tools, operate for longer periods and complete increasingly complex tasks with less human intervention.
The Progress Is Real, but It Tells Only Part of the Story
It would be difficult to argue that capabilities are simply stagnating: across several categories of measurable tasks, progress is rapid and well documented. But this progress does not mean that every limitation has disappeared, nor that each new generation immediately makes the previous one obsolete.
This is precisely where the central paradox of this report begins. While existing systems can already perform tasks that were beyond their reach only a few years ago, the leading players continue to invest in going much further. The pursuit of additional capabilities is driving a considerable increase in demand for computing power, infrastructure and capital.
The next question is therefore no longer simply whether AI is progressing, but why that progress now requires an industrial race on such a massive scale.
Why the Race for More Power Continues
Additional Computing Power Continues to Improve Capabilities
The current race is not based solely on the idea that a larger model is automatically better. It reflects a broader reality: increasing the resources devoted to training, improving architectures and allowing models to use more compute when generating responses continue, in many cases, to produce gains in capabilities.
Computing power now plays a role at several stages. It is required to train models, but increasingly also when they are being used. Reasoning models can devote more compute to certain complex problems, perform additional intermediate steps, use tools or verify parts of their work before producing an answer.
This changes the logic of the race. It is no longer simply a matter of building an extremely expensive model once and then operating it at relatively low cost. Some performance improvements can also depend on the resources used each time the model is called. Training and inference therefore represent two distinct sources of demand for computing power.
The Gains Become Particularly Important When AI Can Operate for Longer
This additional computing power is being pursued because the objective of AI developers has itself evolved. Producing a correct answer to an isolated question is no longer the only benchmark. Systems are increasingly being used as agents capable of chaining actions, recovering from errors, interacting with tools and pursuing a task through to completion.
METR's measurements show that the duration of tasks frontier systems can complete at a given success rate has increased significantly. For public models released since 2024, METR estimated in 2026 a trend corresponding to a doubling of the “time horizon” approximately every 105 days. This metric remains limited primarily to technical and well-defined environments, but it illustrates why even incremental improvements can have significant consequences: they may not simply produce a better answer, but enable a system to complete a longer and more complex task.
This is one reason developers continue to pursue additional capabilities even when existing models already appear highly capable. Moving from a system that can provide occasional assistance on a task to one capable of handling an entire sequence of operations can represent a much greater change than an improvement of a few percentage points on a benchmark.
Competition Between Major Players Is Also Driving the Acceleration
This technical dynamic is accompanied by an industrial one. When one company significantly improves the capabilities, cost or speed of its models, its competitors must decide whether they can afford to slow down. AI laboratories, cloud providers and chip manufacturers are therefore engaged in a competition in which today's capabilities matter, but the ability to train and operate the next generations of systems is itself becoming strategic.
This competition makes a voluntary slowdown in the race for more power difficult. Even when an existing model may already be sufficient for many applications, a company that chooses not to develop the next generation risks allowing a competitor to achieve better performance, lower costs or enable applications that remain impossible with current systems.
But More Compute Does Not Guarantee Proportional Progress
The relationship between computing power and performance is neither simple nor unlimited. Giving a model more resources can improve certain results without producing the same gains everywhere. Research has even identified situations in which allowing a model to reason for longer can reduce performance, for example when it becomes distracted by irrelevant information or locks itself into an incorrect interpretation of the problem.
The race is therefore no longer simply about adding ever more compute. It also concerns the efficiency of architectures, the quality of data, training methods, the tools available to agents and the ability to use computing resources where they actually produce a meaningful gain.
The Question Is Becoming as Industrial as It Is Technological
As long as additional resources continue to unlock new capabilities, the pursuit of more computing power retains a technical rationale. But continuing this trajectory at scale now requires far more than algorithms: chips must be manufactured and purchased, data centers must be built, powered and cooled, and their operation must be financed.
This is where the spectacular progress of AI encounters a much more physical reality. A model is software, but the computing power required to train and operate it is not. The faster the race accelerates, the more it depends on industrial infrastructure capable of keeping pace.
The question is therefore no longer simply how far models themselves can progress. It is also how far the infrastructure required to sustain that progress can expand.
What This Expansion Requires in Infrastructure
AI Progress Is Now Materializing in Massive Infrastructure
Behind the models accessible through a web browser lies an increasingly substantial infrastructure. Training the most advanced systems and then operating them for millions of users requires thousands of specialized processors connected together, high-speed networks, significant storage capacity and data centers capable of continuously powering and cooling the entire system.
The evolution of hardware illustrates this change in scale. New AI platforms are no longer designed around a single processor or even a single server, but at the rack level and then across entire groups of racks operating as a unified system. Nvidia's Vera Rubin platform, which entered production in 2026, illustrates this evolution: its POD configuration can bring together more than a thousand GPUs within a single computing system.
This concentration directly reflects emerging uses. Models must be trained, fine-tuned, evaluated and then operated at scale, while agentic systems can multiply the number of operations required to complete a single task. Infrastructure must therefore keep pace not only with the growing number of users, but also with the increasing amount of compute consumed by certain applications.
Electricity Is Becoming One of the Main Constraints on This Expansion
This infrastructure has an immediate consequence: it consumes electricity. According to the International Energy Agency, data centers consumed approximately 485 TWh of electricity worldwide in 2025. In its central scenario, that consumption would reach around 950 TWh by 2030, nearly doubling in five years and representing about 3% of global electricity consumption.
Not all of this increase comes from artificial intelligence. Data centers also host cloud computing, conventional digital services, storage and many other applications. But electricity consumption from data centers specifically oriented toward AI is growing much faster: the IEA projects that it could triple between 2025 and 2030.
The issue is not simply the total amount of electricity available worldwide. Data centers are concentrated in certain regions and can require enormous amounts of power in a single location. An electricity grid can therefore encounter local constraints long before any global shortage of electricity exists.
The Power Grid Cannot Expand at the Pace of a Processor
This is where an important gap emerges between the digital industry and the infrastructure required to support it. A data center can be built within a few years, while new transmission lines, power plants, substations or grid connections often require much longer development timelines.
This gap is already visible in the United States. In the region served by PJM, one of the country's largest electricity systems, demand from data centers is expected to increase sharply by 2030. Their expansion has become significant enough that the grid operator is now seeking to better track these facilities and anticipate their power consumption.
AI expansion therefore depends on infrastructure that extends far beyond the technology industry itself. Producing more GPUs is not enough if the data centers intended to house them cannot be connected quickly enough, if the grid cannot transmit the required electricity or if additional generation capacity is not available in the right locations.
The Race Is Also About Efficiency, Not Just Consumption
These growing requirements do not mean that every new generation of hardware consumes more energy to perform the same task. A significant part of the technological competition is instead focused on obtaining more useful compute from each unit of energy consumed.
Manufacturers are simultaneously working on processors, memory, interconnects, cooling systems and the overall architecture of data centers. Nvidia, for example, says its Vera Rubin generation delivers substantial improvements in performance per watt compared with the previous generation. Software advances can also reduce the amount of compute required to achieve the same result.
These efficiency gains are essential, but they do not guarantee a decline in total electricity consumption. If the cost of performing a task falls while the number of models, users and operations grows even faster, overall demand can continue to rise despite individual systems becoming more efficient.
AI Is Entering a Phase in Which Its Growth Depends on the Physical World
The AI race is therefore no longer taking place only inside laboratories. It now depends on component supply chains, industrial capacity, land, data centers, electricity grids, energy production and access to capital capable of financing this infrastructure before the revenue it will ultimately generate is fully known.
This does not mean that the expansion will necessarily stop. Investment remains substantial and new infrastructure continues to be announced and built. But a fundamental difference has emerged compared with the early years of generative AI: increasing capabilities now requires technological, industrial, energy and financial challenges to be addressed simultaneously.
This is precisely where the first tensions are beginning to emerge. Some projects are moving forward, while others are being scaled back or delayed, as financing costs and questions about return on investment become increasingly important.
The First Economic and Operational Limits
The Challenge Is No Longer Just Financing Expansion, but Justifying Its Returns
Investment in artificial intelligence continues to increase, but the way it is financed and evaluated is beginning to change. As long as the leading players could fund a large share of their infrastructure through their own cash flows, the race could largely be viewed as a very long-term industrial bet.
The growing scale of data center, processor and energy requirements is gradually changing that equation. An increasing share of the infrastructure now relies on debt, very long-term contracts and financial arrangements whose profitability depends on demand for computing power remaining high for years.
In 2026, this issue became more visible in financial markets. Companies whose expansion relies heavily on debt and long-term commitments are more exposed to questions about the pace of data center construction, their future utilization rates and the revenue required to make these infrastructures profitable.
This does not mean that AI investments have become unprofitable or that a broad reversal is underway. Major cloud providers continue, on the contrary, to emphasize the strength of demand. But every massive investment now comes with an additional question: will future revenue grow quickly enough to justify the infrastructure being built today?
In Business, Having Capable AI Is Not Enough to Create Value
A different limitation appears within the organizations using these technologies. A model can work correctly and the project can still fail. Organizations also need usable data, integration with existing software and processes, mechanisms to monitor results, trained teams and a clear understanding of what the technology is actually improving.
Data published by Gartner in 2026 illustrates the scale of this gap. By the end of 2025, at least half of generative AI projects had been abandoned after the proof-of-concept stage, notably because of insufficient data, rising costs, inadequate risk controls or difficulty demonstrating economic value.
This abandonment rate shows that the challenge no longer lies only in gaining access to the technology. It also emerges when attempting to turn a promising experiment into an application that is reliable, integrated and profitable enough to be deployed sustainably.
The Most Ambitious Uses Encounter Reliability Limits More Quickly
The problem becomes particularly visible when AI is asked not merely to assist a user, but to act autonomously across complex processes. The more steps, exceptions, decisions or difficult-to-reverse consequences a task involves, the greater the risk that an error at an intermediate stage will compromise the final result.
In IT infrastructure, Gartner reported in April 2026 that only 28% of the use cases studied fully achieved their return-on-investment objectives, while 20% failed. Difficulties were observed in particular with ambitions involving self-healing systems and agent-managed workflows when the actual capabilities of AI had been overestimated.
This does not call into question the applications that already work. The same research shows that positive results emerge when AI is integrated into clearly defined processes and tied to a measurable objective. The limitation therefore appears to lie less in the technology's inability to create value than in the gap between what organizations ask it to do and the conditions under which it can currently operate with sufficient reliability.
A Technology Can Advance Very Quickly While Still Facing Economic Limits
This is essential to understanding the current situation. Financial difficulties, an abandoned project or a disappointing return on investment do not necessarily mean that the technology itself has stopped progressing. Conversely, increasingly capable models do not guarantee that every piece of infrastructure built to operate them will be profitable.
Both developments can coexist perfectly well: technical capabilities can continue to improve while some business models, infrastructure projects or applications prove too expensive, premature or poorly suited to their intended purpose.
This is precisely what makes the current period significant. The first phase of generative AI was primarily defined by the discovery of new capabilities. The next phase increasingly requires determining which of those capabilities are actually worth industrializing, at what cost and with what level of risk.
Projects Delayed, Scaled Back or Abandoned: What Do These Setbacks Show?
The First Visible Constraint Is Not Demand, but the Ability to Build Fast Enough
The first slowdowns observed in AI expansion do not necessarily reflect weaker demand. In many cases, projects are first encountering a much more concrete difficulty: turning billions of dollars in investment into fully operational data centers.
In 2026, Amazon, Microsoft, Alphabet and Meta are expected to spend approximately $630 billion on data centers and equipment required for AI, according to Morgan Stanley estimates reported by Reuters. Across the eleven largest cloud and infrastructure providers, planned spending would reach approximately $811 billion.
But having access to capital is no longer enough. Projects must secure land and permits, obtain connections to the power grid, receive transformers and cooling equipment, and then accommodate generations of chips that sometimes evolve faster than the buildings designed to house them.
Some of the current delays therefore stem from a paradox: the industry has considerable demand and enormous budgets, but the physical world in which this computing power must be installed cannot necessarily expand at the same pace.
Access to Electricity Is Beginning to Directly Affect Some Projects
The energy constraint is no longer theoretical. In the United States, requests to connect very large data centers to the grid have reached levels in some regions that existing electricity systems cannot accommodate without additional generation and transmission capacity.
In the region served by PJM, which supplies electricity to approximately 67 million people, data center power demand could triple to around 29 GW by 2030. The pressure has become significant enough for the grid operator to establish a dedicated registry to track their locations, expected commissioning dates and ramp-up schedules.
PJM had also proposed allowing some new data centers to be temporarily disconnected during critical conditions if they did not bring enough additional power capacity of their own. The U.S. federal energy regulator rejected that part of the proposal in October 2026, while leaving PJM the option of submitting a revised proposal.
The fact that such a measure is now being discussed illustrates the change in scale. Processor availability is no longer the only determining factor: guaranteed access to several hundred megawatts, or even more, can now determine the timetable and viability of a project.
Some Companies Are Therefore Beginning to Secure Their Own Power Generation
Faced with long grid-connection timelines, several companies are seeking to bypass the constraint by directly securing new generation capacity. In October 2026, American Electric Power, for example, announced approximately $4 billion in investment in a 702 MW fuel-cell project designed to supply Google in Wyoming under a twenty-year agreement.
xAI has taken a different approach to rapidly power its infrastructure. For its Colossus 2 project in the Memphis area, the company installed dozens of gas turbines to provide off-grid electricity capacity. This strategy, however, has led to regulatory and environmental disputes over permits and emissions from the installations.
These examples therefore do not show that expansion is stopping. On the contrary, they demonstrate how far some companies are willing to go to sustain it. But they also reveal that building additional computing capacity can now require directly solving an energy generation problem.
Financial Risk Increases When Infrastructure Is Committed for Decades
Another constraint is less visible on construction sites: a considerable share of future capacity has already been contracted. An analysis of financial filings conducted by Reuters in August 2026 found that Microsoft, Meta, Oracle, Amazon and Alphabet together had approximately $1.09 trillion in future commitments related to leases that had not yet begun, primarily for data centers.
These amounts do not represent $1.09 trillion in debt that is immediately due. They are contractual payments spread over many years, and their accounting treatment is different. Their scale nevertheless shows that a significant share of future expansion has already been committed before the corresponding facilities are even operational.
Oracle provides a particularly clear example of this risk. The company reported approximately $260 billion in lease commitments that had not yet commenced, mainly related to data centers scheduled to come online between 2027 and 2029. The company itself warns that the duration and terms of these leases may not exactly match those of its customer contracts.
If demand continues to grow as expected, these commitments will provide the capacity required for expansion. If demand slows or shifts toward different types of infrastructure, however, they could become a burden that is difficult to reduce quickly.
CoreWeave Shows What Happens When Growth Depends Heavily on Debt
CoreWeave provides a particularly clear example of this tension. The company expanded very rapidly by financing AI-specialized infrastructure before receiving the revenue generated by its operation.
In October 2026, CoreWeave still had a backlog of more than $100 billion, and demand for its capacity remained strong. Yet its shares, and especially its bonds, fell sharply as investors became concerned about its debt levels and the pace of data center construction.
Reuters estimated its liabilities at approximately $50 billion at the time. The yield on some bonds maturing in 2030 approached 16%, indicating that markets were pricing in a significantly higher level of risk.
The case is particularly interesting because the problem is not a lack of customers. CoreWeave has major contracts with several of the largest players in AI. The difficulty lies in the gap between the money that must be committed immediately to build infrastructure and the revenue that will be received gradually as that infrastructure is operated.
There Is No Broad Retreat Yet, but the Margin for Error Is Shrinking
Taken individually, none of these developments demonstrates that the expansion of artificial intelligence is coming to an end. Investment remains enormous, new data centers continue to be announced, and demand for computing power remains high.
What is changing is the margin for error. A project may have processors and customers but lack electricity. It may have access to electricity but face a wait of several years for a grid connection. It may be technically viable but become far more expensive as interest rates rise. And infrastructure that is profitable at very high utilization rates may become much less attractive if future demand has been overestimated.
The setbacks observed so far therefore do not amount to the collapse of a technology. They reveal something more subtle: the race is entering a phase in which announcing new capacity is no longer enough. That capacity must now be built, powered, financed and kept sufficiently utilized for long enough to justify its cost.
But this raises another question. While the industry is mobilizing hundreds of billions of dollars to further increase available capacity, are businesses and users already making full use of what today's models are capable of?
Are We Really Using the Capabilities Already Available?
The Technological Frontier Is Advancing Faster Than Real-World Adoption
There is a significant gap between what the most advanced artificial intelligence systems can accomplish and the way they are integrated into most organizations. Models can now work with documents, analyze data, produce and modify code, use tools and chain together multiple operations to achieve an objective. Yet having access to these capabilities does not mean they are immediately integrated into working processes.
A global survey published by Gartner in September 2026 illustrates this gap. Among 1,303 respondents from organizations generating at least $50 million in annual revenue, only 22% said they had successfully deployed AI across multiple business units or had adopted a truly AI-centric approach.
At the same time, 85% of the functional leaders surveyed planned to increase their AI spending in 2026. The industry is therefore investing in increasingly powerful capabilities while a large share of organizations have not yet succeeded in deploying the capabilities that already exist at scale.
Using a Chatbot Is Not the Same as Transforming a Workflow
Part of this difference comes down to what is meant by “use.” Asking a model to summarize a document, draft a message, search for information or suggest a few lines of code can already save time. But these uses remain occasional: the user retains control of most of the process and calls on AI at specific stages.
Taking advantage of more advanced capabilities requires something else. The system must be given access to the right data, be able to work with the organization's software, use tools, retain the necessary context and sometimes carry out a sequence of actions on its own. AI then stops being merely an interface that answers questions and becomes a component of the workflow itself.
This transition is already underway. A study conducted by Anthropic and Material involving more than 500 technical leaders found that 57% of the organizations surveyed were using agents for multi-step workflows. But only 16% reported having reached cross-functional processes involving multiple teams.
These results are not directly comparable with Gartner's findings, because the surveyed populations and definitions differ. They nevertheless point to the same transition: advanced uses already exist, but moving from an isolated task to deep integration across an organization remains another step entirely.
The Main Obstacle Is Not Always the Model
This gap also helps explain why immediately demanding a more powerful model does not necessarily solve the problem an organization is facing. If the AI does not have access to the necessary data, cannot communicate with existing software or is operating within a poorly defined process, a few additional points of model performance may have far less impact than better integration.
In the Anthropic and Material survey, 46% of respondents cited integration with existing systems as an obstacle to deploying agents. Access to data and data quality were cited by 42%, implementation costs by 43%, security or compliance concerns by 40%, and training requirements or employee resistance by 39%.
Some of the limitations attributed to AI therefore actually stem from the environment in which it must operate. A model may be capable of performing an operation even though the organization is not yet ready to provide the access, data or level of autonomy required for it to perform that operation in production.
Reliability Also Requires Limits on Autonomy
Not every apparent case of underuse, however, represents a delay that necessarily needs to be overcome. In some areas, deliberately limiting AI autonomy remains a rational choice. The more a system can act on data, accounts, code or sensitive infrastructure, the greater the potential consequences of an error that may be difficult to reverse.
Capabilities are therefore advancing faster than the level of trust that can reasonably be placed in them in certain environments. An organization may technically be able to allow an agent to perform more actions while still choosing to require human approval before an important operation.
Security adds another constraint. Connecting an agent to internal tools gives it precisely the access it needs to be useful, but it also increases what must be protected: identities, permissions, sensitive data, internal services and the actions that a compromised or misdirected system could trigger.
Previous Models Do Not Become Useless With Every New Generation
The technology race can easily create the impression that a model becomes obsolete as soon as its successor appears. For many applications, however, that is not how the problem should be approached. A classification, extraction, summarization, translation or simple automation task may not require the most powerful model available.
As the range of available models expands, choosing the right model becomes a form of optimization in itself. A smaller, faster or less expensive system may be preferable when it already delivers the required level of quality. The most advanced models can then be reserved for tasks where their additional capabilities make a meaningful difference.
AI progress therefore does not necessarily mean constantly replacing everything that came before. It can also lead to an ecosystem of models specialized according to cost, speed, reasoning capability or data sensitivity.
The Next Breakthrough Could Come as Much From Usage as From the Models Themselves
This may be the most important paradox of the current situation. The industry is spending hundreds of billions of dollars to build the capacity required for the next generations of systems, while part of the economic and operational potential of the generations already available has yet to be fully integrated.
This does not make the pursuit of more capable models unnecessary. Some tasks remain beyond their reach, reliability still needs to improve, and new capabilities may enable applications that do not exist today. But some of the value created over the next few years could also come from a less spectacular shift: learning how to connect, manage and properly use capabilities that already exist.
The race for more power must therefore be considered in both directions. How much further will model capabilities need to increase? But also: how far can we go with the capabilities we already have before additional power is truly necessary?
How the Situation Has Evolved
This timeline brings together the developments that have materially changed the trajectory described in this report: advances in capabilities, the growing scale of infrastructure, the emergence of new economic or energy constraints and changes in how AI is being used. It is not intended to catalogue every new model or every investment announcement.
2022-2023: Generative AI Becomes a Mainstream Product
The rapid spread of conversational assistants marks the first major change in scale. Capabilities that had previously been seen mainly in laboratories become accessible to hundreds of millions of users and begin to be integrated into software, search engines and professional tools.
At this stage, competition still focuses largely on the models themselves: response quality, code generation, image processing, context-window size and the first multimodal capabilities.
2024: Competition Shifts Toward Reasoning and Infrastructure
Improving models no longer depends solely on training. A new generation of systems begins using more compute when solving certain problems and develops the ability to sustain longer reasoning processes. At the same time, demand for processors, data centers and energy becomes increasingly important.
The question therefore begins to change: it is no longer simply about who has the best model, but who has the computing capacity required to train and operate the next generations of systems.
2025: AI Agents Become the New Goal
The shift toward systems capable of using tools, chaining together multiple operations and pursuing a task with less human intervention becomes one of the industry's main areas of development.
At the same time, the gap between demonstration and large-scale deployment becomes more visible. Many companies are experimenting with AI, but turning those experiments into reliable and profitable processes remains much more difficult than simply providing access to a model.
Early 2026: Capabilities Continue to Advance Rapidly
Evaluations focused on AI agents show that some systems can now complete much longer technical tasks than before. Progress remains particularly visible in software development and in environments where results can be precisely verified.
This development strengthens the prospect of systems capable of handling increasingly large portions of a workload, but it also preserves an essential distinction: successfully completing long, well-defined tasks does not mean possessing general autonomy in complex and ambiguous professional environments.
2026: Investment Reaches a New Scale
Competition in AI is simultaneously becoming an industrial competition. The world's largest technology companies are committing hundreds of billions of dollars to the data centers, processors, networks and energy capacity required for the next generations of systems.
Electricity supply and grid connections are themselves becoming determining factors. Some projects are now seeking to directly secure their own power generation or develop off-grid capacity rather than wait for existing electricity infrastructure to expand.
2026: The First Economic Strains Become More Visible
This expansion is not stopping, but its financing is coming under greater scrutiny. Long-term commitments related to data centers are increasing sharply, and companies that depend most heavily on debt are becoming more sensitive to market concerns.
Among companies using AI, the same gap is appearing in another form: spending continues to increase while the transition from experimentation to applications deployed at scale remains much slower.
October 17, 2026: Where Does the Situation Stand?
There is currently no clear evidence that progress in AI or the investment supporting it is coming to a halt. Capabilities continue to improve, new infrastructure is being built and the leading players remain engaged in the race.
The situation is no longer the same, however, as it was during the early years of generative AI. Technological progress now depends on considerable physical and financial infrastructure: processors, data centers, energy, power grids and capital commitments extending over several years.
At the same time, available capabilities are advancing faster than their integration across a large share of organizations. The next stage therefore does not depend solely on even more powerful models. It also depends on making existing systems sufficiently reliable, accessible, integrated and economically useful.
This is the dual trajectory that now needs to be watched. On one side is the question of how far technical capabilities and infrastructure can continue to advance. On the other is whether AI adoption and the revenue it generates can grow quickly enough to sustainably justify this race for more power.
Going further: data and monitoring
Alongside its reports, Palks Studio also tracks a range of indicators using public data, making it possible to observe changes over time across areas such as energy, infrastructure, digital systems and major flows.
Explore the trackers and public data →
Follow developments more closely
To follow important developments more directly, you can also find L'Actu sans le Bruit on Telegram. The channel follows new events as they unfold, while this report is updated only when those developments materially change the situation.
The channel is published in French, but Telegram's translation feature can be used to translate posts into English.