A phone screen is a grid of icons. Dozens of applications, each a separate door to a service, a bank, a newspaper, a map, a shop. That grid is the map of a fragmented economy and, inconvenient as it is, fragmentation has a virtue we tend not to see: it distributes control across thousands of different actors. For some time now, the industry has been telling us that the grid is destined to disappear. In its place, a single conversational interface, an assistant you ask in natural language for what you want and which takes care of the rest: booking, buying, searching, writing. You no longer open apps, you express an intention. The promise is convenience, the risk is concentration.
In a previous article I tried to explain what an artificial intelligence system actually does, and why it is not neutral. It reduces pieces of reality to numbers, learns by successive approximation and, to do so, depends on three very material things: data, that is, reality already turned into numbers, computing power, and the definition of the objective, that is, the choice of which problem the system should solve. The conclusion, there, was that whoever governs this material base ultimately decides what the technology does for people and, to some extent, to people as well.
The possible disappearance of the grid of icons adds a fourth face to the same dependency, and perhaps the most intimate of all: mediation. If artificial intelligence becomes the single interface between us and every piece of software, then what matters is not only who owns the data, the compute and the objective, but who controls the point of contact, the “place” where human intention meets the machine. And here something happens that the enthusiasm for the end of apps tends to obscure. The layer that fades is not software, but its direct interface with the user. Control over access to software, on the other hand, what is known in the jargon as gatekeeping, the guardian function that decides who passes and on what terms, does not evaporate. It moves up a layer.
Today that control is unevenly distributed: infrastructural gatekeeping is already concentrated in a few decisive chokepoints, above all operating systems and app stores, but the final relationship with the user remains distributed across thousands of applications and services. A universal conversational interface could concentrate this second layer too: not just access to the device, but the everyday choice of which service to use.
That power would end up in the hands of whoever manages the relationship with the user and decides which models, applications and services to call upon. I will call this ensemble the mediation layer: the interface through which the user formulates their intention, the memory that preserves its context, and the system that decides which models, applications and services to activate.
Whoever produces the model may coincide with whoever controls this layer, and the large platforms are working to make them coincide. But it is not inevitable: model, assistant, operating system and service orchestration are technically separable layers, and precisely this separability, as we shall see, opens the space for a response.
The question “do we really need so many applications” is, on the surface, a matter of convenience. Underneath, it is the politically densest question this technology poses: who controls the single interface that replaces them? It is the same materialist pattern, the same dependency on computing capacity and infrastructure I have already written about, shifted from the plane of hardware to that of the everyday mediation between people and technology. And it is how an apparently technical question, about the future of apps, becomes a question about sovereignty.
One clarification before proceeding: what follows is not the description of an already accomplished present. Apps have not disappeared and assistants have not yet replaced software. It is the diagnosis of a movement under way, and the structure that could bring it to completion, as we shall see, is already visible. To understand why power behaves this way, why control over access to software “moves up a layer” instead of dispersing, it is worth starting from a theory formulated some time ago, which has already described this same movement twice.
The mechanism: why power "moves up a layer"
Much is being said about the relationship between artificial intelligence and software, but almost always around two questions: whether developers become more productive, and how many jobs will be lost. These are legitimate questions, and the data around them is more contradictory than the debate admits, but they remain inside a perimeter that takes the unfolding of that relationship for granted. The larger stake, control of the interface that mediates between people and technology, is left out of the frame.
To bring it into focus, it is in my view useful to recall a theory Ben Thompson first formulated in 2015, Aggregation Theory, which has since become one of the most widely used keys to understanding the economics of the internet. The idea is simple and powerful. In the pre-digital economy, power lay in control of distribution: whoever owned the printing presses, the frequencies, the shelves decided what reached the public. The internet slashed the marginal cost of digital distribution, and power shifted: no longer towards those who control supply, but towards those who aggregate demand. Whoever owns the relationship with the user, whoever can consider themselves “the point the user starts from”, can make every downstream supplier interchangeable, and impose their own market conditions upon them.
Recent history offers numerous examples of this same “backwards” movement. Search did it to publishers: Google has never written a news story, but by being the point from which people set off to find news online, it made newspapers, from the standpoint of access, largely substitutable suppliers of content, capturing the value of mediation. The App Store did it to software: Apple does not write most of the applications it distributes, but for years it controlled the only authorised distribution channel on the iPhone and was able to impose its rules and its commission on developers. The DMA has begun to open alternative channels in the European Union, and the precedent remains instructive twice over: it shows how much power derives from control of the access point, and it shows that such control can be made contestable through regulation. In both cases the underlying layer did not disappear, newspapers still exist (as do their publishers), and so do apps (along with their actual developers). What moved upwards, towards the demand aggregator, was the power to define the very conditions of their existence.
The conversational interface is shaping up as the third act of the same story and the most radical. When interaction with technology shifts from navigating menus to delegating goals, when you tell the assistant “book me a table” instead of opening the restaurant’s app, aggregation no longer concerns a category of content or a platform, it concerns human intention as such. The assistant becomes the point you start from for anything, and everything behind it, services, applications, catalogues, becomes interchangeable supply. Apps do not die, they recede: from interfaces with a direct relationship with the user to backend services answering the model’s calls, visible only if and when the aggregator decides to consult them, on the terms the aggregator sets.
This prediction, however, must immediately be stripped of its veneer of technical inevitability, because on one point enthusiasm runs faster than substance. Natural language is very good at expressing intent, but it cannot replace the formal, executable semantics software is made of: state, data models, backwards compatibility, correctness guarantees do not go away because the user speaks instead of clicking. Those who have seriously analysed the hypothesis of “disposable” software generated on the fly have reached the same conclusion: as soon as robustness is required, the ephemeral part shrinks to surface details and everything that makes the system work remains persisted, engineered, maintained. Software, then, does not fade. But this objection, which seems to shrink the thesis, in fact strengthens it: if the software layer remains, and what changes is only who controls access to it, then the main stake is not the disappearance of software, but the power of mediation over its access. Value does not shift because software disappears, it shifts because the direct relationship between those who make software and those who use it disappears.
And there is a precise reason why this time the aggregator is harder to contest than in the past. Google could be avoided by typing a website’s address, the App Store at least had the Android ecosystem as an alternative, albeit within another closed model. An interface that absorbs intention, that learns habits, that accumulates the context of whoever uses it, raises exit costs of a new kind: switching does not mean changing supplier, it means starting over at making yourself known. The lock-in concerns not only the data stored, but the inferred preferences, the permissions granted, the routines built and the history through which the assistant has learned to interpret ever more implicit requests. It is the kind of position economics calls, with some understatement, hard to contest. European law has a more honest word for such positions.
The legal precedent: gatekeeping is not a metaphor
Gatekeeping, in Europe, is not a figure of speech. It is a legal category, with quantitative criteria, obligations and sanctions. The digital markets regulation, the Digital Markets Act, allows the Commission to designate as gatekeepers the undertakings that provide “core platform services”, the platform services that act as a gateway between businesses and end users. The list of covered services is instructive, because it is the exact map of the mediation layers that power has occupied in previous eras: search engines, operating systems, app stores, browsers, social networks, intermediation services. Every entry on that list is a chapter of the story told above, a point where demand aggregated and power concentrated enough to require a regulatory response.
And the list already contains the entry that interests us: virtual assistants have been among the core platform services provided for by the regulation since its adoption. The European legislator, in other words, had already foreseen in 2022 the possibility that the conversational assistant might become a gateway as search and app stores had been. To date, there is no standalone designation of a conversational assistant as a core platform service, for reasons that combine the quantitative thresholds, the classification of services and the integration of assistants inside already designated platforms, but the box exists, waiting to be filled.
The point is that the technology is running towards that box faster than the thresholds can measure, not least because the quantitative thresholds, centred mostly on economic scale and user numbers, risk missing the quality of the intermediation: how many intentions an assistant captures, how much personal context it accumulates, how far it can steer the choice of downstream services, how much the businesses passing through it depend on it to reach their customers.
The first review of the regulation, published by the Commission in April 2026, confirmed the framework, recognised that artificial intelligence services may already fall within the regulation when they are integrated into designated services or, where distinct, be assessed as virtual assistants, and declared its intention to explore how the DMA applies to artificial intelligence and cloud, while Parliament, through its internal market committee, has explicitly called for artificial intelligence services to be closely monitored, observing that they are tightly interdependent with the cloud services already covered.
And it is no longer just exploration. In November 2025 the Commission opened market investigations into Microsoft Azure and Amazon Web Services and, on 25 June 2026, communicated its preliminary position: both should be designated as gatekeepers, the first application of the regulation to cloud infrastructure. Two details make this step important for the argument of this piece. First: the position rests on a qualitative assessment, lock-in, high switching costs, the breadth of the ecosystems, and not on the regulation’s quantitative thresholds, which the two services do not meet, confirming that the quality of intermediation can weigh more than measurable size. Second: among the reasons given is, explicitly, the growing weight of artificial intelligence tools and partnerships in cloud purchasing decisions. The final decision has not yet been adopted, but the direction is clear: European law is already reading the connection between infrastructure and the new mediation layer.
The discussion, in other words, is no longer whether artificial intelligence services can perform gatekeeping functions, but how to qualify the different layers, the model, the assistant, the operating system, the orchestration, and when to apply the existing instruments to them. Those who present the conversational interface as a new, uncharted space, foreign to the categories of regulation, do not see, or prefer not to see, that the categories of existing European law are already capable of functionally qualifying these new services: they were written precisely by observing how this movement ended the two previous times.
There is, however, a difference between past eras and this one, and it is worth stating clearly, because it is the reason why the legal framework alone is not enough. The DMA was written against aggregators that were already mature, after fifteen years of entrenched positions, when, to use the old saying, the horse had long since bolted. With the conversational interface the window is different: the concentration is forming now, in plain sight, and there is no shortage of data to describe it.
The fact: the new mediation is being born into an already concentrated market
If power moves up towards the model layer, the empirical question is what that layer looks like. And the answer, by now documented by a substantial body of institutional analysis, is that it already displays an oligopolistic structure, tightly interwoven with the infrastructure beneath it.
The UK competition authority, the CMA, moved from curiosity to concern after its review of foundation models, describing an “interconnected web” of more than ninety partnerships and strategic investments revolving around the same names, Google, Apple, Microsoft, Meta, Amazon, Nvidia, the same firms that already hold the gatekeeping positions of the previous era. This is neither an accident nor a conspiracy, it is structure: developing the most capable models requires very high fixed costs, privileged access to compute and distribution at scale, conditions that generate strong economies of scale and reward the actors already integrated into the large digital ecosystems. In markets built this way, concentration is not a possible degeneration, it is a structural tendency.
Beneath the model layer sits compute, and there the concentration is even starker. The British authority, at the close of its market investigation, found AWS and Microsoft to hold significant market power in cloud and documented barriers to switching, data egress costs and licensing practices; the response that arrived in 2026 was more selective than the diagnosis, but its stated rationale is itself a confirmation of this piece’s thesis: acting now serves to prevent the power accumulated in cloud from transferring to the emerging markets of artificial intelligence. The most telling fact, however, is how the two layers have welded together: Microsoft’s investment in OpenAI was born around a privileged cloud relationship, and Anthropic has built deep infrastructure partnerships first of all with Amazon, then with Google and more recently with Microsoft and Nvidia. The layer that produces the models and the layer that owns the compute are not two merely adjacent markets: they are becoming a single system of cross-dependencies. Those who have read my previous article will recognise the foundation: according to Epoch AI’s estimates, roughly three quarters of the global performance of large GPU clusters for artificial intelligence sits in the United States, while the European Union stops at around five per cent. The interface rests on that subsoil.
This concentration also produces a systemic fragility: if a disproportionate number of businesses and services depends on the same few providers of models and compute, the failure or suspension of a single one propagates along the entire chain. Concentration is not only a question of who is in charge, it is a question of what happens when the single point of passage stops.
And there is one case worth more than any statistic, because it shows the harm this dependency can do at the most sensitive point of all, the security of critical infrastructure. In August 2025 DARPA’s AI Cyber Challenge came to a close, the competition with which the US government demonstrated that autonomous systems can find and fix vulnerabilities in software relevant to critical infrastructure. According to the results presented by DARPA, in the final the systems identified 54 of the 63 synthetic vulnerabilities set by the competition and fixed 43 of them, as well as discovering 18 real vulnerabilities in widely used open source projects, as subsequent accounts also report. A remarkable result, presented as a step towards the automatic defence of the software we all depend on. But those systems were born and raised inside an ecosystem materially sustained by credits, models and expertise provided by Anthropic, Google, OpenAI, Microsoft and other partners. Even the exercise with which a State demonstrates it can defend its own critical infrastructure was made possible, in its development phase, by resources and models supplied by the same few private actors that dominate the foundation model layer. And it is true that the finalist systems were released as open software, but what is open is the system built around the models, not the model layer it depends on. The transparency of the application code does not remove the dependency on the layer beneath it. The case shows that even a US public programme dedicated to the security of critical software was developed by drawing on the resources of the few actors that dominate models and compute. For countries and institutions with lesser industrial capacity, this dependency is plausibly even more marked.
This is the plane onto which power is climbing. Not an open, contested market in which the best interface wins, but a layer consolidating around a very small number of actors, vertically integrated with compute, geographically concentrated above all in the United States.
It is worth specifying, however, that although these phenomena can be explained through the same theory, they are at different stages of realisation: the aggregation of search and app stores is what has already happened, the concentration of the model layer is what is happening today, the single interface is what risks happening tomorrow.
The end of the grid of icons, seen in this light, is therefore not the user’s liberation from fragmentation, but the transfer of the direct relationship with the user from thousands of competing applications and services to a handful of intermediaries able to decide which of those services to make visible and on what terms: guardians the law has not yet qualified as such.
The enclosure: the same pattern, one layer up
There is a word that captures this mechanism, and it comes from the scholar who has looked most deeply into the materiality of artificial intelligence. Kate Crawford, in Atlas of AI, advances a deliberately provocative thesis: AI is neither artificial nor intelligent, it is an extractive industry, made of natural resources, labour, infrastructure and logistics, and its infrastructures reflect the beliefs and serve the interests of a very small group of people. The pattern she identifies has a name taken from economic history: enclosure, the process by which the English commons were turned into private property. The result, she writes, is not the democratisation of intelligence but the enclosure of sovereignty.
I find it an effective word for the idea, because it describes exactly the nature of the movement under way on the interface. The ability to make oneself understood by a machine in natural language, to have things explained, translated, assisted, is born of a largely collective cognitive heritage: texts, images, software and knowledge produced by generations of people, together with decades of research often publicly funded. The transformation of that heterogeneous heritage into an owned service, delivered through private infrastructure, optimised towards objectives decided elsewhere, is a contemporary form of enclosure: something that belonged to the common landscape becomes a plot of land with a gate, and the gate has a definite owner.
In my previous article I wrote that whoever governs the material base (data, compute and objective) ultimately decides what the technology can do for people, but also what it can do to people.
Mediation carries that conclusion one layer up and makes it a daily matter.
Whoever controls the single interface does not necessarily decide how each model was trained, but decides which models to employ, what context to give them, which services to make accessible. It decides what we see when we ask, which services are offered to us and which remain invisible, which answers are possible and which are not, in what order the world is presented to us. It is the power that search and app stores had, but extended from access to information and software to the whole spectrum of human intention, and made more opaque, because a generated answer has no results page to inspect, no visible ranking to argue about, and does not let us see what it discarded.
And the enclosure does not stop at use, it reaches production too. A study published by Anthropic on its own usage data for agentic coding tools, with the limits inherent in a study conducted by a vendor on the usage data of its own product, suggests that success does not depend only on knowing how to code, but on how well one knows the problem being solved: agents do not replace domain expertise, they reward it. Read against the light, it is a datum about the distribution of power: if writing software becomes describing what one wants, value tends to divide between those who hold the domain expertise and those who control the horizontal layer through which that expertise is turned into software. The former changes in every sector, the latter can serve all sectors at once. Every developer, business or administration that builds durably through a proprietary interface transfers value and dependency, in varying measure, to whoever controls the mediation layer. The rent of the enclosure is levied upstream of everyone.
This is the point at which the question stops being a curiosity of industrial economics and becomes what the title of this piece promises. Sovereignty, in its most concrete definition, is the capacity to decide the material conditions in which a community works, learns, exercises its rights. If the point of contact between people and the entire digital sphere, the place where what they know, what they buy, what they ask for takes shape, belongs to a handful of private actors outside European industrial control and subject to decision-making centres located in other jurisdictions, however regulated when they operate in the European market, then the enclosure of sovereignty is not an academic metaphor. It is a technical description of the state of affairs.
The response: an open and contestable mediation
If the diagnosis is right, the response can be neither the rejection of this new technology nor the hope that the market developing it will correct itself, because a market marked by strong economies of scale, vertical integration and high entry costs rarely corrects itself.
The response must be sought, then, where the highest stake has been identified: in the control and governance of the mediation layer. And here, as usual, there are two planes, one technological and one political, and they must be held together.
The technological plane has a precise name: malleable software. The independent laboratory Ink & Switch, in one of the most lucid essays of recent years, recalls that the original promise of personal computing was a new “clay”, a tool each person could shape to their own needs, and that the promise was betrayed: we got appliances, built far away, sealed, unmodifiable. Artificial intelligence, they observe, can still go either way: it can complete the transformation of the user into the consumer of a closed service, or it can be the technology that finally makes the “clay” accessible, because it lowers as never before the barrier between having a need and building the tool that meets it. The difference between the two outcomes does not lie in the model, it lies in the architecture around it: tools the user can inspect, recombine and modify, data that remains theirs, components that talk to each other through open standards rather than through a single gate. It is the same distinction I have drawn elsewhere between the two outcomes of the workforce management algorithm: the technology is identical, the control is not.
The political plane is the territory of public digital infrastructure, and Europe, here, has more experience than is commonly assumed. A public digital infrastructure is a common base, democratically governed, on which the public sector and the market build their services: India has done it with UPI payments, Brazil with PIX, Europe is doing it with the European Digital Identity Wallet and, on the digital currency, with the digital euro, which I have discussed in a dedicated piece. And the wallet is worth a closer look, because it is more pertinent than it seems.
The European identity wallet is a mediation layer subject to public rules: it is designed to connect the citizen to a plurality of public and private services through common standards that anyone can implement, without entrusting a single private platform with control of the whole ecosystem. Its public quality lies in its governance, certification and interoperability, not necessarily in the ownership of the code or the operator. It offers a concrete institutional precedent: a mediation function can be designed and governed as common infrastructure rather than as rent, even when its implementation involves public and private actors. And the model is not theory: UPI and PIX show that infrastructures governed through common rules and components can operate at the scale of hundreds of millions of people, while the European wallet, which Member States will make available by the end of 2026, attempts to carry a similar logic, with its own characteristics, to the most delicate layer of all, the identity of citizens.
Bringing this logic to the conversational interface does not mean imagining a “state-owned ChatGPT”, which would be the wrong answer to a right question. It means acting on three levers: architecture, public capacity, regulation.
The first lever is architecture, and it is the most important, because the whole diagnosis of this piece says that architecture matters more than the ownership of any single component. A genuinely contestable interface should allow the user to change model without losing memory and configurations, to choose which services may be called upon, to verify why a supplier was selected, to transfer identity, permissions and habits to another assistant. And separability must also hold towards downstream services: we need open protocols to describe services’ capabilities, authorise their use, invoke them and return a verifiable trace of what was done, protocols that are in fact already emerging in industry practice, otherwise it will be possible to replace the model, but not the system that decides which services exist, which may be called and on what terms. The user’s sovereignty does not coincide with public ownership of every component: it consists in the effective possibility of replacing them.
The second lever is public capacity. Full openness means making the process inspectable, not just distributing the weights. European public initiatives are moving in this direction with varying degrees of openness, from fully reproducible projects such as Apertus to publicly backed models distributed under more restrictive terms, by way of ALIA, Minerva and GPT-NL. None of these projects is, on its own, an alternative at the mediation layer, because an open model is a component, not an interface. But they are components that are substitutable and inspectable down to the choices, not just the result, on which open assistants can be built, preventing mediation from depending on a single proprietary model. On commercial open weights, the caution I have already argued still holds: opening the weights democratises use, not production, and we can now add the layer that was missing: it does not democratise mediation either, because if the interface that aggregates intention remains in the hands of a single actor, re-centralisation happens anyway, one layer above the weights. To the same lever belongs public procurement, which the Commission’s “sovereignty package” has finally taken up: it can do for the interface what I have argued it should do for training, not buy sealed boxes, but impose conditions of interoperability, auditability and portability of the user’s context, with the possibility of replacing the underlying model without losing everything. It is how public demand creates a market for the contestable architecture of the first lever. And alongside public demand can stand public revenue: adequate taxation of the digital rents generated in Europe, with the proceeds directed to industrial policy and to the building of European alternatives, including State aid where justified.
The third lever is regulation. The gatekeeping framework must be kept ready: the virtual assistant box exists, the criteria for filling it must be adapted to an aggregator whose strength is measured not only in registered users, but also in captured intentions and accumulated context, and this time, as the cloud case shows, regulation can arrive while the position is forming, not fifteen years later. Nor does this approach renounce competition law’s structural weapon, dismemberment: it keeps it as a last resort. The separability of the first lever is, prospectively, its ex ante transposition, preventing the layers from welding into a single block rather than breaking it apart once the damage is done, and it also provides the map along which any future break-up would have to cut. The DMA, for that matter, already provides for structural remedies, up to the divestiture of businesses, in cases of systematic non-compliance.
The three levers, finally, presuppose an upstream condition: that a sufficient degree of pluralism be maintained in the underlying market, a task that belongs to competition law and to a DMA applied more ambitiously than it is today. None of these three levers, on its own, stops the movement I have described, and that is not the point. The point is to prevent the movement from completing itself without an alternative, because it is the absence of an alternative, not the existence of the aggregator, that turns a market position into a de facto quasi-constitutional power: the private capacity to set the conditions through which people and businesses access a growing share of the digital sphere. The grid of icons, with all its flaws, always offered an alternative: another app, another store, the browser. The political task of the coming years is to make sure the conversational interface has one too, be it public, open or malleable.
I closed my previous article by writing that the question is not technological, it is political. This piece adds a specification: it is also the oldest question in politics, the one about who controls the controllers. Artificial intelligence as a universal interface can become a guardian of unprecedented reach between people and the world of their possibilities. Deciding how to distribute and govern its control while that position is still taking shape, rather than merely recording its effects afterwards, is the difference between choosing and enduring. Once again, between sovereignty and enclosure.
The information and views set out in this article are those of the author and do not necessarily reflect the official opinion of the European institutions.